Innovations in Bio-Inspired Computing and Applications: Proceedings of the 12th International Conference on Innovations in Bio-Inspired Computing and ... (Lecture Notes in Networks and Systems, 419) 3030962989, 9783030962982

Citation preview

Lecture Notes in Networks and Systems 419

Ajith Abraham · Ana Maria Madureira · Arturas Kaklauskas · Niketa Gandhi · Anu Bajaj · Azah Kamilah Muda · Dalia Kriksciuniene · João Carlos Ferreira   Editors

Innovations in Bio-Inspired Computing and Applications Proceedings of the 12th International Conference on Innovations in Bio-Inspired Computing and Applications (IBICA 2021) Held During December 16–18, 2021

Lecture Notes in Networks and Systems Volume 419

Series Editor Janusz Kacprzyk, Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Advisory Editors Fernando Gomide, Department of Computer Engineering and Automation—DCA, School of Electrical and Computer Engineering—FEEC, University of Campinas— UNICAMP, São Paulo, Brazil Okyay Kaynak, Department of Electrical and Electronic Engineering, Bogazici University, Istanbul, Turkey Derong Liu, Department of Electrical and Computer Engineering, University of Illinois at Chicago, Chicago, USA Institute of Automation, Chinese Academy of Sciences, Beijing, China Witold Pedrycz, Department of Electrical and Computer Engineering, University of Alberta, Alberta, Canada Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland Marios M. Polycarpou, Department of Electrical and Computer Engineering, KIOS Research Center for Intelligent Systems and Networks, University of Cyprus, Nicosia, Cyprus Imre J. Rudas, Óbuda University, Budapest, Hungary Jun Wang, Department of Computer Science, City University of Hong Kong, Kowloon, Hong Kong

The series “Lecture Notes in Networks and Systems” publishes the latest developments in Networks and Systems—quickly, informally and with high quality. Original research reported in proceedings and post-proceedings represents the core of LNNS. Volumes published in LNNS embrace all aspects and subfields of, as well as new challenges in, Networks and Systems. The series contains proceedings and edited volumes in systems and networks, spanning the areas of Cyber-Physical Systems, Autonomous Systems, Sensor Networks, Control Systems, Energy Systems, Automotive Systems, Biological Systems, Vehicular Networking and Connected Vehicles, Aerospace Systems, Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems, Robotics, Social Systems, Economic Systems and other. Of particular value to both the contributors and the readership are the short publication timeframe and the world-wide distribution and exposure which enable both a wide and rapid dissemination of research output. The series covers the theory, applications, and perspectives on the state of the art and future developments relevant to systems and networks, decision making, control, complex processes and related areas, as embedded in the fields of interdisciplinary and applied sciences, engineering, computer science, physics, economics, social, and life sciences, as well as the paradigms and methodologies behind them. Indexed by SCOPUS, INSPEC, WTI Frankfurt eG, zbMATH, SCImago. All books published in the series are submitted for consideration in Web of Science. For proposals from Asia please contact Aninda Bose ([email protected]).

More information about this series at https://link.springer.com/bookseries/15179

Ajith Abraham Ana Maria Madureira Arturas Kaklauskas Niketa Gandhi Anu Bajaj Azah Kamilah Muda Dalia Kriksciuniene João Carlos Ferreira •

•

•

•

•

•

•

Editors

Innovations in Bio-Inspired Computing and Applications Proceedings of the 12th International Conference on Innovations in Bio-Inspired Computing and Applications (IBICA 2021) Held During December 16–18, 2021

123

Editors Ajith Abraham Scientific Network for Innovation and Research Excellence Machine Intelligence Research Labs (MIR Labs) Auburn, WA, USA Arturas Kaklauskas Department of Construction Management and Real Estate Vilnius Gediminas Technical University Vilnius, Lithuania Anu Bajaj Scientific Network for Innovation and Research Excellence Machine Intelligence Research Labs (MIR Labs) Auburn, WA, USA Dalia Kriksciuniene Vilnius University Kaunas, Lithuania

Ana Maria Madureira Departamento de Engenharia Informática Instituto Superior de Engenharia do Port Porto, Portugal Niketa Gandhi Scientific Network for Innovation and Research Excellence Machine Intelligence Research Labs (MIR Labs) Auburn, WA, USA Azah Kamilah Muda Faculty of Information Communication Technology Universiti Teknikal Malaysia Melaka Durian Tunggal, Melaka, Malaysia João Carlos Ferreira Lisbon University Institute Lisbon, Portugal

ISSN 2367-3370 ISSN 2367-3389 (electronic) Lecture Notes in Networks and Systems ISBN 978-3-030-96298-2 ISBN 978-3-030-96299-9 (eBook) https://doi.org/10.1007/978-3-030-96299-9 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Preface

Welcome to the 12th International Conference on Innovations in Bio-Inspired Computing and Applications (IBICA 2021) and 11th World Congress on Information and Communication Technologies (WICT 2021) held online during December 16–18, 2021. The aim of IBICA is to provide a platform for world research leaders and practitioners, to discuss the full spectrum of current theoretical developments, emerging technologies, and innovative applications of bio-inspired computing. Bio-inspired computing is currently one of the most exciting research areas, and it is continuously demonstrating exceptional strength in solving complex real-life problems. WICT provides an opportunity for researchers from academia and industry to meet and discuss the latest solutions, scientific results, and methods in the usage and applications of ICT in the real world. Innovations in ICT allow us to transmit information quickly and widely, propelling the growth of new urban communities, linking distant places and diverse areas of endeavor in productive new ways, which a decade ago was unimaginable. Thus, the theme of this World Congress is “Innovating ICT For Social Revolutions.” IBICA–WICT 2021 brings together researchers, engineers, developers, and practitioners from academia and industry working in all interdisciplinary areas of intelligent systems, nature-inspired computing, big data analytics, real-world applications, and to exchange and cross-fertilize their ideas. The themes of the contributions and scientific sessions range from theories to applications, reflecting a wide spectrum of the coverage of intelligent systems and computational intelligence areas. IBICA 2021 received submissions from 15 countries, and each paper was reviewed by at least five reviewers in a standard peer-review process. Based on the recommendation by five independent referees, finally 33 papers were presented during the conference (acceptance rate of 28%). WICT 2020 received submissions from 24 countries, and each paper was reviewed by at least five reviewers in a standard peer-review process. Based on the recommendation by five independent referees, finally 47 papers were presented during the conference (acceptance rate of 34%).

v

vi

Preface

Many people have collaborated and worked hard to produce the successful IBICA–WICT 2021 conference. First, we would like to thank all the authors for submitting their papers to the conference, for their presentations and discussions during the conference. Our thanks go to the program committee members and reviewers, who carried out the most difficult work by carefully evaluating the submitted papers. Our special thanks go to the following plenary speakers, for their exciting plenary talks: • • • • • • • • • •

Yukio Ohsawa, The University of Tokyo, Japan Juergen Branke, University of Warwick, UK Cengiz Toklu, Beykent University, Turkey Günther Raidl, Technische Universität Wien, Austria Kalyanmoy Deb, Michigan State University, USA Oscar Cordon, University of Granada, Spain Andries Engelbrecht, University of Stellenbosch, South Africa Antônio de Padua Braga, Federal University of Minas Gerais, Brazil Frédéric Guinand, Le Havre Normandy University, France Marco Dorigo, Université Libre de Bruxelles, Belgium

Our special thanks to the Springer Publication team for the wonderful support for the publication of these proceedings. We express our sincere thanks to the session chairs and organizing committee chairs for helping us to formulate a rich technical program. Enjoy reading the articles! Ajith Abraham Ana Maria Madureira Arturas Kaklauskas General Chairs Dalia Kriksciuniene João Carlos Ferreira Nuno Bettencourt Azah Kamilah Muda Program Chairs

IBICA – WICT Organization

General Chairs Ajith Abraham Ana Maria Madureira Arturas Kaklauskas

Machine Intelligence Research Labs (MIR Labs), USA Instituto Superior de Engenharia do Porto, Portugal Vilnius Gediminas Technical University, Lithuania

Program Chairs Dalia Kriksciuniene João Carlos Ferreira Nuno Bettencourt Azah Kamilah Muda

Vilnius University, Lithuania ISCTE-IUL, Portugal Instituto Superior de Engenharia do Porto, Portugal Universiti Teknikal Malaysia Melaka, Malaysia

Publication Chairs Niketa Gandhi Kun Ma

Machine Intelligence Research Labs (MIR Labs), USA University of Jinan, China

Special Sessions Chairs Maria Leonilde Varela Gabriella Casalino Catarina Reis

Universidade do Minho, Portugal University of Bari, Italy Polytechnic Institute of Leiria, Portugal

vii

IBICA – WICT Organization

viii

Publicity Chairs Aswathy S. U. Pooja Manghirmalani Mishra Mahendra Kanojia Anu Bajaj

Jyothi Engineering College, Kerala, India University of Mumbai, Maharashtra, India MVLU College, Maharashtra, India Machine Intelligence Research Labs (MIR Labs), Washington, USA

International Publicity Team Mabrouka Salmi Phoebe E. Knight Marco A. C. Simões Hsiu-Wei Chiu Serena Gandhi

National School of Statistics and Applied Economics (ENSSEA), Kolea, Tipaza, Algeria UTeM, Malaysia Bahia State University, Brazil National University of Kaohsiung, Taiwan USA

Program Committee Agostino Forestiero Andre Santos Antonio J. Tallón-Ballesteros Anu Bajaj Aswathy S. U. Ayush Goyal Bruno Cunha Carlos Pereira Elizabeth Goldbarg Gabriella Casalino Gianluca Zaza Isaac Chairez Isabel S. Jesus Ivo Pereira János Botzheim José Everardo Bessa Maia Kaushik Das Sharma Lee Chang-Yong Mohd Abdul Ahad Niketa Gandhi Norberto Díaz-Díaz Oscar Castillo Patrick Siarry Pooja Manghirmalani Mishra

CNR-ICAR, Italy School of Engineering, Polytechnic Institute of Porto (ISEP/IPP), Portugal University of Huelva, Spain Machine Intelligence Research Labs, USA Jyothi Engineering College, Kerala, India Texas A&M University - Kingsville, USA GECAD-ISEP, Portugal ISEC, Portugal Federal University of Rio Grande do Norte, Brazil University of Bari, Italy University of Bari “Aldo Moro,” Italy UPIBI-IPN, Mexico Institute of Engineering of Porto, Portugal University Fernando Pessoa, Portugal Eötvös Loránd University, Hungary State University of Ceará, Brazil University of Calcutta, West Bengal, India Kongju National University, South Korea Jamia Hamdard, New Delhi, India Machine Intelligence Research Labs, USA Pablo de Olavide University, Spain Tijuana Institute of Technology, Mexico Universit de Paris 12, France University of Mumbai, Maharashtra, India

IBICA – WICT Organization

Poonam Ghuli Radu-Emil Precup Rajashree Shettar Shankru Guggari Thatiana C. Navarro Diniz Ankit Gupta Antonello Florio Anurag Rana Catarina I. Reis Daniele Schicchi Deepika Koundal Gaurav Gupta Gianluca Zaza Giuseppe Coviello Hiren Jayantilal Dand Kapil Sethi Kingsley Okoye Mahendra Kanojia Mohd Abdul Ahad Nuno Bettencourt Preety Baglat Sindhu P. M. Vishesh Shrivastava

ix

RV College of Engineering, India Politehnica University of Timisoara, Romania RV College of Engineering, Karnataka, India BMS College of Engineering, Karnataka, India Federal Rural University of the Semi-Arid, Brazil Interactive Technologies Institute (ITI)/LARSyS, Caminho da Penteada, Portugal Politecnico di Bari, Italy Shoolini University, Himachal Pradesh, India Escola Superior de Tecnologia e Gestão de Leiria, Portugal Institute for Educational Technology - National Research Council of Italy, Italy University of Petroleum and Energy Studies, India Shoolini University, Himachal Pradesh, India University of Bari “Aldo Moro,” Italy Politecnico of Bari, Italy Mulund College of Commerce, India Bahra University, Himachal Pradesh, India Tecnologico de Monterrey, Mexico Sheth L.U.J. and Sir M.V. College, India Jamia Hamdard, New Delhi, India Polytechnic of Porto (ISEP/IPP), Portugal Interactive Technologies Institute (ITI)/LARSyS, Caminho da Penteada, Portugal Nagindas Khandwala College, Maharashtra, India B.K. Shroff College of Arts and Commerce, India

Contents

Bio-inspired Computing and Applications BIC Algorithm for Heineken Brand Awareness in Vietnam Market . . . Nguyen Thi Ngan and Bui Huy Khoi Comparison of Ant Colony Optimization Algorithms for Small-Sized Travelling Salesman Problems . . . . . . . . . . . . . . . . . . . Arcsuta Subaskaran, Marc Krähemann, Thomas Hanne, and Rolf Dornberger

3

15

Tamilnadu Omnibus Travels Evaluation Using TOPSIS and Fuzzy TOPSIS Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . S. M. Vadivel, A. H. Sequeira, Sunil Kumar Jauhar, and V. Chandana

24

A Hybrid Feature Extraction Method Using SeaLion Optimization for Meningioma Detection from MRI Brain Image . . . . . . . . . . . . . . . . S. U. Aswathy, Divya Stephen, Bibin Vincent, and P. Prajoon

32

An Effective Integrity Verification Scheme for Ensuring Data Integrity in Cloud Computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Minakshi Kamboj and Sanjeev Rana

42

A Survey on Arrhythmia Disease Detection Using Deep Learning Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . George C. Lufiya, Jyothi Thomas, and S. U. Aswathy

55

Comparison of Different Machine Learning Methods to Detect Fake News . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tanishka Badhe, Janhavi Borde, Vaishnavi Thakur, Bhagyashree Waghmare, and Anagha Chaudhari A Tool for Air Cargo Planning and Distribution . . . . . . . . . . . . . . . . . . Diana Costa, André S. Santos, João A. Bastos, Ana M. Madureira, and Marlene F. Brito

65

78

xi

xii

Contents

A Markov Model for Improving the Performance of COVID-19 Contact Tracing App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Abdessamad Bellouch, Ahmed Boujnoui, Abdellah Zaaloul, Abdelkrim Haqiq, and Aboul Ella Hassanien NLP and Logic Reasoning for Fully Automating Test . . . . . . . . . . . . . . Nesrine Bnouni Rhim and Mouna Ben Mabrouk

88

98

A Comparative Study of Three LoRa Collision Resolution Schemes: A Markov Model-Based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 Abdellah Amzil, Abdessamad Bellouch, Ahmed Boujnoui, Mohamed Hanini, and Abdellah Zaaloul Production Scheduling Using Multi-objective Optimization and Cluster Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 Beatriz Flamia Azevedo, Maria Leonilde R. Varela, and Ana I. Pereira Data Prediction Model in Wireless Sensor Networks: A Machine Learning Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 Khushboo Jain, Manali Gupta, and Ajith Abraham A Study of Version Control System in Software Development Management Concerning PLC Environments . . . . . . . . . . . . . . . . . . . . 141 Domingos Costa, Senhorinha Teixeira, and Leonilde R. Varela REGION: Relevant Entropy Graph spatIO-temporal convolutional Network for Pedestrian Trajectory Prediction . . . . . . . . . . . . . . . . . . . . 150 Naiyao Wang, Yukun Wang, Changdong Zhou, Ajith Abraham, and Hongbo Liu An Analysis of Multipath TCP for Improving Network Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 Virendra Dani, Sneha Nagar, and Vishal Pawar Improving 3D Plankton Image Classification with C3D2 Architecture and Context Metadata . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 Nassima Benammar, Haithem Kahil, Anas Titah, Facundo M. Calcagno, Amna Abidi, and Mouna Ben Mabrouk Livestock Application: Naïve Bayes for Diseases Forecast in a Bovine Production Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 Aline Neto, Susana Nicola, Joaquim Moreira, and Bruno Fonte Development of a Workstation Assessment APP, Integrating Performance with the Worker’s Health and Well-Being . . . . . . . . . . . . 193 Pedro C. Ribeiro, Marlene F. Brito, and Ana L. Ramos Deep Learning for Big Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 Filipe Correia, Ana Madureira, and Jorge Bernardino

Contents

xiii

A Review on MOEA and Metaheuristics for Feature-Selection . . . . . . . 216 Duarte Coelho, Ana Madureira, Ivo Pereira, and Ramiro Gonçalves Detection and Classification of Age-Related Macular Degeneration Using Integration of DenseNet169 and Convolutional Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226 F. Ajesh and Ajith Abraham An Augmented Lagrangian Artificial Bee Colony with Deterministic Variable Selection for Constrained Optimization . . . 239 Marco Antônio Florenzano Mollinetti, Bernardo Bentes Gatto, and Otávio Noura Teixeira Remote Monitor System for Alzheimer Disease . . . . . . . . . . . . . . . . . . . 251 Luis B. Elvas, Daniel Cale, Joao C. Ferreira, and Ana Madureira Detection of Social Distance and Intimation System for Covid-19 . . . . . 261 S. Anandamurugan, M. Saravana Kumar, K. Nithin, and E. G. Prashanth Automatic Shoe Detection Using Image Processing . . . . . . . . . . . . . . . . 270 K. R. Prasanna Kumar, D. Pravin, N. Rokith Dhayal, and S. Sathya Recognition of Disparaging Phrases in Social Media . . . . . . . . . . . . . . . 278 K. R. Prasanna Kumar, P. Aswanth, A. Athithya, and T. Gopika Machine Learning Model for Identification of Covid-19 Future Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286 N. Anitha, C. Soundarajan, V. Swathi, and M. Tamilselvan Alzheimer’s Disease Detection Using Machine Learning and Deep Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296 K. Sentamilselvan, J. Swetha, M. Sujitha, and R. Vigasini Crime Factor Anaysis and Prediction Using Machine Learning . . . . . . . 307 N. Anitha, S. Gowtham, M. Kaarniha Shri, and T. Kalaiyarasi Detection of Fake Reviews on Online Products Using Machine Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314 H. Muthu Krishnan, J. Preetha, S. P. Shona, and A. Sivakami Deep Neural Network Model for Automatic Detection of Citrus Fruit and Leaf Disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320 S. Anandamurugan, B. Deva Dharshini, J. Ayesha Howla, and T. Ranjith Reducing Time Complexity of Fuzzy C Means Algorithm . . . . . . . . . . . 332 Amrita Bhattacherjee, Sugata Sanyal, and Ajith Abraham Information and Communication Technologies Kubernetes for Fog Computing - Limitations and Research Scope . . . . 351 R. Leena Sri and Divya Vetriveeran

xiv

Contents

Design and Simulation of 2.4 GHz Microstrip Parallel Coupled Line Low Pass Filter for Wireless Communication System . . . . . . . . . . 362 Shamsuddeen Yusuf, Shuaibu Musa Adam, Adamu Idris, David Afolabi, Vijayakumar Nanjappan, and Ka Lok Man A New Cascade-Hybrid Recommender System Approach for the Retail Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371 Miguel Ângelo Rebelo, Duarte Coelho, Ivo Pereira, and Fábio Fernandes A Novel Deep Neural Network Based Approach for Alzheimer’s Disease Classification Using Brain Magnetic Resonance Imaging (MRI) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 381 Ruhul Amin Hazarika, Debdatta Kandar, and Arnab Kumar Maji Classification of Cognitive Ability from Multichannel EEG Signals Using Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391 Nilima Salankar Heterogeneous DBSCAN for Emergency Call Management: A Case Study of COVID-19 Calls Based on Hospitals Distribution in Saudi Arabia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402 Naila Aziza Houacine, Lydia Sonia Bendimerad, and Habiba Drias A Survey on the Quality of Service and Metaheuristic Based Resolution Methods for Multi-cloud IoT Service Selection . . . . . . . . . . . 412 Ahmed Zebouchi and Youcef Aklouf Critical Success Factors for Information Technology and Operational Technology Convergence Within the Energy Sector . . . . . . . . . . . . . . . . 425 Thabani Dhlamini and Tendani Mawela A New Structured Model for ICT Competencies Assessment Through Data Warehousing Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435 Vladimir Dobrynin, Michele Mastroianni, and Olga Sheveleva Automated Evaluation Tools for Web and Mobile Accessibility: A Systematic Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447 João Dias, Diana Carvalho, Hugo Paredes, Paulo Martins, Tânia Rocha, and João Barroso My Buddy: A 3D Game for Children Based on Voice Commands . . . . . 457 Diana Carvalho, Tânia Rocha, and João Barroso Demography of Machine Learning Education Within the K12 . . . . . . . 467 Kehinde Aruleba, Oluwaseun Alexander Dada, Ibomoiye Domor Mienye, and George Obaido

Contents

xv

Educational Workflow Model for Effective and Quality Management of E-Learning Systems Design and Development: A Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475 Kingsley Okoye Designing Green Routing and Scheduling for Home Health Care . . . . . 491 Hossein Shokri Garjan, Alireza Abbaszadeh Molaei, Fariba Goodarzian, and Ajith Abraham The Bi-level Assembly Flow-Shop Scheduling Problem with Batching and Delivery with Capacity Constraint . . . . . . . . . . . . . . . . . . . . . . . . . 505 Hossein Shokri Garjan, Alireza Abbaszadeh Molaei, Nazanin Fozooni, and Ajith Abraham Building Trust with a Contact Tracing Application: A Blockchain Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517 Tomás Honório, Catarina I. Reis, Marco Oliveira, and Marisa Maximiano Immunity Passport Ledger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 527 Marco Oliveira, Tomás Honório, Catarina I. Reis, and Marisa Maximiano Computer Graphics Rendering Survey: From Rasterization and Ray Tracing to Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537 Houssam Halmaoui and Abdelkrim Haqiq A Sentiment-Based Approach to Predict Learners’ Perceptions Towards YouTube Educational Videos . . . . . . . . . . . . . . . . . . . . . . . . . 549 Rdouan Faizi ChatBots and Business Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557 Teresa Guarda and Maria Fernanda Augusto A Modified Feature Optimization Approach with Convolutional Neural Network for Apple Leaf Disease Detection . . . . . . . . . . . . . . . . . 567 Vagisha Sharma, Amandeep Verma, and Neelam Goel Ontology Based Knowledge Visualization for Domestic Violence Cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579 Tanaya Das, Abhishek Roy, and Arun Kumar Majumdar Application for the Management of Sports Performance in CrossFit Supported by an Artificial Intelligence Cognitive Service . . . . . . . . . . . . 590 J. Oliveira, S. Nicola, P. Graça, S. Martins, and T. Gafeira Perceptions of Cloud Computing Risks in the Public Sector . . . . . . . . . 599 Bonginkosi Mkhatshwa and Tendani Mawela

xvi

Contents

Mitigating Security Problems in Fog Computing System . . . . . . . . . . . . 612 Shruti and Shalli Rani A Detailed Review of Organizational Behavior of College Employees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623 V. I. Roy and K. A. Janardhanan Non-invasive Flexible Electromagnetic Sensor for Potassium Level Monitoring in Sweat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 633 Gianvito Mevoli, Claudio Maria Lamacchia, and Luciano Mescia A Study on Sequential Transactions Using Smart Card Based Cloud Voting System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643 Roneeta Purkayastha and Abhishek Roy Detecting Spinal Abnormalities Using Multilayer Perceptron Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654 Arju Manara Begum, M. Rubaiyat Hossain Mondal, Prajoy Podder, and Subrato Bharati Wireless Sensor Networks Time Synchronization Algorithms and Protocols Message Complexity Comparison: The Small-Size Star-Topology Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665 Giuseppe Coviello, Gianfranco Avitabile, and Antonello Florio Trust Management Model in IoT: A Comprehensive Survey . . . . . . . . . 675 Muhammad Saeed, Muhammad Aftab, Rashid Amin, and Deepika Koundal Sentiment Classification and Comparison of Covid-19 Tweets During the First Wave and the Second Wave Using NLP Techniques and Libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685 Sareeta Mugde, Garima Sharma, Aditya Singh Kashyap, and Swastika Swastik Automatic Modulation Recognition Models Based on Transfer Learning and Simulated Radio Signals in AWGN Channels . . . . . . . . . 700 Jamiu R. Olasina, Emmanuel Adetiba, Abdultaofeek Abayomi, Obiseye O. Obiyemi, Surendra Thakur, and Sibusiso Moyo Analysis of the Access to the Financing of the Ecuadorian Companies in the Framework of the Sanitary Emergency of COVID 19 and the Economic Sectors of Unemployment . . . . . . . . . . 713 Marcelo León, Carlos Redroban, Vinicio Loaiza, and Paulina León State of the Art of Wind and Power Prediction for Wind Farms . . . . . . 723 Ricardo Puga, José Baptista, José Boaventura, Judite Ferreira, and Ana Madureira

Contents

xvii

State of the Art on Advanced Control of Electric Energy Transformation to Hydrogen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 733 Ricardo Puga, José Boaventura, Judite Ferreira, and Ana Madureira Dynamic Modelling of a Thermal Solar Heating System . . . . . . . . . . . . 743 José Boaventura-Cunha and Judite Ferreira A Review of Unpredictable Renewable Energy Sources Through Electric Vehicles on Islands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 751 Juliana Chavez, João Soares, Zita Vale, Bruno Canizes, and Sérgio Ramos Fish Control Process and Traceability for Value Creation Using Blockchain Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 761 Joao C. Ferreira, Ana Lucia Martins, Ulpan Tokkozhina, and Berit Irene Helgheim Impact of Socio-Economic Factors on Students’ Academic Performance: A Case Study of Jawahar Navodaya Vidyalaya . . . . . . . . 774 Kapila Devi, Saroj Ratnoo, and Anu Bajaj Techno-Economic Feasibility Analysis and Optimal Design of Hybrid Renewable Energy Systems Coupled with Energy Storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 786 Shirin Cupples, Amir Abtahi, Ana Madureira, and José Quadrado Information Technology Roles and Their Most-Used Programming Languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 797 Oluwaseun Alexander Dada, Kehinde Aruleba, Abdullahi Abubakar Yunusa, Ismaila Temitayo Sanusi, and George Obaido Automated Fingerprint Biometric System for Crime Record Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 806 Muyideen AbdulRaheem, Sanjay Misra, Joseph Bamidele Awotunde, Idowu Dauda Oladipo, and Jonathan Oluranti Estimation Techniques for Scrum: A Qualitative Systematic Study . . . . 818 Diaz Jorge-Martinez, Sanjay Misra, Shariq Aziz Butt, and Foluso Ayeni Development of Students’ Results Help Desk System for First Tier Tertiary Institutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 830 Abraham Ayegba Alfa, Sanjay Misra, Blessing Iganya Attah, Kharimah Bimbola Ahmed, Jonathan Oluranti, Robertas Damaševičius, and Rytis Maskeliūnas

xviii

Contents

Performance Evaluation of Machine Learning Techniques for Prescription of Herbal Medicine for Obstetrics and Gynecology Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 842 Oluwasefunmi Arogundade, Adeniyi Akanni, Sanjay Misra, Temilade Opanuga, Oreoluwa Tinubu, Muhammad Akram, and Jonathan Oluranti Data Science for COVID-19 Vaccination Management . . . . . . . . . . . . . 852 Elham Rezaei, Kajal Ghoreyshi, and Kazi Masum Sadique Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 863

Bio-inspired Computing and Applications

BIC Algorithm for Heineken Brand Awareness in Vietnam Market Nguyen Thi Ngan

and Bui Huy Khoi(B)

Industrial University of Ho Chi Minh City, Ho Chi Minh City, Vietnam {nguyenthingan,buihuykhoi}@iuh.edu.vn

Abstract. The building to recognize the Heineken brand in the international market and the Vietnamese Market in particular, in the current period, has a lot of potentials, but still many potential challenges. The author has based on the theoretical basis got from previous studies and specific information from recent research papers to propose a research model comprising 5 independent factors: (1) Perceived quality, (2) Brand image, (3) Price, (4) Loyalty, (5) Brand association and a dependent variable is Brand awareness in Vietnam market. Analytical data was collected through an online survey with 196 samples. The analysis results show that 5 factors all have a positive influence on the brand awareness of Heineken in the Vietnam market. Perceived quality (PQ), Loyalty (LOY), Brand image (BI), Brand association (BA), and Price (PE) impact Brand awareness (BAW) is 28.8. BIC finds the optimal choice and five variables have a probability of 70.7%. From there, we give the necessary and management implications for each factor to help managers help businesses improve and improve efficiency and make customers remember the Heineken brand. This study uses the optimal choice of the BIC Deep Learning Algorithm for Heineken brand awareness in Vietnam Market. Keywords: BIC deep learning algorithm · Heineken · Brand · Awareness

1 Introduction Big Data Analytics and Deep Learning are too high-focus on data science. Big Data has become important as many organizations, both public and private, have been collecting massive amounts of Deep Learning Algorithm for Heineken brand awareness in Societal Issue [1]. With a young population structure and a growing middle class, Vietnamese people’s beer-drinking habits have risen steadily in recent years, and the country now has the highest beer consumption in Southeast Asia. Third in Asia, only after Japan and China, and in the world’s top 25 according to a report of the Association of Beer – Alcohol – Beverage Vietnam (VBA), the Vietnamese market consumed about 3.4 billion liters of beer in 2015 and this number will probably increase to around 4.4 billion liters in 2017 [1]. However, it is forecasted that the beer market in Vietnam will be saturated in the coming years and stabilize at around 5 billion liters per year. Brand rivalry in the Vietnamese beer industry is expected to become increasingly severe. Currently, besides domestic brands, most famous international beer brands have increasingly penetrated © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 3–14, 2022. https://doi.org/10.1007/978-3-030-96299-9_1

4

N. T. Ngan and B. H. Khoi

the Vietnamese market such as Beer Tiger, Biere Larue, Heineken, Sapporo, Budweiser, Carlberg. Among these brands, Heineken has the reputation for product quality and wide distribution systems along with impressive messages and especially with famous brands that always receive a lot of love and recognition with an important position in the minds of Vietnamese consumers. According to the development plan for 2020, Vietnam’s beer industry will reach an output of about 4.1 billion liters in 2025 to 4.6 billion liters and 5.5 billion liters in 2035, and the end of 2018 has reached over 4 billion liters [2]. Vietnam’s beer industry is maintaining a growth rate of over 5% per year while many other markets have been moving sideways or having negative growth in recent years. Vietnam is becoming an attractive beer market that makes domestic and international beverage businesses struggle. Winning in the beer industry largely depends on technical marketing secrets rather than finance and product formula. Enterprises capture market share with unique marketing strategies, from advertising, communication to promotions on all channels, such as on-site, off-site, modern on-site, modern off-site, and construction brand at the initial customer level. Heineken refers to the communication and promotion in the beer market. Many customers and consumers always love Heineken Vietnam Brewery’s many activities such as Heineken Formula 1, Tiger Street Football, Loyalty Trade, Refrigerator, and Banquet… Therefore, a brand, especially Heineken’s inherent brand components in the Vietnamese beer market, influences the purchasing behavior of Vietnamese consumers and subsequently contributes to the Vietnamese market. Nam is the third-largest in the Heineken market, following only Mexico and Nigeria [3]. Building a model to recognize corporate brands in the international market and the Vietnamese market in particular, in the current period, has quite a lot of potential, but there are still many potential challenges. Therefore, the paper has synthesized many models suitable for the Vietnamese market to increase the efficiency when researching a model of factors affecting the brand equity of the company [3], research model of factors affecting brand awareness according to Ozasomer and Altaras [4] and the factors include: (1) Perceived quality, brand equity, brand loyalty, brand association, (2) perceived quality, purchase decision, brand loyalty, brand experience, (3) brand loyalty, brand association, brand perception, brand exclusivity, (4) Brand perception, brand reputation, brand investment, brand association, price perception, (5) Brand satisfaction, brand image, brand loyalty, brand impact brand. The article uses the optimum selection by the BIC Deep Learning Algorithm for Heineken brand awareness in the Vietnam market.

2 Literature Review 2.1 Brand Awareness (BAW) Keller’s study [2] defines brand awareness as the customer’s ability to recall and recognize a brand as reflected because customers recognize the company’s brands under different conditions and can associate the brand name, logo, or symbol, in the customer’s memory. Brand awareness has several levels, starting from the lowest level, namely unrecognized brand and brand recognition, the flashback stage in the customer’s mind. Brand awareness signifies the existence, commitment, and core that are so important to

BIC Algorithm for Heineken Brand Awareness in Vietnam Market

5

a company. So, if brand knowledge is high, then the brand presence can always be felt. Several factors often cause a brand to achieve high brand awareness: constant advertising and association with the existence and distribution of products reaching different groups [3]. A well-managed brand can deliver customer satisfaction and customer value [4]. Brand awareness has a familiar key indicator [5], such as giving a sense of satisfaction and pride [6], easily receiving recognition [7], and can influence buyer decisions [3]. Brand awareness is a potential buyer’s capacity to recognize or recall that a brand is part of a particular product category. Chaney et al. [8] state that brand awareness is the ability of consumers to identify brands under different conditions, reflected in rebranding and performance recall. In short, “brand awareness is understood as the brand that delivers what they promised” for Heineken. They always care about all customers but do not forget about their obligations and duties. itself is always providing necessary and timely products and services to meet the needs of Heineken consumers, “the product claims of this brand are credible” Heineken consistently offers products that meet consumer standards according to regulations guaranteeing quality in each of their products, “this brand has a name you can trust, this brand doesn’t pretend to be something must be themselves” fair competition without conflicts and deception is the business standard of Heineken, which always puts the brand’s trust value first and considers this a standard measure in business. 2.2 Perceived Quality (PQ) According to Alfred [9], Perceived quality is demarcated as “the customer’s perception of the overall quality or superiority of a product/service for its intended purpose over alternatives”. As a result, perceived quality is the quality that is based on the consumer’s perception. It is not the product’s or production’s real quality that is productor production-oriented [10]. According to Garvin [10], brand value will depend on perceived quality because the product must be positively evaluated about the brand in the customer’s memory. Therefore, brand perception is a crucial part of brand equity. Perceived quality gives value to a brand in several ways: high perceived quality influences consumer choice, can support higher prices, and can lead to bigger profit margins and leads to enhanced brand equity [11]. The influence of perceived quality on brand equity has been confirmed by many previous research results [11]. Therefore, the following hypothesis about the relationship between perceived quality and brand equity is proposed. Perceived quality is the consumer’s perception of the overall quality or superiority of a product or service concerning its intended purpose [12]. Perceived quality is defined as the customer’s perception of the overall quality or superiority of a product or service for its intended purpose, relative alternatives [13]. According to Aaker [14], quality perception is the consumer’s perception of a product or service’s overall quality or superiority for the intended purpose. Delivering high quality, understanding consumer indications of quality, identifying crucial features of quality, and successfully communicating the message of quality are the keys to high-quality perception [14]. Investing in boosting a brand’s genuine objective quality is the best approach for it to improve perceived quality. Furthermore, the company must connect the quality of its brands through quality signs in its marketing activities. As a result, consumers judge brand quality based

6

N. T. Ngan and B. H. Khoi

on the company’s direct experience with the brand and data gathered from environmental elements [15, 16]. 2.3 Loyalty (LOY) According to Chaudhuri and Holbrook [17], Brand loyalty is the size of customer association or closeness to a brand. Torres et al. [12] define brand loyalty in terms of behavior, attitude, and opinion choice. The longitudinal perspective combines customer preferences and attitudes toward the brand, whereas the behavioral approach is based on the number of purchases for a certain brand. Brand loyalty is described as an organization’s long-term commitment to acquiring or repurposing a chosen product or service, thereby causing co-branding or repurchase activity, regardless of situational influences and marketing efforts that have the potential to induce customer behavior change [15]. Yoo and Donthu [11] defined brand loyalty as the tendency to remain loyal to a focal brand, as demonstrated by the intention to purchase that brand as a primary option. Brand loyalty is a measure of a customer’s association or closeness with a brand [12]. Javalgi and Moberg [18] defined brand loyalty in terms of behavior, opinion, and choice. While the behavioral perspective is based on the number of purchases for a particular brand, the fundamental view combines consumer preferences and inclinations towards the brand. Repeated purchases of the same or the same brand regardless of situational influences and marketing efforts are likely to induce switching behavior [15]. Yoo and Donthu [11] defined brand loyalty as the tendency to be loyal to a focal brand, as demonstrated by the intention to purchase the brand as the main choice. 2.4 Brand Image (BI) Brand image is often referred to as the psychological aspect of the image or impression that has been built in the subconscious of consumers through expectations and experiences when talking about the brand. through a product or service [19]. Therefore, forming a positive brand image is becoming more and more important for the company/business. According to Amelia [20], a good perception of product and service quality will promote consumers to form a positive brand image. Brand image has a direct impact on consumer-based purchasing behavior [21]. According to Maholtra [22], brand image is also known as a customer’s perception of reason or rationale or through more emotion towards a particular brand. A positive brand image will help the marketing program be liked and can create unique associations with the brand and help the brand stay alive in customer retention. Brand image is the consumer’s perception of the brand. The goal of working strategically with the brand image is to ensure that consumers hold strong and favorable brand associations in their minds. The brand image often includes many concepts: perceived, because the brand is perceived; awareness, because the brand is perceived as perceived; and finally attitude, because consumers continuously after feeling and evaluating what they perceive will form brand attitudes [2, 14]. Brand image is the crux of a consumer-based approach.

BIC Algorithm for Heineken Brand Awareness in Vietnam Market

7

2.5 Brand Association (BA) According to Boisvert and Burton [23], brand awareness was defined as a consumer’s capacity to recall that a brand belongs to a product category. When customers attach a favorable memory to a brand, the brand association is stronger [14]. Perceived brands influence consumer decision-making by influencing the degree of association between brands in mind [23]. The higher the consumer’s brand awareness, the more obvious the brand is [24]. Customers’ brand association is consequently likely when they have a strong level of brand awareness concerning other brands. Brand association is stronger when consumers positively associate them in their memory with the brand [14]. Brand awareness influences consumer decision-making by influencing the intensity of brand associations in their mind [23]. The higher the level of consumer brand awareness, the clearer the consumer brand association [24]. Therefore, consumers’ brand association is likely to become stronger when they have a high level of brand awareness. 2.6 Price (PE) Price is a significant factor in customer decision-making because it is indicative of a product’s quality on the market [25]. Dodds et al. [26] suggest that using price as a sign of product quality is not unreasonable but represents the belief that market prices are determined by the interplay of different factors of competitive demand and supply such forces would lead to a “natural” ordering of competing products on a price scale, leading to a strong, real positive relationship between price and product quality [26]. As a result, a greater price sends the message that the product is of much higher quality, leading to positive brand connotations among consumers. As a result, we propose that pricing moderates the relationship between brand awareness and brand association. Price indicates an amount of money that must be paid to achieve something [27]. Price refers to the amount charged for goods and services in return for the advantage of the product based on Kotler and Keller [28]. Kotler and Keller [28] also defined Price as the amount charged for a product or service, the sum of the values that customers exchange to obtain or use the product or service. To see if price strengthens the relationship between brand awareness and brand association, we included perceived price as a moderating variable in our research. Consumer decision-making is heavily influenced by price, as it is an important indicator of the quality of a product in the market [25, 26].

3 Methodology According to Tabachnick and Fidell [29] for the best regression analysis, it is necessary to ensure the sample size as follows: N ≥ 8 m + 50. In which: N: is the sample size, m: is the number of independent variables. Corresponding to the formula N >= 8 m + 50, the number of survey samples is 8 * 6 + 50 = 98 samples. On that basis, to meet the research process, the author chose a convenient sampling method to easily access the model, so the survey sample was distributed with several 196 samples in Ho Chi Minh City, Vietnam. Table 1 shows sample characteristics statistics.

8

N. T. Ngan and B. H. Khoi Table 1. Statistics of sample

Characteristics Sex and Age

Education

Job

Consuming

Amount

Percent (%)

Male

76

38.8

Female

120

61.2

17–25

59

30.1

26–35

76

38.8

36–45

43

21.9

46–55

18

9.2

High school diploma

64

32.7

Diploma

68

34.7

Degree

64

32.7

Other

64

32.7

Student

38

19.4

Officer

85

43.4

Freelance

45

23.0

Business executives

28

14.3

Everyday

87

44.4

Every week

66

33.7

Every month

43

21.9

The above Table 1 shows that in the total number of samples sent, the author received the highest percentage of male sex with 61.2%, followed by the female with a lower rate of 38.8%. The age of using Heineken beer from 26 to 35 years old accounted for the highest rate with 38.8%, and then the age group from 17 to 25 accounted for 30.1%, the age from 36 to 45 accounted for 21.9%, and finally the age group from 46 to 55 accounts for 9.2%. Through the above data, it shows that the level of preference and use of Heineken beer among men and between the ages of 26 and 35 has the most interest and use. The group of people surveyed by education level also showed interest in using Heineken beer: 32.7% for high school diploma, 34.7% diploma. 32.7% degree. Thus, it can be seen that the educational level factor does not affect the difference in using Heineken beer today. All users like to consume Heineken Brand beer. All study staff and respondents were blinded by the length of the trial. The study participants had no touch with anyone from the outside world

4 Results 4.1 Reliability Test Cronbach’s Alpha test is a tool to help the author check whether the observed variables of the crucial factor are reliable or not, and whether or not the variable is good. This

BIC Algorithm for Heineken Brand Awareness in Vietnam Market

9

test reflects whether the criteria for compatibility and concordance among dependent variables in the same major factor are closely related. The higher the coefficient of Cronbach’s Alpha (α), the higher the reliability of the factor. Cronbach’s Alpha value coefficient includes the following values: 0.8 to 1: very good scale, 0.7 to 0.8: good use scale, 0.6 and above: qualified scale. If a measure has a Corrected item-total Correlation (CITC) greater than 0.3, then that variable meets the requirements [30]. Table 2 shows the Cronbach’s Alpha coefficient of Brand awareness (BAW), Perceived quality (PQ), Loyalty (LOY), Brand image (BI), Brand association (BA), and Price (PE) for the Heineken beer brand is all greater than 0.7. This shows that the factors are all reliable. Table 2 shows that all Corrected item-total Correlation of items is greater than 0.3. That shows that the items are all correlated in the factor and they contribute to the correct assessment of the concept and properties of each factor. Therefore, in testing the reliability of Cronbach’s Alpha for each scale, the author found that all the observed variables satisfy the set conditions that the Cronbach’s Alpha coefficient is greater than 0.6 and the Corrected item coefficient – Total Correlation is greater than 0.3, so all items are used for the next test step. 4.2 BIC Deep Learning Algorithm There have been numerous algorithms created and extensively explored for detecting association rules in transaction databases. Other mining algorithms, such as incremental updating, mining of generalized and multilevel rules, mining of quantitative rules, mining of multi-dimensional rules, constraint-based rule mining, mining with multiple minimum supports, mining associations among correlated or infrequent items, and mining of temporal associations, were also presented to provide more mining capabilities [31]. Big Data Analytics and Deep Learning are two areas of data science that are receiving considerable interest. As many individuals and organizations have been gathering enormous amounts of Deep Learning Algorithm for Heineken brand awareness in Societal Issues, Big Data has become increasingly important [1]. BIC (Bayesian Information Criteria) was used to select the best model by R software. In the theoretical environment, BIC has been used to choose models. BIC can be used as a regression model to estimate one or more dependent variables from one or more independent variables [32]. The BIC is an important and useful metric for determining a full and straightforward model. A model with a lower BIC is chosen based on the BIC information standard [32–34]. R report shows every step of searching for the optimal model. BIC selects the best 4 models as Table 3. There are five independent and one dependent variable. Perceived quality (PQ), loyalty (LOY), and Price (PE) influence Brand awareness (BAW) with a probability of 100%. Brand image (BI) and Brand association (BA) have a probability of 93.3% and 92.2%.

10

N. T. Ngan and B. H. Khoi Table 2. Reliability

Factor

α

Item

Code

CITC

Brand awareness (BAW)

0.764

The brand brings what it promised

BAW1

0.719

The brand’s product privileges are trustworthy

BAW2

0.617

The brand has a name you can believe

BAW3

0.717

The brand doesn’t invent to be something that isn’t themselves

BAW4

0.367

Perception of the innovative quality of Heineken beer is an important factor to attract customers

PQ1

0.773

The innovation of Heineken beer brings the brand closer to customers

PQ2

0.834

Heineken features positively impact the brand

PQ3

0.574

Heineken’s promotion is very appreciative

PQ4

0.746

I am loyal to the frequency of appearance of the brand

LOY1

0.435

I usually use this brand as my first choice over another brand

LOY2

0.528

I would recommend the Heineken beer brand to everyone

LOY3

0.470

Heineken does what it promises

BI1

0.651

Heineken raises awareness that I have a desirable lifestyle

BI2

0.597

Heineken makes me feel good

BI3

0.690

Using Heineken suits my needs better than other brands

BI4

0.474

The brand has character

BA1

0.913

The brand is exciting

BA2

0.807

I have a clear picture in my mind of who would use the brand

BA3

0.722

I believe the company that created the brand

BA4

0.732

Reasonably priced Heineken beer

PE1

0.809

Perceived quality (PQ)

Loyalty (LOY)

Brand image (BI)

Brand association (BA)

Price (PE)

0.868

0.662

0.788

0.901

0.852

(continued)

BIC Algorithm for Heineken Brand Awareness in Vietnam Market

11

Table 2. (continued) α

Factor

Item

   2 σ (xi ) k 1− α = k−1 σx2

Code

CITC

The price of Heineken is equal to the PE2 benefit I get

0.619

Heineken offers a flexible seasonal pricing strategy

PE3

0.744

Heineken’s price is equal to the quality I get

PE4

0.660

Table 3. BIC model selection BAW

Probability (%)

SD

model 1

model 2

model 3

model 4

Intercept

100.0

0.38302

−0.06736

0.18297

0.16789

−0.25963

PQ

100.0

0.08952

0.33361

0.40763

0.36474

0.27647

BI

93.3

0.08444

−0.21362

−0.20101

−0.19117 0.25462

0.20741

0.31523

0.27042

PE

100.0

0.10765

0.22312

LOY

100.0

0.07698

0.31704

0.33000

92.2

0.07598

0.18297

0.20528

BA

0.16315

4.3 Model Evaluation

Table 4. Model test Model

nVar

R2

BIC

post prob

model 1

5

0.288

−40.23299

0.707

model 2

4

0.257

−37.10588

0.148

model 3

4

0.252

−35.81494

0.078

model 4

4

0.251

−35.52688

0.067

BIC = −2 * LL + log(N) * k

According to the results from Table 4, BIC shows model 1 is the optimal selection because BIC (−40.23299) is minimum. Perceived quality (PQ), Loyalty (LOY), Brand image (BI), Brand association (BA), and Price (PE) impact Brand awareness (BAW) is 28.8% in Table 4. BIC finds model 1 is the optimal choice and five variables have a probability of 70.7%. The above analysis shows the regression equation below is statistically significant.

12

N. T. Ngan and B. H. Khoi

BAW = −0.06736 + 0.33361PQ − 0.21362BI + 0.22312PE + 0.31704LOY + 0.18297BA

Perceived quality (PQ), loyalty (LOY), Brand association (BA), and Price (PE) positively influence Brand awareness (BAW). Brand image (BI) negatively affects Brand awareness (BAW).

5 Conclusions This study uses the optimal choice by BIC Deep Learning Algorithm for Heineken brand awareness in Societal Issue. According to statistics from the survey sample, 61.2% are male and the remaining 38.8% are female. With these 193 suitable survey samples, the author went into reliability testing, linear regression analysis. From the regression results, the author believes that perceived quality is the factor that has the strongest and most positive impact on the brand awareness of Heineken beer in the Vietnam market with the index β = 0.33361 followed by the loyalty with β = 0.31704, the price with the β index = 0.22312, then the brand image with the β index = −0.21362, and finally the brand association factor with a low β index at most 0.18297. Limitations Although the research results have certain academic and practical contributions to the field of research on brand identity and brand development in the international market in general and Heineken beer brand identity in the Vietnamese market, South in particular. However, there are still some limitations to this study. Firstly, in the world today, the brand has become a specific symbol for each business that intends to build a foothold and create a position in the market. Managers are always interested in details, image attributes, icons, etc. All external factors can affect more or less the process of building and developing a strong and unique brand. This research paper also comes from the ideas of creating a strong brand and increasing brand awareness among customers in one or more different markets. This study has measured most of the factors that can affect the recognition of a brand in the Vietnamese market in recent years and specifically this is the Heineken beer brand in the Vietnamese market. However, these are only factors affecting Vietnam but may not be comprehensive when used for the international market. This may be the next research direction to expand to neighboring areas. Second, for the scope of the study, this study is considered in terms of the Vietnamese market. However, it is still not possible to apply the details and specifics from the specific region in Vietnam such as big cities like Hanoi or Ho Chi Minh City, Vietnam to get a more reliable result. This limitation comes from objective reasons such as funding, time, and the author’s ability. If these factors are specifically surveyed in all areas in Vietnam, the research will be more effective and will be much more useful for managers to carry out the process of brand development and marketing to bring brands closer to consumers. Third, this study was conducted in terms of surveying all sources of customers from all ages and classes in today’s society. From students to the officer, freelance, or businessmen,… In general, the research has comprehensively surveyed most of the departments that need to use Heineken beer but still have not been able to use Heineken beer to confirm its accuracy and capability, this is also the direction for us to continue to expand

BIC Algorithm for Heineken Brand Awareness in Vietnam Market

13

the research to receive a more relevant and effective result. Finally, in terms of research methods, this study has used most of the required tests to analyze the received results, tested the theoretical model by linear regression, but has not used it yet. Other methods are more accurate such as the linear structural model (SEM), this model is both to test the hypothesis and to better define the cause-and-effect relationship between the research concepts. All the limitations mentioned above will open many directions for further research when carried out in the future, especially for researchers in the field of branding in a dynamically changing market such as Vietnam.

References 1. Najafabadi, M.M., Villanustre, F., Khoshgoftaar, T.M., Seliya, N., Wald, R., Muharemagic, E.: Deep learning applications and challenges in big data analytics. J. Big Data 2(1), 1–21 (2015). https://doi.org/10.1186/s40537-014-0007-7 2. Keller, K.L., Apéria, T., Georgson, M.: Strategic Brand Management: A European Perspective. Pearson Education (2008) 3. Mashur, R., Gunawan, B.I., Fitriany, F., Ashoer, M., Hidayat, M., Aditya, H.P.K.P.: Moving from traditional to society 5.0: case study by online transportation business. J. Distrib. Sci. 17(9), 93–102 (2019) 4. Macdonald, E.K., Sharp, B.M.: Brand awareness effects on consumer decision making for a common, repeat purchase product: a replication. J. Bus. Res. 48(1), 5–15 (2000) 5. Ha, H.Y., Perks, H.: Effects of consumer perceptions of brand experience on the web: brand familiarity, satisfaction and brand trust. J. Consum. Behav. Int. Res. Rev. 4(6), 438–452 (2005) 6. Biel, A., Aaker, D.: Brand equity and advertising. Laurence Erlbaum Associates (1993) 7. Balmer, J.M.: Corporate branding and connoisseurship. J. Gen. Manag. 21(1), 24–46 (1995) 8. Chaney, I., Hosany, S., Wu, M.-S.S., Chen, C.-H.S., Nguyen, B.: Size does matter: effects of in-game advertising stimuli on brand recall and brand recognition. Comput. Hum. Behav. 86, 311–318 (2018) 9. Alfred, O.: Influences of price and quality on consumer purchase of mobile phone in the Kumasi Metropolis in Ghana a comparative study. Eur. J. Bus. Manag. 5(1), 179–198 (2013) 10. Garvin, J.: Managing with total quality management-theory and practice. Int. J. Manpow. (1998) 11. Yoo, B., Donthu, N.: Developing and validating a multidimensional consumer-based brand equity scale. J. Bus. Res. 52(1), 1–14 (2001) 12. Torres, P.M., Augusto, M.G., Lisboa, J.V.: Determining the causal relationships that affect consumer-based brand equity: the mediating effect of brand loyalty. Marketing Intelligence and Planning (2015) 13. Kirmani, A., Zeithaml, V.: Advertising, perceived quality, and brand image. Brand equity and advertising: Advertising’s role in building strong brands, pp. 143–161 (1993) 14. Aaker, D.A.: Building Strong Brands The Free Press, pp. 598–614. New York (1996) 15. Gil, R.B., Andres, E.F., Salinas, E.M.: Family as a source of consumer-based brand equity. J. Prod. Brand Manag. (2007) 16. Sugiyarti, G., Mardiyono, A.: The role of brand equity in increasing buying interest. Manag. Sci. Lett. 11(7), 1999–2010 (2021) 17. Chaudhuri, A., Holbrook, M.B.: The chain of effects from brand trust and brand affect to brand performance: the role of brand loyalty. J. Mark. 65(2), 81–93 (2001) 18. Javalgi, R.R.G., Moberg, C.R.: Service loyalty: implications for service providers. J. Serv. Mark. (1997)

14

N. T. Ngan and B. H. Khoi

19. Bakri, M., Krisjanous, J., Richard, J.E.: Decoding service brand image through user-generated images. J. Serv. Mark. (2020) 20. Amelia, S.: The Effect of perceived quality, brand awareness, and brand loyalty toward brand equity of Beer Bintang in Surabaya. CALYPTRA 7(1), 899–918 (2018) 21. Nasar, A., Hussani, S.K., Karim, E., Siddiqui, M.Q.: Analysis of influential factors on consumer buying behavior of youngster towards branded products: evidence from Karachi. KASBIT Bus. J. 5(1), 56–61 (2012) 22. Malhotra, N.K.: Essentials of Marketing Research: A Hands-on Orientation. Pearson Essex (2015) 23. Boisvert, J., Burton, S.: Towards a better understanding of factors affecting transfer of brand associations. J. Consum. Mark. (2011) 24. Homburg, C., Klarmann, M., Schmitt, J.: Brand awareness in business markets: when is it related to firm performance? Int. J. Res. Mark. 27(3), 201–212 (2010) 25. Zeithaml, V.A.: Consumer perceptions of price, quality, and value: a means-end model and synthesis of evidence. J. Mark. 52(3), 2–22 (1988) 26. Dodds, W.B., Monroe, K.B., Grewal, D.: Effects of price, brand, and store information on buyers’ product evaluations. J. Mark. Res. 28(3), 307–319 (1991) 27. Abrate, G., Quinton, S., Pera, R.: The relationship between price paid and hotel review ratings: expectancy-disconfirmation or placebo effect?. Tour. Manag. 85, 104314 (2021) 28. Kotler, P., Keller, K.L.: Marketing Management 12e. Edition Pearson Education, France (2006) 29. Tabachnick, B., Fidell, L.: Using Multivariate Statistics, 4th edn., pp. 139–179. HarperCollins, New York (2001) 30. Nunnally, J.C.: Psychometric Theory 3E. Tata McGraw-Hill Education (1994) 31. Gharib, T.F., Nassar, H., Taha, M., Abraham, A.: An efficient algorithm for incremental mining of temporal association rules. Data Knowl. Eng. 69(8), 800–815 (2010) 32. Raftery, A.E., Madigan, D., Hoeting, J.A.: Bayesian model averaging for linear regression models. J. Am. Stat. Assoc. 92(437), 179–191 (1997) 33. Kaplan, D.: On the quantification of model uncertainty: a Bayesian perspective. Psychometrika 86(1), 215–238 (2021). https://doi.org/10.1007/s11336-021-09754-5 34. Raftery, A.E.: Bayesian model selection in social research. Sociol. Methodol. 111–163 (1995)

Comparison of Ant Colony Optimization Algorithms for Small-Sized Travelling Salesman Problems Arcsuta Subaskaran, Marc Krähemann, Thomas Hanne(B)

, and Rolf Dornberger

University of Applied Sciences and Arts Northwestern Switzerland, Basel, Muttenz, Olten, Switzerland [email protected]

Abstract. This paper deals with Ant Colony Optimization (ACO) applied to the Travelling Salesman Problem (TSP). TSP is a well-known combinatorial problem which aim is to find the shortest path between a designated set of nodes. ACO is an algorithm inspired by the natural behavior of ants. When travelling from the nest to a food source, ants leave pheromones behind. This algorithm can be applied to TSP in order to find the shortest path. In this paper, the variants of ACO are shortly explained and a new Improved Ant Colony Optimization (IACO) algorithm is proposed. The IACO is applied to small-sized TSP. It is shown that the proposed IACO performs better in some cases, especially when there are more cities in the TSP. Keywords: Ant Colony Optimization · Travelling Salesman Problem · Improved Ant Colony Optimization · Comparison · Metaheuristics

1 Introduction The Travelling Salesman Problem (TSP) is a combinatorial NP-hard problem. The aim of the TSP is to find the shortest path between a set of nodes by visiting each node, except the starting node, exactly once [1]. There are many approaches to solve the problem, including metaheuristics such as Swarm Intelligence or Genetic Algorithms. The aim of these methods is to get as close to an optimal solution (thus the shortest route) as possible. One method which is often used in solving the TSP is Ant Colony Optimization (ACO), which makes use of the natural behavior of ants [2]. ACO originated from the Ant System (AS). Nowadays, ACO is hybridized with other algorithms to provide even better results. One of the main problems with the ACO algorithm is that it often gets stuck in a local minimum. Given this problem, a lot of Improved Ant Colony Optimization (IACO) algorithms were developed and proposed in the literature [3]. In this paper, a new IACO approach is suggested. We investigate how it solves the TSP with different parameter settings. The results are then compared with a classic ACO.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 15–23, 2022. https://doi.org/10.1007/978-3-030-96299-9_2

16

A. Subaskaran et al.

2 Research Problem and Method The research problem in this paper is to solve the TSP with ACO. There are many variants of ACO. This paper focuses on one IACO which was introduced in [3]. The algorithm was adapted, but the basic principle behind it remained the same. It makes use of an adjustable pheromone evaporation rate. In this paper, the focus is set on small to mid-sized TSP with n ≤ 30. Two main aspects are investigated. In a first step, the parameters of the IACO are changed and adjusted to find the best settings. In a second step, the same TSP is solved with the classic ACO to compare the results. To conduct the analysis, the software NetLogo [4, 5] was used. Existing code for ACO was used and adapted for the IACO. NetLogo is an open-source programmable modelling environment, which is used by many researchers. The set of cities was created with the aim of representing distances of Swiss cities on a smaller scale. The investigated TSPs are all symmetric. In Sect. 3, a literature review is provided on the TSP (Sect. 3.1) and ACO (Sect. 3.2). Sections 3.3 and 3.4 introduce the IACO used as a basis and the adapted version used in this paper. In Sect. 4, the results or conducted computation experiments including comparisons of ACO and IACO are shown and discussed. The paper ends with the conclusions in Sect. 5.

3 Literature Review and Suggested Method 3.1 Travelling Salesman Problem The aim of the TSP is to find the shortest path between designated cities (nodes). A salesman needs to visit each city exactly once, except for the starting city, to which he returns in the end. The TSP can be scaled with different numbers of cities. Therefore, it can be divided into different categories [6] such as small, medium, or large scale problems. The more cities are depicted in the problem, the more difficult it is to find an optimal solution to the problem. The following formula describes the TSP [7]: min

n  n 

cij xij

i=1 j=1

(1)

subject to n 

xij = 1

j = 1, 2, . . . n, i = j

(2)

xij = 1

i = 1, 2, . . . n, i = j

(3)

i=1 n  j=1

ui − uj + nxij ≤ n − 1 xij ∈ {0, 1}

i, j = 2, 3, . . . n, i = j i, j = 1, 2, . . . n, i = j

(4) (5)

Equation (1) shows that the TSP is a minimization problem, i.e., we want to find the shortest route. The aim is to optimize the route so that the cost is as low as possible.

Comparison of Ant Colony Optimization Algorithms

17

The restrictions (2)–(5) help assure that every city is visited exactly once. When a city is visited, the cost of the respective connection is multiplied with 1, otherwise with 0. At the end, the costs of the selected connections are added to determine the total cost of the route. 3.2 Ant Colony Optimization The idea of the AS comes from the natural behavior of ants. When ants are looking for food, they manage to find the shortest path between the nest and a food source using some concepts from Swarm Intelligence. At the beginning, ants choose different ways with the same probability. When going to a food source, ants leave pheromones behind, which evaporate over time. They use pheromones as a communication channel. The shorter a path, the higher the pheromone concentration at the edges. Therefore, the shortest path will be chosen with a higher probability over time. This eventually leads to finding the shortest path [8]. The first algorithm considering ant colonies was introduced by Dorigo in the 1990s [9]. There were three different versions of AS: 1. Ant-Density, 2. Ant-Quantity and 3. Ant-Cycle (AC). The AC is the algorithm that has been proven to provide the best results [10]. The principle of AC is that ants travel from city to city and leave pheromones behind. The probability to move to another city is stated by the transition probability. At the end of each iteration, the pheromone is updated on each path to favor a previously taken good path, but not too much to avoid getting stuck in local optima [9]. Also, many ACO algorithms are based on AC [11]. The AC algorithm was further developed, and the ACS was introduced [12]. The principle stays the same, but there are some important differences to the AS algorithm. The transition probability was changed into the state transition rule to favor movements to referable cities. The global updating rule has been adapted so that only the best ant can update the pheromone. Finally, a local updating rule was introduced, where ants update the pheromone level directly on the edge they visit [11, 12]. The example of the ACS shows that there is a lot of possible adjustments that can be made to the original algorithm of AS. All algorithms which make use of the natural behavior of ants are summarized under the term ACO. The ACO that is used in this paper consists of the parts shown in the following paragraphs. First, let the ant choose which city it wants to move to. The following equation shows the transition probability pij for the city i to the city j of an ant k: ⎧ ⎨  [τij (t)]α ×[ηij ]β β if j ∈ allowedk α k k∈allowedk [τik (t)] ×[ηik ] pij = (6) ⎩ 0, otherwise where ηij =

1 Lij

τij (t) is the pheromone intensity at edge (i,j), while ηij represents a function that takes the distance between i and j into account. L ij stands for the distance between the cities i and j and thus represents the quality of an edge. α is a parameter for the weighting of the pheromone. β is a parameter for the weighting of the distance between two cities.

18

A. Subaskaran et al.

Those parameters can be adjusted to change the importance of the pheromone and the distance, respectively. allowed k stands for the cities that an ant is still allowed to visit. It can be represented by {N-tabuk }, while N stands for the set of cities. Tabuk represents the tabu list for an ant k, thus the cities that have already been visited. The condition k ∈ allowedk makes sure that the ant is allowed to visit the next chosen city. The other function which is important in the ACO applied in this paper is the pheromone updating rule. It is summarized by the following equation: τijk ← (1 − ρ) × τij +  where τijk

=

m 

τijk

(7)

k=1 1 Lk

if kth ant uses edge (i, j) 0, otherwise

The parameter ρ which is introduced in (7), stands for the evaporation rate. It is a value between 0 and 1, while 0 means no evaporation at all and 1 causes the whole pheromone at an edge to vanish. L k represents the tour length of an ant k. As soon as an iteration is finished the pheromone is updated according to the rule in (7). This ensures that even edges with smaller pheromone mutation have a chance of being further explored by the ants. 3.3 Improved Ant Colony Optimization There are many modifications which can be made to the algorithms. Some examples are the Elitist Ant System [13] or the Max-Min Ant System [14]. As the TSP is an NP-hard problem, researchers are interested in finding new methods to get to an optimal solution faster. This has led to the creation of many Improved Ant Colony Optimization (IACO) algorithms. A common problem which is encountered with ACO is that it easily gets stuck in a local optimum [15]. There are many approaches to solving this, e.g. by using pheromone mutation or by combining ACO with other algorithms [16]. Many efforts have been made by researchers to solve this problem to find the optimal solution for a TSP. One IACO solution was proposed in [3]. One of the improvements made to the initial AS algorithm is the calculation of the parameter ρ: ρij (t) = 1 if t ≤ tearly  ρij (t) =

k1 , τij (t) ≤ C if t > tearly k2 , else

(8)

Equation (8) shows that, when the pheromone intensity is lower at a point t than at an earlier point t early , then ρ is set to 1, otherwise it is set either to value k 1 or k 2 . These two values lie between 0 and 1, while k 1 > k 2 and the distance between them is not too large. C is a value which lies between the average and maximum value of τ ij (t). The basic idea behind this algorithm is that ρ is not just a constant, but an adjustable parameter. This algorithm tries to prevent the solution from being stuck in a local minimum. The edges with lower pheromone intensity are given a smaller evaporation rate

Comparison of Ant Colony Optimization Algorithms

19

and thus they remain attractive for some ants. In this way, they keep exploring other paths which may lead to a better solution [3]. 3.4 Our Improved Ant Colony Optimization The basic idea of the IACO in [3], was adapted for this paper to create a new IACO. The algorithm is also based on adjustable pheromone evaporation, depending on the amount of pheromone left behind. For that reason, three new terms are introduced: 1) Pheromone threshold, 2) ρ low and 3) ρ high . The pheromone update strategy consists of the following steps: Step 1. Set the pheromone threshold to a number between 0 and 1. This value is used to determine whether the pheromone on a path is considered low or high. Step 2. Set the evaporation parameters ρlow and ρhigh to a value between 0 and 1, while ρ low < ρ high . Step 3. At the end of an iteration, the amount of each pheromone is evaluated with regard to the pheromone threshold. If τ ij < pheromone threshold, then the evaporation rate is set to ρ low. Otherwise, the evaporation rate is set to ρ high as shown in Eq. (9). Step 4. The pheromone is updated as shown in Eq. (7), by replacing ρ with either the high or the low evaporation rate.  ρlow τij (t) ≤ C (9) ρij (t) = ρhigh else Our IACO attributes edges with a higher pheromone intensity, a higher evaporation rate. On the other hand, lower pheromone intensity at an edge results in a lower pheromone evaporation rate. Therefore, this algorithm ensures that the current best path is not given too much preference. Thus, it helps that the ACO does not get stuck too easily in a local minimum. It also allows for more variation and the ants also explore paths that were initially considered less favorable.

4 Results To solve the TSP with IACO, the parameters pheromone threshold and the ant colony size are varied to find the best result for different TSP sizes (number of cities). The TSP sizes are categorized as small, medium, and large. In this paper, only small TSP problems are discussed. Hence, the average tour length is calculated from ten repetitive runs using the same parameters. The best tour length is also obtained from these ten runs. Three different sizes of the ant colony and three different pheromone thresholds are studied to find out how much they influence the results. The ant colony size that generated the best results with the shortest tour length and the shortest average tour length for various TSP sizes is indicated in bold in Tables 1, 2, 3 and 4. The parameters of the IACO used are defined as: • C is the value of pheromone threshold, which serves to set either ρ low or ρ high .

20

A. Subaskaran et al.

• ρ low is set when the pheromone value does not reach the pheromone threshold. • ρ high is set when the pheromone value exceeds the pheromone threshold. The remaining parameters have been set with the following values which are typical values from the literature: • ρ low = 0.25 • ρ high = 0.75. 4.1 Results of IACO with Seven Cities From the result of a TSP with seven cities (shown in Table 1), different pheromone thresholds and ant colony sizes show the same best tour length. This is probably because the TSP size is too small, so the best result can be easily obtained. Based on the average of the individual evaluations, it turns out that better results are achieved with the pheromone threshold of 0.1 and an ant colony size of 100. Each testcase is run ten times. Table 1. IACO results for the seven cities problem with a pheromone threshold of 0.1, 0.01, and 0.001. Ant colony size

25

50

100

Average tour length with threshold of 0.1

557

555

549

Best tour length with threshold of 0.1

546

546

546

Average tour length with threshold of 0.01

560

554

548

Best tour length with threshold of 0.01

546

546

546

Average tour length with threshold of 0.001

546

546

549

Best tour length with threshold of 0.001

546

546

546

4.2 Results of IACO with 16 Cities From the result with 16 cities (shown in Table 2), different pheromone thresholds and ant colony sizes show the same best tour length. This is probably still because the TSP size is too small. Using the average of the individual evaluations, it turns out that better results are achieved with the pheromone threshold of 0.01 and an ant colony size of 100. On average, an ant colony size of 100 is better than one of 25 or 50. 4.3 Results of IACO with 26 Cities From the result with 26 cities (shown in Table 3), different pheromone threshold and ant colony size do not show the same best tour length. This TSP size is big enough not to find the best result easily. Compared to the number of colonies, this size of TSP shows that a low average does not automatically lead to the best tour length. But for TSP with size seven and 16 cities, this is the case with one exception. The best tour was achieved with an ant colony size of 50 and with a threshold of 0.01.

Comparison of Ant Colony Optimization Algorithms

21

Table 2. IACO results for the 16 cities problem with a pheromone threshold of 0.1, 0.01 and 0.001. Ant colony size

25

50

100

Average tour length with threshold of 0.1

1007

983

973

Best tour length with threshold of 0.1

893

893

893

Average tour length with threshold of 0.01

991

973

945

Best tour length with threshold of 0.01

893

893

893

Average tour length with threshold of 0.001

1005

985

977

Best tour length with threshold of 0.001

893

893

893

Table 3. IACO results for the 26 cities problem with a pheromone threshold of 0.1, 0.01 and 0.001. Ant colony size

25

50

100

Average tour length with threshold of 0.1

1570

1432

1388

Best tour length with threshold of 0.1

1297

1195

1196

Average tour length with threshold of 0.01

1553

1424

1362

Best tour length with threshold of 0.01

1292

1124

1196

Average tour length with threshold of 0.001

1563

1434

1368

Best tour length with threshold of 0.001

1271

1202

1196

4.4 Comparison of ACO and IACO To compare the IACO algorithm with the ACO algorithm, we evaluated the best values of the IACO from Tables 1, 2 and 3 with the same initial situation and ten runs for each of the three different TSP sizes. Table 4. Comparison of ACO and IACO. Number of cities

7

16

26

ACO

546

893

1212

ACO average

549

956

1431

IACO

546

893

1196

IACO average

549

945

1362

Referring to Table 4 and Fig. 1, the result of the IACO best tour length is with 26 cities better than the ACO. It is also clear from the average of the tour length that IACO is not only with 26 cities better, but also with 16 cities. The evaluation with the number

22

A. Subaskaran et al.

of seven cities and 16 cities clearly shows that the IACO and ACO are equally good for this small size of TSP.

Fig. 1. Comparison of ACO and IACO. For different problem sizes (number of cities), the resulting tour lengths are shown.

5 Conclusions The results show that IACO performs better than ACO from a certain TSP size. However, the evaluation comparison based on the best results obtained in IACO. In the evaluation of IACO, the same size of ant colonies and pheromone thresholds were used for comparability. The pheromone thresholds were tested and selected, but there was a striking difference only for the size of 26 cities. For future work, we suggest investigating a larger number of TSP problems. As the impact seems to get larger with a higher number of cities, it would be interesting to evaluate the suggested algorithm for medium- and large-scale TSP problems. It would also be useful to investigate the influences of the ant colony size and the evaporation rate in more details. The largest ant colony size does not always give the best result. This, of course, would have to be tested again in combination with pheromone threshold and ant colony size. Also, the number of runs have to be increased depending on the size to achieve reliable and better results.

References 1. Liu, J., Li, W.: Greedy permuting method for genetic algorithm on traveling salesman problem. In: Proceedings of the 2018 8th International Conference on Electronics Information and Emergency Communication (ICEIEC), pp. 47–51. IEEE, Piscataway (2018). https://doi.org/ 10.1109/ICEIEC.2018.8473531 2. Dewantoro, R.W., Sihombing, P., Sutarman: The combination of ant colony optimization (ACO) and tabu search (TS) algorithm to solve the traveling salesman problem (TSP). In: 2019 3rd International Conference on Electrical, Telecommunication and Computer Engineering (ELTICOM), pp. 160–164. IEEE, Piscataway (2019). https://doi.org/10.1109/ELTICOM47 379.2019.8943832

Comparison of Ant Colony Optimization Algorithms

23

3. Zhang, J., Liu, H., Tong, S., Wang, L.: The improvement of ant colony algorithm and its application to TSP problem. In: 2009 5th International Conference on Wireless Communications, Networking and Mobile Computing, pp. 1–4. IEEE, Piscataway (2009). https://doi. org/10.1109/WICOM.2009.5301753 4. Tisue, S., Wilensky, U.: Netlogo: A simple environment for modeling complexity. In: International Conference on Complex Systems, vol. 21, pp. 16–21 (2004) 5. Wilensky, U., Rand, W.: An Introduction to Agent-Based Modeling: Modeling Natural, Social, and Engineered Complex Systems with NetLogo. MIT Press, Cambridge (2015) 6. Shetty, A., Shetty, A., Puthusseri, K.S., Shankaramani, R.: An improved ant colony optimization algorithm: Minion Ant (MAnt) and its application on TSP. In: 2018 IEEE Symposium Series on Computational Intelligence (SSCI), pp. 1219–1225. IEEE, Piscataway (2018). https://doi.org/10.1109/SSCI.2018.8628805 7. Cheong, P. Y., Aggarwal, D., Hanne, T., Dornberger, R.: Variation of ant colony optimization parameters for solving the travelling salesman problem. In: 2017 IEEE 4th International Conference on Soft Computing & Machine Intelligence (ISCMI), pp. 60–65. IEEE, Piscataway (2017). https://doi.org/10.1109/ISCMI.2017.8279598 8. Bullnheimer, B., Hartl, R.F., Strauss, C.: A new rank based version of the ant system – a computational study. CEJOR 7(1), 25–38 (1999) 9. Dorigo, M., Maniezzo, V., Colorni, A.: Ant system: optimization by a colony of cooperating agents. IEEE Tran. Syst. Man Cybern. B: Cybern. 26(1), 29–41 (1996). https://doi.org/10. 1109/3477.484436 10. Zeghida, D., Bounour, N., Meslati, D.: The ant-step algorithms: Reloading the ant system heuristic and the overlooked basic variants. In: 2020 IEEE 2nd International Conference on Electronics, Control, Optimization and Computer Science (ICECOCS), pp. 1–6. IEEE, Piscataway (2020). https://doi.org/10.1109/ICECOCS50124.2020.9314375 11. Jangra, R., Kait, R.: Analysis and comparison among Ant System; Ant Colony System and Max-Min Ant System with different parameters setting. In: 2017 3rd International Conference on Computational Intelligence & Communication Technology (CICT), pp. 1–4. IEEE, Piscataway (2017). https://doi.org/10.1109/CIACT.2017.7977376 12. Dorigo, M., Gambardella, L.M.: Ant colony system: a cooperative learning approach to the traveling salesman problem. IEEE Trans. Evol. Comput. 1(1), 53–66 (1997). https://doi.org/ 10.1109/4235.585892 13. Prakasam, A., Savarimuthu, N.: Metaheuristic algorithms and probabilistic behaviour: a comprehensive analysis of ant colony optimization and its variants. Artif. Intell. Rev. 45(1), 97–130 (2015). https://doi.org/10.1007/s10462-015-9441-y 14. Joshi, S., Kaur, S.: Comparative analysis of two different ant colony algorithm for model of TSP. In: 2015 International Conference on Advances in Computer Engineering and Applications, pp. 669–671. IEEE, Piscataway (2015). https://doi.org/10.1109/ICACEA.2015.716 4775 15. Chen, H., Tan, G., Qian, G., Chen, R.: Ant colony optimization with tabu table to solve TSP problem. In: 2018 37th Chinese Control Conference (CCC), pp. 2523–2527. IEEE, Piscataway (2018). https://doi.org/10.23919/ChiCC.2018.8483278 16. Ratanavilisagul, C.: Modified ant colony optimization with pheromone mutation for travelling salesman problem. In: 2017 14th International Conference on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology (ECTI-CON), pp. 411–414. IEEE, Piscataway (2017). https://doi.org/10.1109/ECTICon.2017.8096261

Tamilnadu Omnibus Travels Evaluation Using TOPSIS and Fuzzy TOPSIS Methods S. M. Vadivel1(B)

, A. H. Sequeira2

, Sunil Kumar Jauhar3

, and V. Chandana1

1 Department of Industrial and Production Engineering, The National Institute of Engineering,

Mysuru 570008, India [email protected] 2 School of Management, National Institute of Technology Karnataka, Surathkal 575025, India [email protected] 3 Indian Institute of Management Kashipur, Kundeshwari, Uttarakhand 244713, India [email protected]

Abstract. This research uses an effective fuzzy multi-criteria method to evaluate the performance of eight omnibus operators in Tamilnadu, India’s southernmost state (MA). The data was gathered from passengers (regular travelers) and the management of omnibus travel operators. Passengers provided qualitative data, while management offered quantitative data. To estimate the omnibus travels effectiveness, the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) technique and Fuzzy TOPSIS were employed to analyze the subjective assessments of both quantitative and qualitative data. Using linguistic concepts, the subjectiveness and inaccuracy of the performance assessment methods are expressed as fuzzy numbers. Traditionally, fuzzy MA models are often used to find the degree of possibility of an alternative of each attribute and sub-attributes to avoid complicated and inaccurate comparisons of fuzzy numbers. Then, grounded on the fuzzy techniques, it is converted into a weighted fuzzy performance index for each alternative. This technique is computationally efficient, and the principles behind it are basic and straightforward to comprehend. Keywords: TOPSIS · Fuzzy TOPSIS · Multicriteria Decision Making (MCDM) · Performance measurement · Omnibus transportation

1 Introduction Efficient performance assessment is a vital way to promote the efficiency of urban public transport systems and their level of service (Gomes 1989). It is necessary to address two primary obstacles. In general, the assessment criteria are complex and frequently arranged into multi-level hierarchies. Second, the assessment procedure generally included subjective evaluations, which led to qualitative and imprecise data. The fuzzy set theory introduced by Zadeh in 1965 and applied MA models provides an active way for defining decisions in a fuzzy situation with subjective and imprecise data. Chang et al. (1998) demonstrate the benefits of applying for fuzzy numbers in modeling traffic © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 24–31, 2022. https://doi.org/10.1007/978-3-030-96299-9_3

Tamilnadu Omnibus Travels Evaluation Using TOPSIS and Fuzzy TOPSIS Methods

25

and transportation systems. Fuzzy numbers can better reflect qualitative research or linguistic terms used to describe subjective and vague evaluations of the decision-making problem. Modeling transportation decision issues with fuzzy MA models has proved to be beneficial. AboulEla et al. (1982) showed the fuzzy MA used to assess transportation project choices at different phases of the planning process. Chang and Shyu (1994) imitated the fuzzy MA model to evaluate the performance of public transportation systems with multilayer hierarchies of criteria. Vadivel et al. (2020) used AHP to evaluate Tamilnadu private bus journeys from the standpoint of passengers. Other Tamilnadu states or southern Indian states are the final locations, with Chennai as the starting spot (Destination). Further, this paper is arranged as follows: Sect. 2 strengthens the literature support relevant to transportation. Section 3 is the recommended applied methodology. Section 4 is a detailed case study on Tamilnadu Omnibus Companies evaluation, and the final section concludes with the limitations and future benefits of this research study.

2 Literature Support Palczewski and Sałabun (2019) applied TOPSIS and FTOPSIS methodologies in the supply chain, environment, energy sources, business, and healthcare. Dhiman and Deb (2020) also used fuzzy-based MCDM techniques within three hybrid wind farms to determine the best approach in the field. Fuzzy TOPSIS and Fuzzy COPRAS were also used for evaluations and to analyze the potential lowering of tabulated rankings. Çalık (2021) pioneered a new group decision-making method based on Industry 4.0 constituents’ pieces designed for evaluating the best green supplier by combining AHP and TOPSIS methods under the Pythagorean fuzzy TOPSIS domains. This method took into account the distances of suppliers to obtain the ranking of the suppliers and determine the most satisfactory one. Prabhu et al. (2020) applied the fuzzy TOPSIS method to determine the factors relevant for overall production execution. Bottani and Rizzi (2006) presented a suitable 3PL service provider using the FTOPSIS methodology. Baki 2020 demonstrated fuzzy AHP and fuzzy TOPSIS to asses hotel websites. Boran (2017) analyzed the performance levels of power plants in Turkey using the FTOPSIS method due to the absence of complete information in the evaluation process. The criteria, installation cost, efficiency, emission of carbon dioxide gas, result from electricity cost, and social acceptance were all considered. Ranganath et al. (2020) developed a risk analysis through the application of the FTOPSIS method. Bera et al. (2021) proposed the interval type-2 fuzzy (IT2F) TOPSIS method to select optimum suppliers in Indian spare parts manufacturing companies. Vadivel et al. (2020) applied TOPSIS and FTOPSIS to an Indian postal service facility design layout for efficiency gains and more productive workflow. Very few research studies have been conducted on TOPSIS and FTOPSIS methodology to apply in Omnibus bus transport system evaluation in the southern part of India, Tamil Nadu state (Fig. 1).

26

S. M. Vadivel et al.

3 Research Methods

To find the optimum performance omnibus companies.

To find the operational criterion on attributes.

A survey from passengers, managers of omnibus travels, employers, etc.

Apply TOPSIS, Fuzzy (to reduce uncertainty) TOPSIS method

Ranking Alternatives Fig. 1. Recommend Methodology

4 Case Problem – Chennai, Tamilnadu, Southern India The study aims to evaluate omnibus transportation in Chennai omnibus travels. The assessment must be cost-effective for the passengers while still providing excellent customer service. 4.1 Application of TOPSIS Method Hwang and Yoon (1981) discovered that the TOPSIS approach requires the chosen option to be the smallest distance from the positive ideal solution and the furthest distance from the perfect negative solution. Due to a lack of comprehensive information (Zadeh, 1965). The advantage of implementing a fuzzy method is that fuzzy numbers are exact numbers used to give relative value to attributes. The fuzzy set theory represents decision-maker preferences using linguistic terms. Further, Vadivel and Sequeira (2020) has provided additional TOPSIS and FTOPSIS procedures to evaluate this transportation problem (Fig. 2) (Tables 1, 2, 3, 4, 5, 6, 7, 8 and 9).

Tamilnadu Omnibus Travels Evaluation Using TOPSIS and Fuzzy TOPSIS Methods

27

Table 1. TOPSIS Decision matrix Weightage value Alternatives

KPN MGM MJT Parveen Rathimeena RPN SRM Yogalakshmi

0.5

0.4

0.2

0.4

0.2

0.3

Safety 9 8 7 7 8 8 8 7

Comforts 9 8 7 8 8 7 9 7

Operation 8 6 7 7 8 7 9 6

Service 7 6 6 7 7 6 8 6

Social benefits 8 4 6 6 5 6 8 6

Finance 7 4 6 6 5 5 8 5

Alternative

8

m

Attribute

8

n

Xij

The score of option i concerning criterion j

X

xij

8x8

Table 2. TOPSIS normalized decision matrix BC1

0.41

0.40

0.39

0.37

0.45

0.42

KPN MGM MJT Parveen Rathimeena RPN SRM

0.36 0.32 0.32 0.36 0.36 0.36 0.32

0.36 0.31 0.36 0.36 0.31 0.40 0.31

0.29 0.34 0.34 0.39 0.34 0.44 0.29

0.32 0.32 0.37 0.37 0.32 0.42 0.32

0.22 0.33 0.33 0.28 0.33 0.45 0.33

0.24 0.36 0.36 0.30 0.30 0.48 0.30

Table 3. Results of TOPSIS methods Bus Companies

KPN MGM MJT Parveen Rathimeena RPN SRM Yogalakshmi

Si' 0.0953 0.0289 0.0435 0.0516 0.0459 0.0379 0.1079 0.0287

Si* 0.0295 0.1028 0.0881 0.0703 0.0736 0.0859 0.0224 0.0869

Ci* 0.7637 0.2195 0.3308 0.4233 0.3843 0.3058 0.8283 0.2482

TOPSIS Priorities 2 8 5 3 4 6 1 7

28

S. M. Vadivel et al.

Fig. 2. Fuzzy triangular membership function

Table 4. Fuzzy triangular membership function Rank

Grade

Membership function

Very Poor (VP) Poor (P) Neutral (N) Good (G) Very Good (VG)

I II III IV V

(0.00,0.10,0.25) (0.15,0.30,0.45) (0.35,0.50,0.65) (0.55,0.70,0.85) (0.75,0.90,1.00)

Table 5. Normalized decision matrix Bus Companies

C1

C2

C3

C4

C5

C6

KPN MGM MJT Parveen Rathimeena RPN SRM Yogalakshmi

0.41 0.36 0.32 0.32 0.36 0.36 0.36 0.32

0.40 0.36 0.31 0.36 0.36 0.31 0.40 0.31

0.39 0.29 0.34 0.34 0.39 0.34 0.44 0.29

0.37 0.32 0.32 0.37 0.37 0.32 0.42 0.32

0.45 0.22 0.33 0.33 0.28 0.33 0.45 0.33

0.42 0.24 0.36 0.36 0.30 0.30 0.48 0.30

Table 6. Fuzzy linguistic variables from Decision Matrix Bus Companies

KPN MGM MJT Parveen Rathimeena RPN SRM Yogalakshmi Weight

Safety

Comforts

Operation

Service

Social benefits

Finance

VG G G G VG G VG G VG

VG G G VG G G VG G G

VG N G G G G VG N P

G N M G G N G N G

VG P G G N N G N P

G N N G N N VG N N

Tamilnadu Omnibus Travels Evaluation Using TOPSIS and Fuzzy TOPSIS Methods

29

Table 7. Fuzzy decisions matrix and fuzzy attribute weights Bus Companies

C1

KPN MGM MJT Parveen Rathimeena RPN SRM Yogalakshmi

(0.75,0.90,1.00) (0.55,0.70,0.85) (0.55,0.70,0.85) (0.55,0.70,0.85) (0.75,0.90,1.00) (0.55,0.70,0.85) (0.75,0.90,1.00) (0.55,0.70,0.85) (0.75,0.90,1.00)

Weight

C2 (0.75,0.90,1.00) (0.55,0.70,0.85) (0.55,0.70,0.85) (0.75,0.90,1.00) (0.55,0.70,0.85) (0.55,0.70,0.85) (0.75,0.90,1.00) (0.55,0.70,0.85) (0.55,0.70,0.85)

C3

C4

C5

C6

(0.75,0.90,1.00) (0.35,0.50,0.65) (0.55,0.70,0.85) (0.55,0.70,0.85) (0.55,0.70,0.85) (0.55,0.70,0.85) (0.75,0.90,1.00) (0.35,0.50,0.65) (0.15,0.30,0.45)

(0.55,0.70,0.85) (0.35,0.50,0.65) (0.35,0.50,0.65) (0.55,0.70,0.85) (0.55,0.70,0.85) (0.35,0.50,0.65) (0.55,0.70,0.85) (0.35,0.50,0.65) (0.55,0.70,0.85)

(0.75,0.90,1.00) (0.15,0.30,0.45) (0.55,0.70,0.85) (0.55,0.70,0.85) (0.35,0.50,0.65) (0.35,0.50,0.65) (0.55,0.70,0.85) (0.35,0.50,0.65) (0.15,0.30,0.45)

(0.55,0.70,0.85) (0.35,0.50,0.65) (0.35,0.50,0.65) (0.55,0.70,0.85) (0.35,0.50,0.65) (0.35,0.50,0.65) (0.75,0.90,1.00) (0.35,0.50,0.65) (0.35,0.50,0.65)

Table 8. Analysis of FTOPSIS Bus Companies

KPN MGM MJT Parveen Rathimeena RPN SRM Yogalakshmi

C1 (0.56,0.81,1.00) (0.41,0.63,0.85) (0.41,0.63,0.85) (0.41,0.63,0.85) (0.56,0.81,1.00) (0.41,0.63,0.85) (0.56,0.81,1.00) (0.41,0.63,0.85)

C2 (0.41,0.63,0.85) (0.30,0.49,0.72) (0.30,0.49,0.72) (0.30,0.49,0.72) (0.30,0.49,0.72) (0.30,0.49,0.72) (0.30,0.49,0.72) (0.30,0.49,0.72)

C3

C4

(0.11,0.27,1.00) (0.05,0.15,0.29) (0.08,0.21,0.38) (0.08,0.21,0.38) (0.08,0.21,0.38) (0.08,0.21,0.38) (0.11,0.27,1.00) (0.05,0.15,0.29)

(0.30,0.49,0.72) (0.19,0.35,0.55) (0.19,0.35,0.55) (0.30,0.49,0.72) (0.30,0.49,0.72) (0.19,0.35,0.55) (0.30,0.49,0.72) (0.19,0.35,0.55)

C5 (0.11,0.31,0.45) (0.02,0.10,0.20) (0.08,0.21,0.38) (0.08,0.21,0.38) (0.05,0.15,0.29) (0.05,0.15,0.29) (0.08,0.21,0.38) (0.05,0.15,0.29)

C6 (0.19,0.35,0.55) (0.12,0.25,0.42) (0.12,0.25,0.42) (0.19,0.35,0.55) (0.12,0.25,0.42) (0.12,0.25,0.42) (0.26,0.45,0.65) (0.12,0.25,0.42)

A*

(1,1,1)

(1,1,1)

(1,1,1)

(1,1,1)

(1,1,1)

(1,1,1)

A-

(0,0,0)

(0,0,0)

(0,0,0)

(0,0,0)

(0,0,0)

(0,0,0)

W

(0.75,0.90,1.00)

(0.15,0.30,0.45)

(0.55,0.70,0.85)

(0.15,0.30,0.45)

(0.35,0.50,0.65)

(0.55,0.70,0.85)

Table 9. Fuzzy TOPSIS results

Bus Companies

Di*

Di'

Cci

Rank

KPN MGM MJT Parveen Rathimeena RPN SRM Yogalakshmi

3.2548 4.0756 3.9095 3.6882 3.7031 3.9655 3.3476 3.8936

3.3122 2.3910 2.3789 2.7753 2.8110 2.5161 3.2102 2.7932

0.5044 0.3697 0.3783 0.4294 0.4315 0.3882 0.4895 0.4177

1 8 7 4 3 6 2 5

30

S. M. Vadivel et al. Table 10. Comparative study results

Bus Companies

KPN MGM MJT Parveen Rathimeena RPN SRM Yogalakshmi

AHP Rank

TOPSIS Rank

2 5 6 3 4 7 1 8

2 8 5 3 4 6 1 7

Fuzzy TOPSIS 1 8 7 4 3 6 2 5

5 Results of the TOPSIS, FTOPSIS Technique Table 10 shows the results of the comparison analysis. We have found out that SRM Company Travels has the finest selection among passengers, followed by KPN Travels. The AHP, TOPSIS, and FTOPSIS methodologies have the highest global weightage (0.1732, 0.8283, 0.5044). In AHP, TOPSIS techniques, SRM is the primary priority, while KPN is the second. KPN is the top priority in the FTOPSIS approach, followed by SRM. This study did not conduct any sensitivity analysis considered as a limitation.

6 Conclusion The difficulty of evaluating the performance of metropolitan public transportation networks is subjective and inaccurate in nature. The most evident and successful technique for the DM to employ in the evaluation process can use fuzzy assessments articulated in linguistic terms. This paper introduces a compelling fuzzy MA approach that gives fresh positioning results to the assessment issue. A detailed investigation of 8 transport organizations in the Tamilnadu omnibus vehicle evaluation has been completed to exemplify the methodology. Public transportation is a day-to-day usage in our daily life, and it brought it into human wellbeing and safety in a city. The proposed TOPSIS FTOPSIS technique is advantageous as far as productivity and time-investment methods. The benefits of the methodology work with its implementation with the help of computer DSS for addressing viable execution assessment problems in a fuzzy nature. The limit of this review is concerned transport course in the southern piece of India. Later on, analysts can apply other MCDM strategies for this review in a viable way.

References Aboul-Ela, M.T., Stevens, A.M., Wilson, F.R.: A multiple criteria decision making methodology for transportation policy analysis. Logist. Transp. Rev. 18(3), 279–295 (1982)

Tamilnadu Omnibus Travels Evaluation Using TOPSIS and Fuzzy TOPSIS Methods

31

Baki, R: Evaluating hotel websites through the use of fuzzy AHP and fuzzy TOPSIS. Int. J. Contemp. Hosp. Manag. (2020) Bera, A.K., Jana, D.K., Banerjee, D., Nandy, D.: A group evaluation method for supplier selection based on interval type-2 fuzzy TOPSIS method. Int. J. Bus. Perform. Supply Chain Model. 12(1), 1–26 (2021) Boran, K.: An evaluation of power plants in turkey: fuzzy TOPSIS method. Energy Sources B 12(2), 119–125 (2017) Bottani, E., Rizzi: A fuzzy TOPSIS methodology to support outsourcing of logistics services. Supply Chain Manag. Int. J. (2006) Çalık, A.: A novel pythagorean fuzzy AHP and fuzzy TOPSIS methodology for green supplier selection in the Industry 4.0 era. Soft. Comput. 25(3), 2253–2265 (2020). https://doi.org/10. 1007/s00500-020-05294-9 Chang, Y.-H., Shyu, T.-H.: The application of fuzzy multicriteria decision making to the transit system performance evaluation. In: Proceedings of the Tenth International Conference on Multiple Criteria Decision Making, Taipei, pp. 351±360 (1994) Chang, Y-H., Yeh, C-H., Cheng, J-H.: Decision support for bus operations under uncertainty: a fuzzy expert system approach. Omega 26(3), 367±380 (1998) Dhiman, H.S., Deb, D: Fuzzy TOPSIS and fuzzy COPRAS based multi-criteria decision making for hybrid wind farms. Energy 202, 117755 (2020) Gomes, L.F.A.M.: Multicriteria ranking of urban transportation system alternatives. J. Adv. Transp. 23(1), 43±52 (1989) Palczewski, Sałabun: Application of TOPSIS and Fuzzy TOPSIS in the Supply Chain, Environment, Energy Sources, Business, Healthcare (2019) Prabhu, M., Abdullah, N.N., Ahmed, R.R., Nambirajan, T., Pandiyan, S: Segmenting the manufacturing industries and measuring the performance: Using interval-valued triangular fuzzy TOPSIS method. Complex Intell. Syst. 6, 591–606 (2020) Ranganath, N., Sarkar, D., Patel, P., Patel, S: Application of fuzzy TOPSIS method for risk evaluation in development and implementation of solar park in India. Int. J. Constr. Manag. 1–11 (2020) Vadivel, S.M., Sequeira, A.H.: Enhancing the operational performance of mail processing facility layout selection using multi-criteria decision-making methods. Int. J. Serv. Oper. Manag. 37(1), 56–89 (2020) Vadivel, S.M., Sequeira, A.H., Jauhar, S.K., Baskaran, R., Robert Rajkumar, S.: Application of multi-criteria decision-making method for the evaluation of Tamilnadu private bus companies. In: Pant, M., Kumar Sharma, T., Arya, R., Sahana, B.C., Zolfagharinia, H. (eds.) Soft Computing: Theories and Applications. AISC, vol. 1154, pp. 209–222. Springer, Singapore (2020). https://doi.org/10.1007/978-981-15-4032-5_21

A Hybrid Feature Extraction Method Using SeaLion Optimization for Meningioma Detection from MRI Brain Image S. U. Aswathy1 , Divya Stephen1(B) , Bibin Vincent2 , and P. Prajoon1 1 Jyothi Engineering College, Thrissur, Kerala, India 2 Providence College of Engineering, Thrissur, Kerala, India

Abstract. The Brain is the one that has a significant impact on the control and managing of the entire body. The ability to see, hear, think, walk, talk, feel, remember, and a lot more, and also the breathing which is the essential part to stay alive is controlled by the Brain. So it is a crucial part to take care of the brain from various diseases. Tumors, which are collections of abnormal growth of cells, can cause damage to the brain and can be malignant or non-cancerous. Here we are focusing on meningiomas, the majority of meningiomas are benign (non-cancerous) and slow-growing, although some are malignant. The detection of these types of tumors can be a daring task. As technology evolved, there are various methods that can detect brain tumors and even classify their types. The proposed work follows a hybrid feature extraction method that fuses PCA and GIST and also uses the SeaLion algorithm for optimization purposes. With the hybrid feature extraction techniques and the SLnO, the designed method shows a better classification accuracy. The paper includes the workflow of the proposed strategy, the first phase is all about the preprocessing of the image using the CLAHE and the anisotropic diffusion followed by the segmentation in the second phase, uses K-means, then the feature extraction in the third phase. The fourth phase deals with the optimization and finally the classification of inputs. The trials were conducted on 100 images from the human brain and a synthetic MRI dataset, with 25 images being healthy and 75 being problematic. On both training and test imagery, the classification performance was found to be 98.56%. Keywords: Brain tumor · Detection · Hybrid feature extraction · CLAHE · Anisotropic diffusion · PCA · GIST · Sea Lion

1 Introduction The brain is the body’s most critical organ, and should be cared for and treated in an extensive way. The tumor refers to an abnormal bunch of cells located inside the brain, tumor can be of two classes, benign and malignant tumor [1]. The proposed method mainly focuses on the meningiomas which are perhaps the most frequent benign intracranial tumors, accounting for 10 to 15% of all brain neoplasms, only with a small fraction of malignant tumors. As the technology evolved, detection and identification of © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 32–41, 2022. https://doi.org/10.1007/978-3-030-96299-9_4

A Hybrid Feature Extraction Method Using SeaLion Optimization

33

various medical illnesses are acquiring a huge demand in the medical sector [2]. Early detection of tumors probably can increase the lifespan of a person. In diagnosing brain tumors, image processing plays an important role. The anomalies found in the brain with respect to its size, position, or structure are re-sectioned and analyzed using an MRI scan or a CT scan. MRI is highlighted as the most efficient one comparing the other imaging techniques; It will not emit any radiation which is known to be dangerous to humans [3]. 1.1 Key Objectives The proposed strategy’s major goal is to effectively detect and categorize the meningiomas by making use of the MRI. As there are many techniques developed by the researchers for detecting the brain tumor, still we can’t point out a single global method for the classification of these tumors, as each method works well for others, So the highlights of our method are; a. A hybrid feature extraction strategy is proposed with SLnO, to aid the radiologist for brain tumor classification. b. We are using hybrid feature extraction, i.e. PCA and NGIST. PCA is one of the most effective techniques for finding patterns in datasets focusing on feature correlation and GIST is a global image descriptor that aids in characterizing an image’s different relevant statistics. c. In the segmentation phase, the k-means is developed to separate the focal region from the rest of the scene. d. We deliver high-quality output by using the Sea Lion optimization (SLnO) algorithm.

2 Proposed Methodology The method begins by using preprocessing tools to prepare the brain images, for that we make use of CLAHE and the Anisotropic diffusion and for segmentation, the Kmeans clustering technique is taken, and then the features are extracted by the hybrid technique, and optimization are done using SLnO. Finally, the classification of input images is carried out by CNN. The entire process is discussed in detail below, and also the diagrammatic representation of the proposed method is shown in Fig. 1. 2.1 Input Data The data for the proposed work is from BRATS 2015, which included 220 MRIs of high-grade gliomas (HGG) and 54 MRI scans of low-grade gliomas (LGG). T1, T1c, T2, and T2 FLAIR are the four MRI modalities. The most often used brain tumor dataset is given in Fig. 2. 2.2 Preprocessing As brain images have been more sensitive than other types of imaging, and highresolution images are required for diagnosis, the main aim of this step is to increase

34

S. U. Aswathy et al.

Fig. 1. Basic block diagram for the proposed method for detecting tumor from MRI brain image

Fig. 2. Data sources

imaging quality. It’s a crucial phase that boosts the quality of brain feature extraction and the final result of imaging analysis. We use CLAHE and anisotropic diffusion, which is a strategy to minimize picture noise without losing significant portions of the source images, like edges, line segments, and some important elements for image interpretation. CLAHE Algorithm CLAHE was designed to improve low-contrast medical images’ contrast [4], The contrast limiting of CLAHE differs from that of standard AHE. To combat the problem of noise amplification, the CLAHE added a clipping limit. Its techniques are as follows, and it delivers good outcomes on medical images [5].

A Hybrid Feature Extraction Method Using SeaLion Optimization

35

• Image is partitioned into numerous non-overlapping areas of almost similar size. Corner regions (CR), border regions (BR), and inner regions (IR) are the three categories of regions that result from this split. • Construct the histogram for every group. • When a clipping value is specified, clip the histogram to that value and then compute the CDFs using the clipped histograms. • For each pixel in the image, find the 4 closest adjacent grid points. • Using the intensity level of that pixel as an index, obtain its mapping at the 4 grid points based on their CDFs. • Interpolate between these values to get the mapping at the present pixel point. This intensity should be mapped to the [min: max] range and used in the final image. The CLAHE method divides an actual image into non-overlapping context-specific parts known as sub-images, tiles, or blocks. The most essential elements in the CLAHE are Block Size (BS) as well as Clip Limit (CL). These two parameters are mostly responsible for improved image quality (Fig. 3) (Table 1). Anisotropic Diffusion Anisotropic diffusion filter reduces picture noise without deleting significant sections of the relevant data in the image, such as borders, lines, as well as other important elements for image interpretation. The anisotropic diffusion equation is written as follows: ∂I = div(c(x, y, t)∇I ) = ∇c.∇I + c(x, y, t)I ∂t

(1)

where  denotes the Laplacian, ∇ denotes the gradient with respect to the space variables, div is the divergence operator and c (x, y, t) is the diffusion coefficient [6].

(a)

(b)

(c)

Fig. 3. (a) Input images (b) Preprocessed images after applying CLAHE (c) Preprocessed images after applying anisotropic diffusion filter

36

S. U. Aswathy et al. Table 1. Objective measures of preprocessed images

Metrics

PSNR

MSE

SSIM

Images

CLAHE

ADF

CLAHE

ADF

CLAHE

ADF

Image 1

31.5347

39.5470

3.369335

0.012112

0.7455

0.8393

Image 2

31.8810

40.6586

3.176397

0.013806

0.6264

0.7363

Image 3

34.3846

39.3762

1.752057

0.024475

0.7167

0.9941

Image 4

32.8765

38.73000

2.8610

1.3863

0.912808

0.996781

Image 5

32.90756

37.64297

3.2808

0.9724

0.879353

0.974054

2.3 Segmentation The major consideration of this segmentation process is the extraction of imperative components. Due to this, the data can be effectively seen. The K-means method divides data into K unique non-overlapping clusters, each with just a single data point, in an iterative process. It tries to provide as many comparable intra-cluster points as feasible while maintaining cluster separation. The data values were assigned to clusters with the goal of keeping the total of the squared distances among them and the cluster’s midpoint (arithmetic mean of all the data points in that cluster) as short as possible. The data points are much more identical if there is less variance within clusters (Fig. 4).

Algorithm for K-means Methodology Input: Number of required clusters & set of data Output: Set of clusters Algorithm Step 1: Choose a random number of clusters from the initial data where k is the size of clusters Step 2: Every Cluster needs a center point, choose random k points from the initial points (Every cluster has only one element present in it. i.e the center point) Step 3: For each point that is not a center point, 3.1 Find the distance between the current point and all center points 3.2 Take the cluster of least distance from the current point and assign the current point to that cluster Step 4: Determine the centroids of newly generated clusters once again. Step 5: Continue iterating until the mean value does not change. i.e. the clustering of data points remains constant.

2.4 Feature Extraction The image’s significant higher-level data or properties, such as statistical, structure, color, and texture aspects, are retrieved in this phase. The extraction of features is a crucial step in reducing the complexity of the classifier that is used to categorize the properties of an image. Furthermore, feature extraction is effectively used to improve the accuracy

A Hybrid Feature Extraction Method Using SeaLion Optimization

37

Fig. 4. Preprocessed image & segmented image

of the diagnosis process by selecting significant traits. The combination of PCA and NGIST is adopted. This feature extraction method combines the PCA approach with the GIST descriptor, which has been normalized using the L2 norm. The normalized GIST (NGIST) descriptor is a more advanced version of the standard GIST descriptor. The NGIST can handle changes in brightness and shadowing by normalizing images using the L2 norm. GIST features are computed from MRI scans using this combination method, which picks the eigenvectors with the maximum eigenvalues and manipulates them onto a novel feature subspace of similar or reduced dimensions [7]. 2.5 Optimization Using SLnO The Sea Lion Optimization (SLnO) technique is used to tackle combinatorial optimization problems and has the capacity to perform exploration and exploitation stages. As a result, it is classified as a global optimizer. The scheduling operation is performed by simulating sea lion hunting behavior, the SeaLion algorithm investigates the hunting habits of “SeaLion,” one of the intelligent animals, whiskers are used to detect prey. To chase the prey, they devise a strategic attack mechanism. This property is incorporated into the key generation method, which is critical in the permutation, diffusion, and transposition processes. The target prey is assumed to be the current best answer to the optimal solution by the SeaLion algorithm [8], see Raja Masadeh’s Sea Lion Optimization Algorithm [9] for more information. Below is the description of the Sea Lion algorithm.

38

S. U. Aswathy et al.

Algorithm for Sea Lion Input: Population Output: Best Solution Algorithm Step 1: Initialize population Step 2: Select a random sea lion ( ) from the present population Step 3: Calculate the fitness function for each search agent Step 5: is the best candidate search agent who has the best fitness Step 6: If (i < max number of iterations), continue, Else Stop Step 7: Calculate using the equation, 1 2

|

Step 8: If

)

2

|

> 0.25, modify the location of the current search agent and go to step 12

Step 9: Else if (|c| < 1), modify the location of the current search agent and go to step 12 Step 10: Choose a random search agent Step 11: The location of the current search agent, should e modified and go to step 12 Step 12: If the search agent doesn’t belong to any then go to step 6 Step 13: Calculate the fitness function for every search agent Step 14: If a good alternative becomes available, modify SL. Step 15: Return (best solution)

In the above algorithm, SPleader indicates the velocity of sound of sea lion leader, and the velocity of sounds in water and air are represented by V1 and V2, respectively. 2.6 CNN Based Classification Classification is a computer vision process that can classify an image based on its visual information. After optimizing feature selection, the selected feature is provided to the classifier for training purposes. CNN classifier is being used in our research. Convolutional Neural Networks (CNNs) are multilayered neural networks with a distinct structure for recognizing complex data patterns. CNN’s are convolutional neural networks that combine non - linear, pooled, and interconnected layers with a convolutional layer to build a deep convolutional neural network and It can be advantageous depending on the application [10]. However, it introduces new training parameters. The backpropagation method is used to train convolutional filters in CNN. The filter structure’s shape is determined by the task at hand. Convolution, ReLu (Rectified Linear Unit), Pooling, Flattening, and Full Connection are all important concepts in CNN [11]. There are a variety of CNN designs available, all of which have played a role in developing algorithms that strengthen AI.

3 Experimental Results The proposed technique was tested and simulated in Python using several MRI scans, BRATS dataset was used in this experiment. Based on tumor segmentation and classification, the performance of suggested and existing approaches is assessed. In terms

A Hybrid Feature Extraction Method Using SeaLion Optimization

39

Table 2. Performance evaluation of proposed segmentation models

Image 1

Image 2

Image 3

Image 4

Image 5

Metrics

K-means

CNN

Dice Coefficient

0.9818

0.9868

Jaccard Coefficient

0.9642

0.9740

MCC

0.9814

0.9864

Accuracy

0.9991

0.9993

Dice Coefficient

0.7846

0.9442

Jaccard Coefficient

0.6456

0.8943

MCC

0.7647

0.9443

Accuracy

0.9989

0.9997

Dice Coefficient

0.7962

0.9326

Jaccard Coefficient

0.6617

0.8737

MCC

0.8117

0.9340

Accuracy(%)

0.9957

0.9979

Dice Coefficient

0.9085

0.9884

Jaccard Coefficient

0.8326

0.9770

MCC

0.9114

0.9882

Accuracy

0.9974

0.9988

Dice Coefficient

0.9818

0.9868

Jaccard Coefficient

0.9642

0.9740

MCC

0.9814

0.9864

Accuracy

0.9991

0.9993

of tumor accuracy, precision, sensitivity, segmentation, F measurement, Jaccard Index, Dice overlapping index, accuracy, and MCC (Matthews’ correlation coefficient). A total of 50 MRI scans were used in the experiment, which was separated into two groups: normal and diseased. A set of 25 random MRI images was chosen from the 50 MRI images. The experimental result in Table 2 shows that the overall accuracy is great when employing the optimized technique. The proposed strategy appears to function well in both scenarios where the items in the image are indistinct and distinct from the backdrop, based on our tests on several photographs. Our proposed technique’s effectiveness is shown by the experimental findings in addressing more segmentation difficulties by improving segmentation quality and precision in a short amount of time. Table 3 demonstrates that the CNN classifiers have higher precision, sensitivity, precision, and specificity. The developed model is compared to the state-of-the-art in terms of brain tumor classification accuracy shown in Table 4. The results reveal that the new strategy is extremely effective and the use of the proposed tumor detection method has been proved to improve clinical practice efficacy and precision.

40

S. U. Aswathy et al. Table 3. Performance value of CNN

Metrics

CNN

ResNet

VGG

Dense Net

Accuracy

0.98

0.965

0.97

0.98

Error

0.01

0.045

0.03

0.02

Sensitivity

0.98

0.965

0.97

0.96

Specificity

0.9833

0.9653

0.9733

0.97

Precision

0.9802

0.9602

0.9702

0.976

F1-Score

0.9789

0.9600

0.9700

0.978

Table 4. Comparison with existing techniques Sl.no

Method

Accuracy

1

GLCM & ANN

96.89

2

Gabor, DWT, GLCM & ANN

97.99

3

PCA-NGIST & RELM

94.233

4

Proposed PCA-NGIST & SLO

98.56

4 Conclusions Brain tumors were very common, and the number of patients climbed year after year. This has put considerable strain on the medical workers working in this area. There is a pressing need to develop a procedure that is effective and precise for segmenting brain tumor images in order to meet escalating demand. The snag face in medical image processing is a fully automated brain tumor segmentation system with high accuracy and precision. Even though much effort has been carried out in finding a better solution for this issue, it still remains unsolved. The complicated structure of the brain maximizes the difficulty of the tumor detection process. To reduce that complication, a hybrid approach is introduced because hybrid methods are more accurate and efficient. The recommended method has an accuracy of 98.56%. Specificity obtained for the proposed system is about 99.88%, which is comparatively higher when compared with existing techniques such as, EM - Level Set and FCM method.

References 1. Hashemzehi, R., Mahdavi, S., Kheirabadi, M., Kamel Tabbakh, S.: Detection of brain tumors from MRI images base on deep learning using hybrid model CNN and NADE. Biocybern. Biomed. Eng. 40 (2020). https://doi.org/10.1016/j.bbe.2020.06.001 2. Sasank, V.V.S., Venkateswarlu, S.: Brain tumor classification using modified kernel based softplus extreme learning machine. Multimed. Tools Appl. 80(9), 13513–13534 (2021). https:// doi.org/10.1007/s11042-020-10423-5

A Hybrid Feature Extraction Method Using SeaLion Optimization

41

3. Mohsen, H., El-Dahshan, E.-S., El-Horbarty, E.-S., Salem, M.A.-B.: Classification using deep learning neural networks for brain tumors. Future Comput. Inform. J. 3. https://doi.org/10. 1016/j.fcij.2017.12.001 4. Khan, S., Hussain, S., Yang, S.: Contrast enhancement of low-contrast medical images using modified contrast limited adaptive histogram equalization. J. Med. Imaging Health Infor. 10, 1795–1803 (2020). https://doi.org/10.1166/jmihi.2020.3196 5. Reza, A.M.: Realization of the contrast limited adaptive histogram equalization (CLAHE) for real time image enhancement. J VLSI Signal Process System 38(1), 35–44 (2004) 6. Perona, P., Malik, J.: Scale-space and edge detection using anisotropic diffusion. IEEE Trans. Pattern Anal. Mach. Intell. 12(7), 629–639 (1990). https://doi.org/10.1109/34.56205 7. Gumaei, A., Hassan, M.M., Hassan, M.R., Alelaiwi, A., Fortino, G.: A hybrid feature extraction method with regularized extreme learning machine for brain tumor classification. IEEE Access 7, 36266–36273 (2019). https://doi.org/10.1109/ACCESS.2019.2904145 8. Latha, H.R., Rama Prasath, A.: Enhanced image security using new sea lion optimization algorithm. International Journal of Innovative Technology and Exploring Engineering (IJITEE), ISSN: 2278–3075, 9(7) (May 2020) 9. Masadeh, R., Mahafzah, B.A., Sharieh, A.: Sea lion optimization algorithm. (IJACSA) International Journal of Advanced Computer Science and Applications, 10(5) (2019) 10. Suryani, D., Doetsch, P., Ney, H.: On the benefits of convolutional neural network combinations in offline handwriting recognition. In: Proceedings of 15th International Conference on Frontiers in Handwriting Recognition (ICFHR), pp. 193–198. Shenzhen, China, October 2016 11. Albawi, S., Mohammed, T., Al-azawi, S.: Understanding of a convolutional neural network. In: Proceedings of 2017 International Conference on Engineering & Technology (ICET’2017), pp. 274–279. Akdeniz University, Antalya, Turkey, August 2017 12. Aswathy, S.U., Devadhas, G., Kumar, S.S.: Quick detection of brain tumor using a combination of EM and level set method. Indian J. Sci. Technol. 8(34) (2015) 13. Aswathy, S.U., Devadhas, G., Kumar, S.S.: A survey on detection of brain tumor from MRI brain images. In: 2014 International Conference on Control, Instrumentation, Communication and Computational Technologies (ICCICCT), pp. 871–877 (2014) 14. Aswathy, S.U., Devadhas, G.G., Kumar, S.S.: An improved tumor segmentation algorithm from T2 and FLAIR multimodality MRI brain images by support vector machine and genetic algorithm. Cogent Eng. 5(1), 1470915 (2019)

An Effective Integrity Verification Scheme for Ensuring Data Integrity in Cloud Computing Minakshi Kamboj1(B) and Sanjeev Rana2 1 CSE Department MMEC, M.M. (Deemed To Be University),

Mullana, Ambala, Haryana, India 2 MMEC, M. M. (Deemed To Be University), Mullana, Ambala, Haryana, India

[email protected]

Abstract. Cloud computing is growing in quality thanks to its ability to supply dynamically ascendible resources provisioned as services no matter user or location device. However, moving knowledge to the cloud means the management of the info is a lot of within the hands of the cloud supplier instead of the info owner. Despite the advantages that the cloud presents, cloud computing technology is moon-faced with a range of legal and technological challenges. Security and privacy are amongst the most important challenges as known. There should be certain type of system to guarantee the integrity of the records. The current Cloud safety model is sited on the interpretation that the client should trust the source. Guaranteeing documents integrity needs a promise of trust among the customer and the supplier. Numerous cryptographic methods have been developed and used to deliver improved safety for the sake of keeping integrity issues. We assume that our file (F) is divided into n shares. Poly1305 could be a cryptographically message authentication code (MAC). It is often wont to verify the info integrity and therefore the legitimacy of a message. Shamir Secret Sharing is used to split client file into n shares and these shares are further encrypted using poly1305. An Effective Integrity Verification Scheme (EIVS) based SSS and Poly-1305 is implemented and compared with SSS-AES. Entropy and Timing analysis of EIVS is done for short, long and larger text files. Keywords: Shamir secret sharing · Effective integrity verification scheme · Advance encryption system · Message authentication code

1 Introduction The management, suppleness, and easiness of employment of Cloud Computing arrive with variety of challenges/issues. Loads of latest applications area unit deployed on net each day and number of individual’s victimization these services area unit growing quickly. A recent studies shows power consumption of server from across the globe which incorporates power consumption by the auxiliary equipment’s and cooling system is around North American country $7.2billion [13]. Within the study it’s been determined that the consumption had been doubled since year 2000. These of surveys has born to a replacement support known as inexperienced computing that is growing with the © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 42–54, 2022. https://doi.org/10.1007/978-3-030-96299-9_5

An Effective Integrity Verification Scheme for Ensuring Data Integrity

43

aim to create the system energy economical and economical utilization of resources. Studies shows average utilization of knowledge centers is nearly 2 hundredth and energy consumed by the idle resources is the maximum amount as hour of the height power. Green computing is clear as the use of computing with the additional environment responsibility. This concept contains the energy efficiency peripherals, energy efficient processors (CPU’s) and servers at many data centers or cloud centers. This also includes of incomplete usage of resources and a device for right ‘e-waste’ management. Green computing can also be mentioned to as the subject of study associated to the application of the computational resources extra efficiently. Even though there is a great worry in the group that Cloud computing can bring about higher energy use by the data centers, the Cloud processing has a green covering. In explicit, security has been extensively reported to represent the most factors that forestall migration to the cloud. With the advent of the internet, protection has become a chief worry. Many corporations secure themselves with the help of firewalls and encryption mechanisms. The corporations can create an “intranet” remains connected to the internet and secured it from possible threats. To maintain integrity of data healthier encryption techniques are required for improved safety. Classic approach of Shamir Secret Sharing creates shares which are smaller in size and integrity checks at combined is not fulfilled accurately. Intruder can damage with particular of the shares. Later reestablishing the (malevolent) secret, it is very difficult to identify that it has been altered. Advance Encryption System with secret sharing scheme (AES-SSS) without using cryptographic key is efficient one in that case. Modifying secret sharing classic algorithm and adding the concept of integrity in the secret sharing scheme, Secure the shared secret with a MAC (Message Authentication Code) wrapping round the secret-sharing part. Poly1305 is a cryptographically communication validation code. It can be castoff to validate the statistics truthfulness and the genuineness of a message. This collection practices a tradition randombytes utility to produce a random encapsulation key, which dialogs straight to the operative system; as a result the shares cannot be altered. The remaining of the article is structured as follows. In Sect. 2, we analyze the present security models proposed in the literature for cloud. In Sect. 3, we present the algorithms of the proposed model and evaluation parameters. In Sect. 4, we present outcomes and In Sect. 5; we conclude the article.

2 Related Work This fragment analyses the associated works from different areas: Sweta K. Parmar et al. (2013) Safety is the most troublesome points of view in the web and web application. Web and systems uses are developing snappy, so the unquestionable quality and the reputation of the swap over reports for the web or additional media kinds are creating. Data safety has been basic issue in data correspondence. Some incident or peril to data can end up being unimaginable adversity to the affiliation. Encryption system shows a key part in data security structure. This article provides a relationship of many encryption figuring and after that finds best available one count for the security framework [1]. Rasheed et al. (2014) looked issue of passed on figuring security surveying from three of view: client surveying necessities, specific methodology for (information) security

44

M. Kamboj and S. Rana

assessing and current cloud expert network capacities with regards to meeting review prerequisites. They additionally isolated explicit keeping an eye on issues into two classes: foundation security surveying and information security breaking down. They found at long last that despite distinctive methodologies accessible to address client exploring worries in the information reviewing zone, cloud suppliers have up ‘til now basically spun around foundation security surveying concerns [2]. Sirohi et al. (2015) proposed structure centers around the encryption and unscrambling methodology enabling the cloud client with information safety affirmation. The suggested strategy just discussed the all-inclusive safety at any rate did not analyze the execution. The game-plan in addition joined the working of sensible virtual machine, malware territory and predictable checking of the framework. This article provides a layout of various safety questions and dangers. An information safety structure besides gave the straightforwardness to both the cloud master focus and the cloud client thusly decreasing data safety dangers in cloud situation [3]. Marquez et al. (2016) demonstrated a pertinent examination about design of the movement of the safety parts of an inheritance application to Cloud suppliers by means of utilizing the framework named SMiLe2Cloud. The Cloud Figuring presented a wide combination of central focuses, yet also a fundamental test from the perspective of safety, in truth safety continued the rule deterrent to progress. Improvement of inheritance frameworks to the cloud enabled them to keep a mind security in heritage structures. The framework named SMiLe2Cloud projected to manage the matter of secure undertaking heritage data structures to cloud [4]. Subha et al. (2017) proposed the safety imperfection of their arrangement when dynamic enemies were related with dispersed capacity. There was various security concerns ought to have been watched out for when the data was kept up by outcast expert center in cloud. The overseer was familiar with survey the uprightness of the data in light of a legitimate concern for the consumer in direction to ensure decency of information. This could be known as as open auditability of information. Starting late, two assurances sparing looking into frameworks named Oruta and Knox were familiar with check the exactness of set away data. A working foe was prepared for adjusting the data set away in cloud abstractly. This data change was not being accepted through the customer and the commentator in the checked technique. They tried to suggest a response for assurance this imperfection by denoting the validation response made on the cloud server side. By then the stamped evidence was directed to the trusted in pariah evaluator (TTPA) for affirmation. The inspector firstly checked the stamp and for the endorsement of the proof. The suggested arrangement was finished up being protected alongside active opponent [5]. Binu V P et al. (2015) propose some effective and simple to actualize mystery sharing plans in such manner dependent on number hypothesis and bitwise XOR. These plans are likewise reasonable for secure sharing of pictures. Mystery picture sharing dependent on Shamir’s plans are lossy and includes convoluted Lagrange insertion [6]. A. A. Raipure et al. (2016) for verifying information re-appropriating homomorphic encryption, information encryption and mystery sharing calculations are the systems utilized broadly to verify information re-appropriating. Overseeing CIA (Confidentiality, Integrity, and Availability) is a primary concern and because of safety matters, clients

An Effective Integrity Verification Scheme for Ensuring Data Integrity

45

are deciding on multi cloud as these are verified with different strategies and one of these systems is mystery sharing calculations. Shamir’s Secret Sharing Algorithm is applied which address potential approaches and answers for will protect information re-appropriating in multi mists. The primary focal point of this article is on information safety and decreasing safety dangers [7]. Deneisha Gayle Tieng et al. (2016) assaults including coordinated effort among inside enemies were not yet examined. This paper intends to perform examination on these cases by checking whether SSSS is inclined to specific assaults made by inside foes. The assaults are put together from the assaults with respect to Harn and Lin (2009) and certain alterations were made [8]. Oriol Farras et al. (2016) effort is devoted to the pursuit of limits on the data proportion of non-impeccable mystery distribution plans and the development of effective straight non-immaculate mystery sharing plans [9]. Selda Calkavur et al. (2018) present a picture mystery sharing technique dependent on Shamir mystery sharing strategy. We utilize the lattice projection to build mystery sharing plan. It is a successful, dependable and secure strategy to keep the mystery picture from being lost, taken and ruined [10]. Fatih Molla et al. (2018) propose another way to deal with build mystery sharing plans dependent on field expansions [11]. Mitsugu Iwamoto et al. (2018) whole number writing computer programs is utilized to appropriate ideally the portions of (t,m)- limit plan to every member of the general access structure. From the optimality, it can generally accomplish lesser coding degree than the combined maps aside from the cases that they give the ideal circulation. A similar strategy is additionally functional to develop SSSs for inadequate entrance constructions and additionally incline contact constructions [12].

3 Proposed Work 3.1 Proposed Work and Algorithms Shares generated by Classical Shamir secret sharing are short in size. Small size shares can be tampered by attacker very easily. When this malicious secret is restored, we can’t find whether any of these shares has been tampered. Cloud reliability is a wide-ranging term comprises safety, confidentiality, and accurateness of statistics. Secure the shared secret with a MAC (Message Authentication Code) wrapper around the secret-sharing part. Because of this, the shares are always a little bit larger than the original data. An Effective Integrity Verification Scheme (EIVS) based on Shamir Secret Sharing and Poly-1305 is implemented and compared with SSS-AES (Table 1).

46

M. Kamboj and S. Rana Algorithm Secret_Message_Split Step 1: Plain _to_HexSharers (User_Secret_Message) Secret Charset= User_Secret_Message Shares Charset = String.hexdigits [0:16] Step 2: Split_Secret_Message (Secret_Charset, Threshold, Count_Shares) Secret_integer = Charset_to_integer (User_Secret_Message, Secret_Charset) Points = Secret_integer_to_points (Secret_integer, Threshold, Count_shares) GShares = [] For p in points GShares. Append (Point_to_Share_String (Point, Shares_Charset)) Return GShares Step 3: H_Key = (Pf, User_Secret_Message, Salt, C, D_K_Len) Where Pf is a pseudorandom utility User_Secret_Message is from which a resulting key is formed Salt is a sequence of bits, known as a cryptographic salt C is preferred count of repetitions D_K_Len is the ideal extent of the resulting key H_Key is the formed resulting key Step 4: Create_Key_Shares (&H_Key, count, Threshold)

Algorithm Shares_Combine Step 1: Read shares into memory line by line Step 2: Interpret each line If (line.len () mod 2 0) then Print “Shares are of length” Exit (0) Step 3: Splitting off the keyshares Mine Key_Shares from Cipher_texts Key_Shares = With_Capacity (Interpreted_lines.len ()); Cipher_texts = With_Capacity (Interpreted_lines.len ()); Step 4: Check whether any of shares is corrupt Print “Error if the Cipher_texts are not all the similar” Exit (0) Step 5: Return the Ecryption_Key Key = Match Combine_Key_Shares (&Key_Shares)

3.2 Evaluation Parameters Table 1. Performance metrics Parameter name Description Split time

Time required generating shares from secret message or filing

Combine time

Time required combining shares to recover secret message or file

Entropy

Entropy is catalog of data and is considered as bits per character. It measures hardness of key

An Effective Integrity Verification Scheme for Ensuring Data Integrity

47

4 Results 4.1 Generating Shares of Large 1 MB Text File (SSS-Poly1305) $ time echo shares.txt Here demo is text file of size 1 MB and shares.txt is output file name to stores shares. Total no. of shares to be generated is 3 and threshold is given as 3 (Fig. 1).

Fig. 1. Generating shares of large text file

Fig. 2. Shares.txt

Figure 2 shows the contents of shares.txt, 3-shares of demo.txt file. 4.2 Splitting Shares (SSS-Poly1305) Here shatres.txt is divided into three parts (xaa, xab, xac) as shown in Fig. 3. Here xaa, xab and xac are three text files.

Fig. 3. Splitting of Shares

48

M. Kamboj and S. Rana

4.3 Distributing Shares Among Clients Now these three separate shares files are distributed among different clients as shown in Figs. 4, 5 and 6.

Fig. 4. Received share at Client-1 (xaa)

Fig. 5. Received share at Client-2 (xab)

4.4 Receiving Shares from Clients and Combining (SSS-Poly1305) time head −n 3 0, the problem (10) has a unique solution. Existence of the Solution and Deadlock State: The critical case pr = 0 in which no retransmission is considered leads to a deadlock state. This state is related to the absorbing states of the Markov process, which can be reached with a nonnegative arrival probability. The schemes considered in this paper have different absorbing states, and the probability of reaching one of these states depends on the initial distribution of the considered scheme. Since these absorbing states represent the system’s bottleneck, we shall exclude the case of pr = 0 in the optimization problem (10). The optimization problem (10) admits a solution in [δ, 1], where δ > 0 since the steady-state probabilities πN (pf , pr ) are continuous over a closed interval [δ, 1].

6

Numerical Results

In this section, we present the numerical results associated with our system using the three proposed schemes, to evaluate the performance metrics. Thus, we give a comparative analysis of these metrics taking into account LoRa Standard as a reference. To better understand how a large number of devices influence our

A Comparative Study of Three LoRa Collision Resolution Schemes

117

Fig. 1. Normalized throughput and average delay, depending on M .

system performances, we consider the example of 1000 end-devices transmitting with a probability of 0.1 [packet/slot]. Figure 1a shows the total throughput as a function of the number of enddevices. The results show that the L-CE scheme maintains a throughput greater than 0.8 for any number of end devices. For the L-SC scheme, we can observe that it maintains the throughput between 0.5 and 0.6 for all end-devices. However, for the L-ZD scheme, a small drop in throughput is noted when the number of end devices tends towards 1000, where it takes a value close to 0.7. In addition, LCE outperforms L-ZD in terms of throughput. This is because the L-ZD scheme operates only when a collision occurs between two simultaneous transmissions, unlike the L-CE scheme, which operates when two or more packets are sent simultaneously. On the other hand, L-ZD outperforms L-SC since this latter scheme can retrieve only one packet from two simultaneous transmissions, unlike the L-ZD scheme which can retrieve both the two collided packets. The results also show a significant improvement of 47.63% in throughput compared with standard LoRa. Figure 1b shows the delay for different numbers of connected end-devices. The delay is given in terms of slots and represents the time between the transmission of the packet and when it is received successfully. As the number of end-devices increases, the delay tends to increase due to collisions resulting from simultaneous transmissions. In addition, L-CE guarantees the lowest delay compared to L-SC and L-ZD. Figure 2a plots the average number of backlogged packets versus the number of end devices. We report that the number of backlogged packets increases almost linearly with the number of end-devices, regardless of the scheme used. This is due to the reason that when the system is backlogged, any new packet will get to collide with the other old packets. Figure 2b shows the optimal retransmission probability p∗r , determined by solving the team problem (10). The results show that p∗r decreases when the number of devices increases. Indeed, it is necessary to decrease the transmission rate when many devices try to access the wireless medium. Even if this probability is low, the amount of traffic generated by all the devices combined, results

118

A. Amzil et al.

Fig. 2. Number of backlogged packet and optimal retransmission probability, depending on M .

in traffic congestion and a dramatic decrease in the performance of the overall system. Therefore, the retransmission probability should be lowered when we are dealing with many devices. Our results also show that the optimal retransmission probability in the case of L-SC is slightly higher than that of L-ZD especially in the case where the number of end-devices is less than 600. In the same context, L-CE has the highest probability value compared with the L-SC and L-ZD schemes. This is because a receiver using the L-CE technique can successfully decode the packet of N simultaneous transmissions (as long as it is the only one transmitted at the highest power). In contrast, the L-ZD and L-SC techniques are only able to decode the packet from two simultaneous transmissions.

7

Conclusion and Perspectives

In this paper, we have proposed a comparative study for the LoRa network enhanced with different collision resolution schemes (CE, ZD, and SC). For each scheme, we constructed a Discrete-Time Markov Chain (DTMC). Then, we derived all the performance metrics of interest using the steady-state probabilities. Finally, we obtained the numerical results using Matlab. Our results showed that (L-CE) outperforms all other schemes. In terms of throughput, we get an improvement of 81, 73%, 67, 25% and 55, 38%, respectively for (L-CE), (L-ZD) and (L-SC). Our delay results, show decrease of 12, 16%, 18, 46% and 15, 32% for (L-CE), (L-SC) and (L-ZD), respectively. Our future work consists of comparing more collision resolution schemes for an ultra dense LoRa network. Besides, we plan to develop a general model where we can switch between different schemes with an optimal decision, taking into account the network load, traffic type, and the required QoS.

References 1. Chaudhari, B.S., Zennaro, M.: LPWAN Technologies for IoT and M2M Applications. Elsevier, London (2020). Google-Books-ID: 4SHXDwAAQBAJ

A Comparative Study of Three LoRa Collision Resolution Schemes

119

2. Bellouch, A., Boujnoui, A., Zaaloul, A., Haqiq, A.: Three-dimensional Markov chain model to help reduce the spread of COVID-19 in IoT environment. J. Comput. Inf. Syst. Ind. Manag. Appl 13, 275–285 (2021). ISSN 2150-7988 3. El Kafhali, S., El Mir, I., Hanini, M.: Security threats, defense mechanisms, challenges, and future directions in cloud computing. Arch. Comput. Methods Eng. 29, 223–346 (2021). https://doi.org/10.1007/s11831-021-09573-y. Springer Science and Business Media LLC 4. El Kafhali, S., Chahir, C., Hanini, M., Salah, K.: Architecture to manage internet of things data using blockchain and fog computing. In: Proceedings of the 4th International Conference on Big Data and Internet of Things. ACM, 23 October 2019 5. Raza, U., Kulkarni, P., Sooriyabandara, M.: Low power wide area network analysis: can LoRa scale. IEEE Wirel. Commun. Lett. 6(2), 162–165 (2017) 6. Mikhaylov, K., Petaejaejaervi, J., Haenninen, T.: Analysis of capacity and scalability of the LoRa low power wide area network technology. In: 22th European Wireless Conference, pp. 1–6 (2016) 7. Zaaloul, A., Aattar, M.B.E.: A mathematical model to evaluate delay and power consumption of S-ALOHA protocol in an IoT environment. J. Theor. Appl. Inf. Technol. 99(15), 3693–3702 (2021) 8. Noreen, U., Clavier, L., Bounceur, A.: LoRa like CSS-based PHY layer, capture effect and serial interference cancellation. In: 24th European Wireless Conference, pp. 1–6 (2018) 9. Faber, M.J., van der Zwaag, K.M., dos Santos, W.G.V., Rocha, H.R.D.O., Segatto, M.E.V., Silva, J.A.L.: A theoretical and experimental evaluation on the performance of LoRa technology. IEEE Sens. J. 20(16), 9480 (2020) 10. Bellouch, A., Boujnoui, A., Zaaloul, A., Haqiq, A.: Modeling and performance evaluation of LoRa network based on capture effect. In: Hassanien, A.E., Darwish, A., Abd El-Kader, S.M., Alboaneen, D.A. (eds.) Enabling Machine Learning Applications in Data Science. Algorithms for Intelligent Systems, pp. 249–263. Springer, Singapore (2021). https://doi.org/10.1007/978-981-33-6129-4 18 11. Nehme, J.A., Nicolas, C., Habib, G., Haddad, N., Duran-Faundez, C.: Experimental study of lora performance: a concrete building case. In: 2021 IEEE International Conference on Automation/XXIV (ICA-ACCA) 12. Abramson, N.: The ALOHA system-another alternative for computer communications. In: Proceedings of the Fall Joint Computer Conference, pp. 281–285, 17–19 November 1970 13. Bellouch, A., Boujnoui, A., Zaaloul, A., Haqiq, A.: Hybrid approach for improving slotted ALOHA based on capture effect and ZigZag decoding techniques. In: Hassanien, A.E. et al. (eds.) Proceedings of the International Conference on Artificial Intelligence and Computer Vision (AICV2021). AICV 2021. AISC, vol. 1377, pp. 218–227. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-76346-6 20 14. Jia, D., Xiao, M., Cao, C., Kuang, J.: Enhanced frameless slotted ALOHA protocol with Markov chains analysis. Sci. China Inf. Sci. 61, 102304 (2018). https://doi. org/10.1007/s11432-017-9296-6 15. Boujnoui, A., Zaaloul, A., Haqiq, A.: Enhanced pricing strategy for slotted ALOHA with ZigZag decoding: a stochastic game approach. J. Comput. Inf. Syst. Ind. Manage. Appl. 13, 160–171 (2021). ISSN 2150-7988 16. Yue, W.: The effect of capture on performance of multichannel slotted ALOHA systems. IEEE Trans. Commun. 39(6), 818 (1991)

Production Scheduling Using Multi-objective Optimization and Cluster Approaches Beatriz Flamia Azevedo1,2(B) , Maria Leonilde R. Varela2 , and Ana I. Pereira1,2 1

2

Research Centre in Digitalization and Intelligent Robotics (CeDRI), Instituto Polit´ecnico de Bragan¸ca, 5300-253 Bragan¸ca, Portugal {beatrizflamia,apereira}@ipb.pt Algoritmi Research Centre, University of Minho, Campus Azur´em, 4800-058 Guimar˜ aes, Portugal [email protected]

Abstract. Production scheduling is a crucial task in the manufacturing process. In this way, the managers need to make decisions about the jobs production schedule. However, this task is not simple to perform, often requiring complex software tools and specialized algorithms to find the optimal solution. This work considers a multi-objective optimization algorithm to explore the production scheduling performance measure in order to help managers in decision making related to jobs attribution in a set of parallel machines. For this, five important production scheduling performance measures (makespan, tardiness and earliness time, number of tardy and early jobs) were combined into three objective functions and the Pareto front generated was analyzed by cluster techniques. The results presented different combinations to optimize the production process, providing to the manager different possibilities to prioritize the objectives considered. Keywords: Parallel machines

1

· Simulation · NSGA · k-means

Introduction

Production scheduling is a decision-making process widely used in many manufacturing and services companies. It can be defined as the process of optimizing, controlling and determination of the limited production system resources (machines, workers, finances resource etc.) [1]. In the past, it was traditionally done manually, using paper and/or spreadsheets. However, today can be more or less easily implemented to obtain alternative production schedules digitally, for reaching alternative solutions, which although not being always optimal ones, can usually be trustful and appropriate. Nowadays, scheduling problems have been widely investigated and several methodologies and innovations have been proposed in this context. In [2] a hybrid multi-population genetic algorithm (GA) and constraint programming approach c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 120–129, 2022. https://doi.org/10.1007/978-3-030-96299-9_12

Production Scheduling Using Multi-objective Optimization

121

is developed, in order to solve a multi-objective model that considers three objectives in production scheduling problem: makespan, maximum machine workload and total tardiness. Besides, a simulation is developed to evaluate the influence of some uncertainties in the model. A multi-objective approach is present in [3], in which the objectives are to minimize the number of workers and the makespan time in a firm. Two exact models of mathematical programming and a multi-objective algorithm to solve the scheduling problem with worker allocation is detailed in the paper. The approach results demonstrated that is possible to optimize the firm resource using a limited staff and finish all jobs in a timely manner. Another relevant approach, considering parallel machine production scheduling is presented in [4], in this case, a multi-factory production network is considered, in which a set of factories and a set of demands is analyzed. The objective of the problem is to identify, the most appropriated factory for each job and it’s schedule. The problem was modeled as mixed-integer programming techniques and a heuristic algorithm was proposed to generate the set of Pareto optimal solutions. A comparison between the heuristic developed and a GA indicated better performance of the proposed algorithm. The application of multi-objective strategies in increasing recently, although the research in this area, in comparison with the single criterion problems, is limited. However, in a real industrial environment, normally it is necessary to deal with several objectives simultaneously [4], being the multi-objective strategies the methodology most indicated. This work considers multi-objective strategies to explore the production scheduling performance measure in order to assist managers in decision making related to job distribution for a set of parallel machines and a cluster algorithm is utilized to support the results understanding and also the decision support.

2

Production Scheduling Performance Measures

To quantify the efficiency and effectiveness of decision actions, it is important to establish performance measures. According to [5,6] performance measures allow evaluating the efficiency of the scheduling program, aiming at a final objective, which may be the maximization or minimization of the criteria used as performance measures. Thus, the optimization criteria allow achieving several goals [5,6], namely: maximization of production flow, the satisfaction of quality requirements and quick response to customers, minimization of production costs or the combination of all previous cases mentioned. The most usual performance measures in the objective function of a scheduling problem are presented below, however other measures can be consulted in [5–8]: Makespan Time: it is designed by Cmax , and it defines the time when the last jobs of a sequence of job is complete, where Cj represents the completion date of the job j (Eq. (1)). For jobs consisting of more than one operation, represents the moment when the last one is executed.

122

B. F. Azevedo et al.

Cmax = max(Cj )

(1)

Tardiness and Earliness Time: the tardiness concerns the difference between the conclusion date of the job Cj and its the due time dj . If this difference is positive, Tj , indicates a tardiness in conclusion (Eq. (2)), but if the difference is negative, Ej , means an anticipation or earliness of the conclusion period (Eq. (3)). Tj = Cj − dj , if Cj > dj

(2)

Ej = Cj − dj , if Cj < dj

(3)

Total Number of Tardy Jobs (N Ltotal ): indicates the total number of tardy jobs, N Ltotal , and it can be formulated by Eq. (4), where n represents the number of jobs available, and when N Lj assumes the value 1 it means that the corresponding job j is late (Lj > 0):  n  1, if Lj > 0; N Ltotal = N Lj = (4) 0, if Lj < 0. j=1 Total Number of Early Jobs (N Etotal ): indicates the total number of early jobs, N Etotal , and it can be formulated by Eq. (5), where n represents the number of jobs available, and when N Ej assumes the value 1 it means that the corresponding job j is early (Lj < 0):  n  1, if Lj < 0; N Etotal = N Ej = (5) 0, if Lj > 0. j=1

3

Methodology

Multi-objective optimization is an area of multiple-criteria decision-making, concerning mathematical optimization problems, involving more than one objective function to be optimized simultaneously [9]. Normally, these objectives are conflicting, [10] and multi-objective optimization techniques can provide multiple solutions, representing weightings between objectives through multiple possible solutions [9]. Most multi-objective optimization algorithms use the concept of dominated and non-dominated solution, in which the non-dominated solution will constitute the called Pareto Front, which represents the possible solutions of a multiobjective optimization problem. To better explain this concepts, first, it is necessary to define the multi-objective optimization problems [11], as: minimize {f1 (x), f2 (x), ..., fk (x},

subject to x ∈ S.

(6)

The multi-objective optimization problem is handle of the form (6) involving k (≥ 2) conflicting objective functions fi : Rn → R that is intended to be

Production Scheduling Using Multi-objective Optimization

123

simultaneously minimized. The decision (variable) vector x = (x1 , x2 , ..., xn )T belong to the nonempty feasible region S ⊂ Rn . Objective vector is the image of decision vector and consist of objective function values z = f (x) = (f1 (x), f2 (x), ..., fk (x))T . Furthermore, the image of the feasible region in the objective space is denoted as feasible objective region Z = f (S) [11]. In multiobjective optimization, objective vectors are considered as optimal if none of their components can be improved without damage at least one of the other components. Mathematically, a decision vector x ∈ S is considered a Pareto optimal if there does not exist another x ∈ S such that fi (x) ≥ fi (x ) for all i = 1, ..., k and fj (x) < fj (x ) for at least one index j [11]. In this approach, the Non-dominated Sorting Genetic Algorithm II (NSGA) was used, through the gamultiobj function implemented on MATLAB Software, to create a set of points on the Pareto front by a controlled elitist genetic algorithm [12]. The elitism is done by controlling the elite members of the population as the algorithm progresses in order to maintain the diversity of population until to converge to an optimal Pareto front. In this work, the Pareto front was evaluated by cluster techniques to analyze the solutions similarities and dissimilarities. The k-means algorithm [13], which is the simplest, most popular and generally the most widely partitioning algorithm was applied. The k-means method assigns the solutions to the class with the nearest centroid. Thence, at the beginning of this algorithm, a set of k points is selected, representatives of the classes or centroids [14]. In this sense, the kmeans clustering algorithm was trained in order to find a structure (patterns) in the data that allows the data division into k groups (clusters). As k-means is a partition method, to start the algorithm it is necessary to define the number of partition, it is the number of cluster k that will be built. Therefore, the silhouette method [15] was used to evaluate the number of cluster that the Pareto front will be divided.

4

Conceptualization and Mathematical Model

When there are multiple machines with similar functionalities and these machines can work simultaneously without affecting each other, the problem is considered a parallel machine scheduling problem [16]. Based on the features of the employed machines, parallel machine scheduling can be further classified into identical and unrelated parallel machine scheduling problems. The machines considered in identical parallel machine scheduling problems are homogeneous, and the processing time of a job at any machine is identical [16]. Whereas, the unrelated parallel machine scheduling problem aims to assign a set of jobs to a set of unrelated machines that can process the jobs in parallel without affecting each other [16], having each job different processing times. This model is addressed to simulate an industrial environment composed of m unrelated parallel machine that must execute n jobs. Multi-objective strategies allow combining different objectives in the same model, thus the five performance measures are presented in Sect. 2 considering to three objective functions. The following parameters are necessary to define the objective functions.

124

B. F. Azevedo et al.

Parameters Definition • • • • • • • •

I is the number of machines available i ∈ I = {1, ..., m}; J is the number of jobs, j ∈ J = {1, ..., n}; Cj defines the makespan; Tj defines the tardiness of the jobs; Ej defines the earliness of the jobs; N Lj is the number of tardy jobs; N Ej is the number of early jobs; α and β are penalty parameters, with α, β ∈ Z+ .

Objective Functions The first objective function, denoted by (7), refers to the makespan, which aims to minimize the maximum execution time of the jobs j, considering each machine i ∈ I = {1, ..., m}. min f1 (x) = max (Cj )

(7)

x

The second objective function, denoted by (8), aims to minimize the number of jobs concluded tardily and the jobs concluded early. Once the quantity of late jobs is more critical than quantity of early jobs, a penalty parameter, α, is assigned to weight more the tardiness of jobs than the earliness of jobs. min f2 (x) = α x

m 

N Lj +

j=1

m 

N Ej

(8)

j=1

Finally, the third objective function, defined by (9), aims to minimize the tardiness and the earliness time of the jobs. The first part of the equation refers to the positive difference between the conclusion date of the job and its due time. And the second part refers to the negative difference between both measures. Similar to Eq. (8), a penalty parameter, β, is used to penalize more the tardiness times than the earliness times. min f3 (x) = β x

5

m 

max(Tj , 0) +

j=1

m 

max(−Tj , 0)

(9)

j=1

Results and Discussion

This work considers a machines simulation in an unrelated parallel machine environment. For this, 5 machines and 50 jobs are considered to evaluate the model developed. The processing time of each job on each machine and the due time, was randomly generated, considering uniform distribution between 1 and 50 (with integer values). Thus, the times considered are presented in Fig. 1, as well as the due time, dj , of each job. The parameters α and β were arbitrary chosen, being α = 5 and β = 2. The results of the optimization model was obtained using an Inter(R) i5(R) CPU @1.60 GHz with 8 GB of RAM, using MATLAB software1 . The algorithm was 1

https://www.mathworks.com, 2019a.

Production Scheduling Using Multi-objective Optimization

125

Fig. 1. Processing time and due time of each job

executed 30 times and the Pareto front solution of each execution was compared each other to evaluate the dominated and non dominated solution among all execution. At the end, a final Pareto front was obtained, as represented by Fig. 2, in which the blue points represent the dominated solution (982 points) and the red points (68 points) represents the non-dominated solution. It is important to mention that some points denotes equal solutions, once the algorithm was executed many times and the solutions were compared to each other, from different iterations.

Fig. 2. Pareto front

The points that represent the non-dominated solutions (in red), correspond to 25 different solutions, are equally optimum. However each of them prioritize more one objective than the others. Thus, it is not possible to define a single optimal solution, the final choice of a solution being up to the manager, based on the weighting of the importance of the objectives and the needs of the production. To assist in the final decision and also with the objective of analyzing patterns of similarity and dissimilarity between the optimal solutions, the Pareto front obtained was submitted to cluster analysis. In this case, the k-means algorithm

126

B. F. Azevedo et al.

was used, and the silhouette method indicated 3 as optimum cluster division number. Figure 3 illustrates the Pareto Front cluster obtained.

Fig. 3. Pareto front cluster

Through the Fig. 3 it is noticeable that cluster 1 presents better solution in terms of the first and the third objective functions; but for the second objective function the solutions were not so good, in relation to the solution of the cluster 2 and 3. When it is necessary to prioritize the objective function 2, the better option is to chose a solution that belongs to cluster 2, however this cluster presents the worst solution in terms of objective 1 and 3. Cluster 1 presents more disperse points than the other, meanly in relation to objective 2, which promotes a gamma of options to weight the objective 2 with small modification in the values of objective 1 and 3. On the other hand, the solutions of cluster 2 and 3 are more compacted, presenting less variability of the values. As can be seen, it is extremely hard to define the final choice that enables to better accomplish current needs or preferences in the production system. Among the possible solutions defined by the Pareto front, 3 of them were chosen to be explored in detailed. Each of these solutions prioritize one of the objectives function considered. The solution 1 results in f (x1 , x2 , x3 ) = (216, 222, 8183), which is the best solution in terms of objective function 1, it is this solution that has the smaller x1 , corresponds to the makespan minimization, among all other optimal solutions. The Gantt chart of this solution is shown in Fig. 4, begin 216 the maximum makespan of the solution. As it can be seen the other machines are close, which means that the machines in general are working at similar times. In relation to objective function 2, the solution presented the number of tardy jobs equal to 7 for the machine 1 and equal to 9 for all other machines; and number of the early jobs equal to 1 for machines 1, 2 and 5 and equal to 2 for machines 3 and 4. Regarding objective 3, the solution presents the tardiness time, in minutes, equal to 822, 932, 711, 738 and 848, and the earliness time, also in minutes, equal to 7, 14, 31, 26 and 3, both for machines 1, 2, 3, 4 and 5, respectively.

Production Scheduling Using Multi-objective Optimization

127

Fig. 4. Makespan of solution 1

The solution 2 results in f (x1 , x2 , x3 ) = (393, 208, 14005), which is the best solution in terms of objective function 2. The solution presented 1, 12, 11, 13 and 3 tardy jobs and 0, 2, 1, 2 and 3 early jobs for machines 1, 2, 3, 4 and 5, respectively. In relation to objective 3 the tardiness times are equal to 48, 1966, 2455, 2333 and 167, and the earliness time equal to 0, 17, 13, 26 and 11, respectively for machines 1, 2, 3, 4 and 5. Regarding to objective 1, the Gantt chart of this solution is shown in Fig. 5, begin 393 the makespan of the solution, however, in this case the maximum completion time of the other machines are not close, generating long time for some machine, such as machines 1 and 5.

Fig. 5. Makespan of solution 2

The solution 3 results in f (x1 , x2 , x3 ) = (233, 234, 7816), which is the best solution in terms of objective function 3. The solution presented the tardiness time of 939, 890, 843, 611 and 600, and the earliness time of 0, 14, 2, 14 and 20, respectively for machines 1, 2, 3, 4 and 5. In relation to the number of tardy and early jobs the solution presented 8 tardy job for machines 1 and 5 and 10 tardiness jobs for the others machines. Regarding to the number of early jobs, the machine 1 had no early jobs, while all other machines had 1 early job. About objective 1, the Grantt chart of this solution is shown in Fig. 6, begin 233 the makespan of the solution, which represents a better balance than solution 2, but worst than solution 1.

128

B. F. Azevedo et al.

Fig. 6. Makespan of solution 3

6

Conclusions and Future Work

This work was addressed to explore production scheduling performance measures through a multi-objective approach to support managers in decision making regarding to the jobs scheduling in a set of unrelated parallel machines that compose a production environment. Though the multi-objective optimization, using NSGA algorithm, it was possible to find a production schedule that satisfies different and contradictory objectives based on performance measures that quantify objectives of the schedule. Using the cluster analyze, based on k-means algorithm, it was possible to better explore the solution generation by clustering and analyze the similarities and dissimilarities of the Pareto front solutions. The approach proved to be effective supporting scheduling decision making, in order to consider the most important performance measures in terms of production scheduling preferences. It is noteworthy that the final decision of the optimum solution is up to the manager, who, through the presented methodology, can define different solutions according to the priority of the production system dynamics. As suggestion for future path of this work, it is intended to apply the methodology and the model in a real problem, and also to explore other multi objective algorithms to compare different Pareto front solutions. Acknowledgments. This work has been supported by FCT - Funda¸ca ˜o para a Ciˆencia e Tecnologia within the R&D Units Project Scope: UIDB/00319/2020. Beatriz Flamia Azevedo is supported by FCT Grant Reference SFRH/BD/07427/2021, EXPL/EME-SIS/1224/2021.

References 1. Ojstersek, R., Brezocnik, M., Buchmeister, B.: Multi-objective optimization of production scheduling with evolutionary computation: A review. Int. J. Ind. Eng. Comput. 11(3), 359–376 (2020) 2. Zhang, S., Tang, F., Li, X., Liu, J., Zhang, B.: A hybrid multi-objective approach for real-time flexible production scheduling and rescheduling under dynamic environment in industry 4.0 context. Comput. Oper. Res. 132, 105267 (2021) 3. Pantuza J´ unior, G.: A multi-objective approach to the scheduling problem with workers allocation. Gest˜ ao & Produ¸ca ˜o 23, 132–145 (2016) 4. Behnamian, J., Mohammad, S., Ghomi, T.F.: Multi-objective multi-factory scheduling. RAIRO. Oper. Res. 55, S1447–S1467 (2021)

Production Scheduling Using Multi-objective Optimization

129

5. Varela, M.L.R., Ribeiro, R.A.: Distributed manufacturing scheduling based on a dynamic multi-criteria decision model. In: Zadeh, L., Abbasov, A., Yager, R., Shahbazova, S., Reformat, M. (eds.) Recent Developments and New Directions in Soft Computing. Studies in Fuzziness and Soft Computing, vol. 317, pp 81-93. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-06323-2 6 6. Santos, A.S., Madureira, A.M., Varela, M.L.R.: An ordered heuristic for the allocation of resources in unrelated parallel-machines. Int. J. Ind. Eng. Comput. 6(2), 145–156 (2015) 7. Pinedo, M.L.: Scheduling: Theory, Algorithms, and Systems. Springer, Boston (2012). https://doi.org/10.1007/978-1-4614-2361-4 8. Reis, P.C.S.O.: Ferramenta de apoio ao escalonamento da produ¸ca ˜o. Master’s thesis, Instituto Superior de Engenharia do Porto - Departamento de Engenharia Mecˆ anica (2020) 9. Chang, K.: Chapter 5 - multiobjective optimization and advanced topics. In: Chang, K.-H. (ed.) Design Theory and Methods Using CAD/CAE, pp. 325–406. Academic Press, Boston (2015) 10. Zheng, Y., Ling, H., Xue, J., Chen, S.: Population classification in fire evacuation: a multiobjective particle swarm optimization approach. IEEE Trans. Evol. Comput. 18(1), 70–81 (2014) 11. Lotov, A.V., Miettinen, K.: Visualizing the pareto frontier. In: Branke, J., Deb, K., Miettinen, K., Slowi´ nski, R. (eds.) Multiobjective Optimization. LNCS, vol. 5252, pp. 213–243. Springer, Heidelberg (2008). https://doi.org/10.1007/978-3540-88908-3 9 12. Kalyanmoy, D.: Multi-Objective Optimization Using Evolutionary Algorithms. Wiley, New York (2001) 13. Yedla, M., Pathakota, S.R., Srinivasa, T.M.: Enhancing k-means clustering algorithm with improved initial center. Int. J. Comput. Sci. Inf. Technol. 1(2), 121–125 (2010) 14. Jain, A.K.: Data clustering: 50 years beyond k-means. Pattern Recogn. Lett. 31(8), 651–666 (2010) 15. Rousseeuw, P.J.: Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. J. Comput. Appl. Math. 20, 53–65 (1987) 16. Lin, D.-Y., Huang, T.-Y.: A hybrid metaheuristic for the unrelated parallel machine scheduling problem. Mathematics 9(7), 768 (2021)

Data Prediction Model in Wireless Sensor Networks: A Machine Learning Approach Khushboo Jain1(B)

, Manali Gupta1 , and Ajith Abraham2,3

1 School of Computing, DIT University, Dehradun, UK, India

[email protected]

2 Machine Intelligence Research Labs (MIR Labs), Auburn, Washington 98071, USA

[email protected] 3 Center for Artificial Intelligence, Innopolis University, Innopolis, Russia

Abstract. In resource constraint wireless sensor networks (WSN), an important design concern is the optimization of the data transmission reduction of each sensor node (SN) to extent the overall network lifetime. Numerous cited works claim that the Data Prediction Method (DPM) is the most competent method for data transmission reduction among data aggregation, data regression, neural networks models, spatiotemporal correlation, clustering methods, adaptive sampling and data compression. The big data is generally communicated across the WSN which leads to packet collisions, packet drops and unnecessary energy consumption. Thus, we propose a machine learning model-based on Data Prediction Method (MLM-DPM) to solve these problems. The proposed work is simple yet efficient in terms of processing and needs a small memory footprint in SN. The proposed approach reduces the data transmission rates while maintaining data accuracy. The proposed work is estimated on real dataset attained from the Life Under Your Feet (LUYF) project and compared to two recent Data Prediction Methods. The simulation results were promising and justify the proposed claims. Keywords: Data prediction method · Energy efficiency · Machine learning · Transmission suppression · Wireless sensor network

1 Introduction In WSN, the continuous monitoring applications sense measurements at a very high rate. Continuous data transmissions by SNs at high frequency cause enormous energy investing and thus reduce the network’s lifetime. Energy preservation becomes an essential topic in WSN applications as the SNs are generally battery-equipped [1]. As data transmissions at SNs deplete more energy than any other task [2], data transmission reduction advances further attentions for conserving the scare energy resources [3, 4]. Beside aggregation [5], data prediction methods (DPM) conserve scarce energy resources by avoiding unnecessary data transmissions. In DPM, each SN will train the models with the recent sensed measurement and sends the prediction model to the Base station (BS) or sink. Then, the SNs employ the same DPM as the sink employs to predict the future data. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 130–140, 2022. https://doi.org/10.1007/978-3-030-96299-9_13

Data Prediction Model in Wireless Sensor Networks: A Machine Learning Approach

131

In the literature, researches have been performed to perform data prediction in wireless networks to reduce the number of data transmissions and thus increase the network lifetime. It is observed in most of the research works that the numerous models and techniques proposed for data prediction are based on machine learning, deep learning, regression etc. or their combination with other techniques such as feature extraction, selection, filtering models etc. Zhao et al. [6] proposed a linear-regression prediction model based on the periodic nature of the sensor data which helps to obtain more accurate predicted values over popular data set of 54 sensor nodes deployed in the Intel Berkeley Research laboratory. Matos et al. [7] proposed a strategy based on linear regression to predict sensor data on the basis of data collected from other sensors. The proposed strategy also considered the outlier sensor data values. Raza et al. [8] proposed a novel derivate based prediction strategy which is simple and easy to use for predicting sensor data in wireless sensor networks over a 13 million data-points in four real world problems. Avinash et al. [9] proposed a model based on Kalman filter to predict the data of sensor nodes in WSN. Cheng et al. [10] proposed multiple nodes multiple features based bidirectional long short-term memory (LSTM) to extract features for data prediction using neural network approach. Further, Cheng et al. [11] focused on making the complete use of spatial as well as temporal correlation between sensor data to predict sensor data. They introduced one-dimensional Convolutional Neural Network and Bi-LSTM (Bidirectional Long and Short-Term Memory) to obtain features responsible for first prediction which is executed recursively to get final predicted value. Soleymani et al. [12] proposed a hybrid model using of decision tree, auto-regressive moving average and Kalman filter technique to predict sampling data of sensors for reducing transmissions in WSN. Apart from predicting data at SNs, CHs and sink have also been used for data prediction in wireless networks. Wu et al. [13] applied least-mean-square dual prediction technique over cluster heads to obtain the next value of sensor nodes in WSN. They further tried to optimize the dual prediction technique by minimizing its number of steps required for data prediction. Further, Krishna et al. [14] proposed an autoregressive model of certain order p and applied it to the sensor nodes as well as base stations to predict the next sensor value. The autoregressive model works on the concept that the sensed values use to change slowly and also follow a pattern. Further, in this direction the author proposed an extended regression model in the research work [15] to predict data at sensor node and base station using popular data set of provided by Intel Berkeley Research lab. Further, to establish correlation between senor data values and thus reduce overhead of prediction large number of values, the author proposed a buffer-based linear filter model in [16] for data prediction. Moreover, to synchronize the predicted sensor data the author adopted a two-vector model using Extended-cosine regression approach [17]. The synchronization helps to reduce error due to continuous prediction which in turn reduces the number of transmissions within the network. To reduce the application overhead, the author proposed a light-weight novel approach called as Data Transmission Reduction Method (DTRM) which employ a dual prediction technique between CHs and sink to predict sensor values. The proposed light-weight DTRM is a model which is based on dual prediction and data aggregation [18].

132

K. Jain et al.

The contributions of the proposed work are as follow: • In-depth analysis of literature work of various data prediction techniques. • Design and implementation of Machine Learning Model based Data Prediction Method (MLM-DPM) for WSNs. • Simulation-based validation of the proposed work (MLM-DPM) on real application dataset along with the two related work to determine the performance attained by the proposed algorithm. The remaining of this work is structured as follows. Section 2 presents the Machine Learning Model based Data Prediction Method (MLM-DPM). Section 3 discusses the implementation of the MLM-DPM for WSN based real applications. The simulation results and discussion are illustrated in Sect. 4. Finally, Sect. 5 concludes the work with future direction.

2 Machine Learning Model Based on Data Prediction Method (MLM-DPM) In this section, we present a Machine Learning Model based Data Prediction Method (MLM-DPM) that examines the data collected from the SNs and determine the pattern from it, to predict the future data. An identical DPM is employed at both SNs and at the BS. The SNs and sink will regularly make the prediction of forthcoming observations on the basis of the same historical data. This method allows SN to avoid unnecessary data transmissions to the BS, while the predictions are within threshold or error limit. The sink assumes that the predicted value of SN reflects the actual value of SN till its value gets updated by SN. The algorithm works in a streamlined manner: as soon as the sink receives SN value it will move in the training phase for detecting the pattern of the environmental attributes, which is examined. As long as variance of that attribute remains constant over time, a trend will be detected. For instance, if the feature i.e., ‘temperature’ declines from 32 °C to 31 °C to 30 °C, then it means that the decreasing trend is 1 °C per timeslot. The SN transmits this trend to the sink along with the “last value” of the learning process once a trend is known. Therefore, not all the sensed data is sent to the sink unless a decisive variation is found based on error limit. The SN and the sink will initiate to predict the next value based on the ‘last value’ sent and the trend found. Meanwhile, the sensing process at the SN will remain unchanged. The SN sends data to the sink for each newly sensed measurement, which matches the actual data with the predicted data. When any variation is perceived, the SN transmits that value to the SINK as a sign that some variation is found. Then again, the sink will try to perceive the new trend and will repeat the training process. Though the sink has more powerful memory and is capable of storing huge amount of data, the SNs uses a vector to store the ‘last reading’ of all SNs, and as soon as a shift occurs, it goes back to the vector to estimate the new trend. In any pattern is found it will send it directly to the sink to start the accurate forecasts. The SN will have to train the model again if any pattern is not observed.

Data Prediction Model in Wireless Sensor Networks: A Machine Learning Approach

133

3 Implementation of Machine learning Model based Data Prediction Method (MLM-DPM) This section proposes the implementation of Machine learning Model based Data Prediction Method (MLM-DPM). We present a stepwise algorithm for SNs to learn and update the proposed model. For the sink we present an algorithm to rebuild and predict the data. 3.1 Training of MLM-DPM In this proposed Machine learning Model based Data Prediction Method (MLM-DPM), only two sensed values are required to build the prediction model and only one value is needed for correction or for updating the model. Initially, the ith SN sends the first two sensed readings ri1 and ri2 of time instance t1 and t2 to the BS. When the sink obtains a “new reading”, it saves it with the corresponding time-stamp in its memory. NR signifies the “new reading”, and tNR is denoted when NR was received. Then, the sink will enter the learning phase for n consecutivedata to find thetrend td in the data. Both the SNsand the sink simultaneously calculates the “trend” by finding the variance between the two readings as follow in Eq. (1) td1 = ri2 − ri1

(1)

After the variance is calculated, the trend denoted as td is used to predict the next data value in the time series. This is done concurrently at the SN and the sink based on of the measurement of the SNs. Because the sensor’s data continues to move slowly and smoothly over time, the readings at neighboring time-ticks are generally very close to 

each-other. Hence, the following Eq. (2) will be used to predict the rik (data reading ri at time k th timeslot). 

rik



= rik−1 + td i

(2)



When the SN compares the predicted data rik with the actual data rik to estimate the difference, if this difference lies between the threshold or prediction error (ε), then the actual data is flushed, and there will be no data transmission considering as a “successful prediction”. When the sink does not obtain data from any SN during that timeslot, it will assume that its predicted data lies within the threshold. Otherwise, if the predictedvalue surpasses the threshold error (ε), then the predicted value will be flushed and the real measurement is sent to the sink. The sink will update td k as calculated in Eq. (3). Again, the previous NR reading is updated by the new received data rik to retain track for possible future updates. td k =

rik − NR tk − tNR

(3)

134

K. Jain et al.

Algorithm 1: Model Training Require: , ( 1 , 2 , … ) , ( 1 , 2 , … , Ensure: ( , μ, ) in WSN for each BS will broadcast to all SNs will sense 1 and 2 Transmit 1 and 2 to BS 2 − 1 1 = 2 N = = 2 ≠ 0 do while SN will do: at time read ̂ = ̂ −1 + μ end while end for

)

3.2 Data Prediction and Model Corrections The sensed data from WSN time-series naturally upsurges, stand-still, and declines, particularly when the sensing attributes change over the time period. Since td k is a linear rate of variation over a precise time interval. Thus, to synchronize and get the stability in the regression line over the data curve, td increased by the times of “correction rate” denoted as μ where 0 ≥ μ ≥ 1. Hereafter, the predicted values are calculated by Eq. (4) as follow instead of Eq. (2). 



rik = rik−1 + μtd k

(4)

The role of “correction factor” is to match the predicted data with the real sensed measurements which is achieved by introducing a penalty in the model, which is defined as correction factor (μ) based on a threshold and response on model accuracy while being unbiased. Unbiased means to find a balance between overestimation and underestimation. Thus, the idea is to reduce the bias in the model as td k . Again, as discussed, the SN will not transmit the actual data to the sink until the predicted data is within the threshold or error budget (ε). Otherwise, when sink is likely to obtain one data reading, and it does not receive it, it decides that the data computed by the prediction model is correct. Though, if a predicted data rin at any time slip is not correct. In such scenario, the SN has to transmit the sensed measurements rin to the sink to rectify the model instantly. This is achieved by Eq. (5) that supervised the “update phase” as presented below. 

rin − NR (5) N To determine the model’s accuracy for the next prediction cycle, the SN and the sink will concurrently estimate the “learning rate” (α) as done in Eq. (6). td n =



r n − rin α= i F

(6)

Data Prediction Model in Wireless Sensor Networks: A Machine Learning Approach

135

The prediction frame (F) produced is defined as the number of successful predicts that are estimated before the update. The smaller the value of α, the more exact will be the value of NR. If each estimation inside F is within the threshold, the α will be always less than ε. Assuming, α doesn’t exceed ε, then both the SN and the sink will compute the percentage P in Eq. (7). P=

α × 100 ε

(7)

If α < 0, the td must be increased by P% so that it better fit the data, thus the ε is increased by P%. Otherwise, if α > 0, td must be reduced by P% as stated below in Eq. (8)  if (α < 0), μnew = μ + (P×μ) new 100 = (8) μ else , μnew = μ − (P×μ) 100

Algorithm 2: Data Prediction and Model Corrections Require: , ( 1 , 2 , … ) , ( 1 , 2 , … , ) Ensure: ( , μ, , ) ≥ | then if | ̂ − to the BS Send − = = =

−̂

if | ≥ | then ×100 = if ( < 0), then ( =μ + μ else =μ− μ end if end if

×μ) 100

( ×μ) 100

end if

4 Simulation Results We conduct simulations on Network Simulator (NS2) to evaluate the performances of the proposed work (MLM-DPM) on real application dataset.

136

K. Jain et al.

4.1 Simulation Setup The work is experimented on real dataset obtained from the Life Under Your Feet (LUYF) project [19] which is collecting data from WSN measuring the environmental conditions of the soil. The LUYF project measure following data: soil temperature, soil water pressure, surface temperature, volumetric soil water content, surface moisture, CO2 flux, light flux, battery voltage, total solar radiation and photo-active radiation. The visualization tools of these projects are grazor, silverKoala and sense Web. For our analysis we have simulated the LUYF project based on the of soil temperatures of four SNs and surface temperatures of ten SNs for the duration of 80 days from the LUYF project. Figure 1 demonstrates the amount of data collected by the LUYF project. The simulation parameters are presented below in Table 1.

Fig. 1. The amount of data collected by the LUYF project [19] deployments

4.2 Transmission Suppression% It can be determined by calculating the ratio of the transmitted data by using any prediction algorithm and original transmitted data without using any prediction algorithm.  Transmission Supression% =

transmitted data by using prediction algorithm Original transmitted data without prediction algorithm

 × 100

(9)

Data Prediction Model in Wireless Sensor Networks: A Machine Learning Approach

137

Table 1. Simulation parameters Representation

Parameter

Cost

E0

Initial energy

100 J

ε

Prediction error

0.1 to 0.9 with the step of 0.1

-

Number of Soil Sensors

4

-

Number of Surface Sensors

10

-

Algorithms

MLM-DPM, ELR, P-PDA

-

Data packet

100 bytes

-

Control packet

48 bytes

{ro1 , ro2 , . . . ro10 }

Data Collection Rounds

10

εmp

Multi-path fading amplifier energy

0.0013 (pJ /bit)/m4

εfs

Free space amplifier energy

10 (pJ /bit)/m2

EDA

Aggregation energy

5 (nJ /bit)/s

EDP

Prediction energy

5 (nJ /bit)/s

ERX

Reception energy

50 nJ /sfor 1 − bit

ETX

Transmission energy

150 nJ /sfor 1 − bit, 10 m

Figure 2 illustrates the results of transmission ratio in percentage by executing the MLM-DPM, ELR and P-PDA ten times by varying the threshold levels. The data transmission is reduced drastically to 22% at threshold level 0.1 and goes till 44% even at higher threshold of 0.9%. The data transmission suppression percentage of ELR and P-PDA is between 30% to 55% and 40% to 77% respectively.

Fig. 2. Transmission Suppression (%) of MLM-DPM, ELR and P-PDA for the threshold levels

138

K. Jain et al.

4.3 Energy Efficiency Since the energy consumption is directly proportional to the number of data transmissions performed by the SNs. The transmission data reduction to the sink would suggestively increases the WSN lifetime. The higher the transmission suppression will be, the less data is communicated over the network and less energy will be depleted. The energy model of this work is based on the work done in research [15].

Fig. 3. Energy consumption of MLM-DPM, ELR and P-PDA for the threshold levels

We have compared the energy consumption of our proposed MLM-DPM technique with ELR and P-PDA for the various threshold levels over ten rounds of communication. The results are demonstrated in Fig. 3. The total energy consumption of MLM-DPM is less than 4 J even at the lower threshold level due to high data transmission and approximately 2 J at the higher threshold level. With ELR model the energy consumption is between 3 J to 10 J by varying the threshold level and with P-PDA the energy consumption is between 4.5 J to 14.3 J by varying the level. 4.4 Cost Factor The accuracy of a good DPM should focus on minimizing the difference between the actual data and the predicted data, while being unbiased. Unbiased means to find a balance between overestimation and underestimation. Therefore, to estimate the performance in terms of the predicted time-series reliability at the BS, the root means square error (RMSE) metric is defined as follow:   N 1  × (rin − rin )2 (10) RMSE = N 

i=1

Figure 4 highlights the RMSE for the various threshold levels averaged over ten rounds of data communications. We have compared the RMSE of our MLM-DPM, ELR

Data Prediction Model in Wireless Sensor Networks: A Machine Learning Approach

139

Fig. 4. RMSE of MLM-DPM, ELR and P-PDA for the threshold levels

and P-PDA compared with P-PDA and ELR, MLM-DPM has attained RMSE less than 0.27 which less than 1% even by varying the threshold levels. The MSE of ELR and P-PDA is less than 0.4 and 0.45 respectively, even at the higher threshold limits but it is comparatively higher than the proposed MLM-DPM. Altogether, MLM-DPM delivers much better performance metrics with reasonable error.

5 Conclusion This paper puts forward a Machine learning based Data Prediction Method applied to a real-world WSNs. To improve energy efficiency by avoiding unnecessary data transmission and maintaining data quality by reducing the cost function, MLM-DPM algorithm is trained based on current data values and is employed to predict the future values as well as to regenerate the historical data. An algorithm based on machine learning based data prediction is proposed for model training in SNs to avoid complex computations and to improve the viability of the proposed work. Another algorithm for data prediction and model correction is proposed for the sink. The simulation results of this proposed MLM-DPM algorithm can (1) provide enhanced energy efficiency, (2) higher transmission suppression ratio and (3) reduced cost function for the real time applications of WSN.

References 1. Agarwal, A., Dev, A., Jain, K.: Prolonging sensor network lifetime by using energy-efficient cluster-based scheduling. Int. J. Sci. Technol. Res. ISSN 2277–8616, 9(4) (2020) 2. Jain, K., Singh, A.: An empirical cluster head selection and data aggregation scheme for a fault-tolerant sensor network. Int. J. Distrib. Syst. Technol. (IJDST) 12(3), 27–47 (2021) 3. Gupta, M., Varma, S.: Optimal placement of UAVs of an aerial mesh network in an emergency situation. J. Ambient. Intell. Humaniz. Comput. 12(1), 343–358 (2020). https://doi.org/10. 1007/s12652-020-01976-2

140

K. Jain et al.

4. Gupta, M., Varma, S.: Optimal placement of UAVs forming aerial mesh networks to handle network issues. Adhoc & Sensor Wireless Networks. 48 (2020) 5. Jain, K., Kumar, A.: Energy-efficient data-aggregation technique for correlated spatial and temporal data in cluster-based sensor networks. Int. J. Bus. Data Commun. Netw. (IJBDCN) 16(2), 53–68 (2020) 6. Zhao, J., Liu, H., Li, Z., Li, W.: Periodic data prediction algorithm in wireless sensor networks. In: China Conference on Wireless Sensor Networks, pp. 695–701. Springer, Berlin, Heidelberg (October 2012) 7. Matos, T.B., Brayner, A., Maia, J.E.B.: Towards in-network data prediction in wireless sensor networks. In: Proceedings of the 2010 ACM Symposium on Applied Computing, pp. 592–596 (March 2010) 8. Raza, U., Camerra, A., Murphy, A.L., Palpanas, T., Picco, G.P.: Practical data prediction for real-world wireless sensor networks. IEEE Trans. Knowl. Data Eng. 27(8), 2231–2244 (2015) 9. Avinash, R.A., et al.: Data prediction in wireless sensor networks using Kalman filter. In: 2015 International Conference on Smart Sensors and Systems (IC-SSS), pp. 1–4. IEEE (December 2015) 10. Cheng, H., Xie, Z., Wu, L., Yu, Z., Li, R.: Data prediction model in wireless sensor networks based on bidirectional LSTM. EURASIP J. Wirel. Commun. Netw. 2019(1), 1–12 (2019). https://doi.org/10.1186/s13638-019-1511-4 11. Cheng, H., Xie, Z., Shi, Y., Xiong, N.: Multi-step data prediction in wireless sensor networks based on one-dimensional CNN and bidirectional LSTM. IEEE Access 7, 117883–117896 (2019) 12. Soleymani, S.A., et al.: A hybrid prediction model for energy-efficient data collection in wireless sensor networks. Symmetry 12(12), 2024 (2020) 13. Wu, M., Tan, L., Xiong, N.: Data prediction, compression, and recovery in clustered wireless sensor networks for environmental monitoring applications. Inf. Sci. 329, 800–818 (2016) 14. Krishna, G., Singh, S.K., Singh, J.P., Kumar, P.: Energy conservation through data prediction in wireless sensor networks. In: Proceedings of 3rd International Conference on Internet of Things and Connected Technologies (ICIoTCT), pp. 26–27 (May 2018) 15. Jain, K., Agarwal, A., Kumar, A.: A novel data prediction technique based on correlation for data reduction in sensor networks. In: Proceedings of International Conference on Artificial Intelligence and Applications, pp. 595–606. Springer, Singapore (2021) 16. Agarwal, A., Jain, K., Dev, A.: Modeling and analysis of data prediction technique based on linear regression model (DP-LRM) for cluster-based sensor networks. Int. J. Ambient Comput. Intell. (IJACI). 12(4) (2021) 17. Jain, K., Kumar, A.: An energy-efficient prediction model for data aggregation in sensor network. J. Ambient. Intell. Humaniz. Comput. 11(11), 5205–5216 (2020). https://doi.org/ 10.1007/s12652-020-01833-2 18. Jain, K., Kumar, A.: A lightweight data transmission reduction method based on a dual prediction technique for sensor networks. Trans. Emerg. Telecommun. Technol. e4345 (2021) 19. Dataset (Online): Life Under Your Feet (LUYF) project Retrieved March 11. From http://lif eunderyourfeet.org/en/src/ (2021)

A Study of Version Control System in Software Development Management Concerning PLC Environments Domingos Costa1,2(B) , Senhorinha Teixeira1,2 , and Leonilde R. Varela1,2 1 Department of Production and Systems, University of Minho, Braga, Portugal

[email protected], {st,leonilde}@dps.uminho.pt 2 Algoritmi Research Centre, University of Minho, Guimarães, Portugal

Abstract. In this work, a more apt tool for the PLC and robotics industries is presented. This study contextualizes the areas of PLCs and Robotics to better understand their industry and to serve as a link to the thematic of version control systems specific for this industry. This study will then help to understand the version control systems tools available for software development in general, as well as, the logic behind those while showing how they work. Then it will give more focus over the specific VCS tool that was being used in this project, by showing its functionalities and why it does not work for graphical and structured text programming in the PLC and Robotics contexts. Also, it will be possible to have a general overview of the problem, alongside with examples of solutions to control versions in software development for PLCs and Robotics. Keywords: Version control system · Programmable logic controller · Robotic · Software development

1 Introduction With the purpose to analyse Version Control Systems in software development management, in the context of robotic and Programmable Logic Controller (PLC) languages and industries, a deep study of the Version Control System (VCS) concept took place, as well as the main VCS used by project at hand, to see if it is fit to deal in this industrial context. The use of VCS is a common and increasingly necessary practice in the software industry due to the fact that the projects can reach hundreds, thousands or more lines of code. A simple and brief way to explain of what a VCS consists of is by looking to the literal definition of the name, meaning that a VCS is a tool that helps, saves, and monitors the different versions of some project’s software. Even though, and despite the same purposes, there are key differences between these VCS tools which revolve around their capability type to store data and capacity to read and analyse different coding languages. The very first creation of such tool dates around 1972 derived from the need to program on an IBM S/370 computer. This VCS was untitled Source Code Control System (SCCS), and by being the first made it the most relevant VCS until the © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 141–149, 2022. https://doi.org/10.1007/978-3-030-96299-9_14

142

D. Costa et al.

Revision Control System (RCS). When the RCS was launched it allowed the automation in terms of data storing along with identification, merging of files and other processes. Despite this progress it would not support entire projects but single files, the structure was difficult to use therefore the branching function would be avoided, it had a limited support regarding binary files and no possibility to work on a project from multiple locations with different users. These limitations led to the creation of one of the most famous VCS, the concurrent versioning system (CVS). The CVS was fresh new tool with a client/server model where it allowed a more complete use of the branching system, including branch imports. This new model consists of the introduction of the central server to save the main data, while the developers use copies on their computers, implying the need of a network connection when doing operations with the main, such as committing changes. This where the branching starts evolving, which can be considered on every copy created. Since these copies are sustained on the developer’s computer on an individual level, it allows them to change and experiment without affecting the main project, and when necessary, it is possible to merge these changes to the main code saved on the central server. This also means multiple users can actually work on the same file, by using this method. The CVS proved its worth and over the time it was naturally upgraded and replaced by a more modern tool called Subversion (SVN). By basically being an improved CVS, the SVN used the same central server model more adapted to the modern days. Alongside the SVN, a more flexible and revigorated tool appeared which was called Git. Both these tools, as well as their models and methodologies were analysed in detail in Ruparelia (2010). One aspect in common to all these VCSs is that they were born from the need to code the major and classic programming languages, which all share some similarity at some degree. The problem is when it starts to distance from these scenarios and moves towards the robotic and PLC industries, which is essentially the issue that this work wants to study.

2 Contextualization of Programmable Logic Controller Since the purpose of this work is to evaluate the current situation of VCSs and check if there is an optimal tool regarding the programming of PLC, a brief summary concerning the concepts is now given. The PLC is a particular form of microprocessor-based control that uses programmable memory to retain information in the form of instructions that allow the execution of functions in order to control machines and processes. These controllers are designed to be simple and intuitive, to be used by engineers who are not computer experts and who don’t have much knowledge of their languages, being designed so that only programmers can configure and change their programs. The main advantages of PLCs are cost reduction, immunity to electromagnetic noise, ease of configuration and programming, greater process control, online monitoring capability and simple maintenance [2–4]. The term logic is used because programming is primarily concerned with the implementation of logic and switching operations (switching Operations). Input devices (such as switches) and output devices (such as motors), while controlled, are connected to the

A Study of Version Control System in Software Development Management

143

PLC and then the controller monitors inputs and outputs according to the machine or process [5]. PLCs consist of two parts, PLC hardware and programming. In terms of hardware, all PLC systems are composed mainly of the same construction components that detect input data, and subsequently process and control multiple outputs. These include the CPU (central processing unit), power supply, input and output module, programming device and memory. When it comes to PLC programming, the most common way to do this is to design the desired control circuit in the form of a logic diagram and then insert that diagram into a programming terminal. The programming terminal is able to convert the diagram into digital codes and then send them to the PLC where it is stored in memory [2, 4, 6].

3 Version Control System Approach With the advancement of Informatics and complexity in software creation, an increasing need to replicate, reuse and confirm large amounts of code and information has been growing increasingly. Therefore, the process of creating software and storing/using data is aided by the use of code repository and version control services such as Sourceforge, Bitbucket, Gitlab and Github. These tools allow the internal sharing of code and programming tasks among all team members/contributors, remotely and efficiently [7]. But before going into detail about these repository services mentioned, let us start with the databases that represent these same services. A VCS, in its literal sense, has the functionality to manage and control changes in information repositories, from documents to programming codes. In this way, a VCS also has the ability to recover old versions of a certain document, program or file. Regarding the context of the use of VCS currently, and considering the purpose of the project, VCS is used in software development as it allows an automated control of large amounts of source code and documentation. Some advantages include the accurate detection of bugs, the lifting of previous versions, if the current version is corrupt or if you want to create a new route, even allowing several team members to work simultaneously in different parts of the same project. Therefore, a VCS should have the ability to back up and restore versions and save files as they are edited, as well as the ability to sync, share information with the rest of the team, monitor and verify changes along with the responsible username and comments justifying the change. In addition to these functionalities, a good VCS should also be able to make changes in an isolated location for the purpose of tests and drafts, also known as branching, along with the ability to merge, which means reinserting these changes into the original code for the whole team to view [1, 8]. Turning now to the analysis of VCS storage models, these can be of 4 types: Base Snapshot, Weaves, Delta Base and Skip-Delta Base. Starting with the Snapshot model, this one stores each version as a true snapshot, no matter the old versions of that same file. This model also uses run-length encoding to compress the files in order to save space without losing information. Thus, access to the version of a file is always done at a single common level, that is, it does not depend on the history, number of branches or any other detail. VCSs like Git and Bazaar use this model. In relation to the Weaves model, all versions of a file are stored in a single file (hence the Weave terminology), which is completed with Metadata inserted into each block or

144

D. Costa et al.

piece of code to describe the changes in that same block. The extraction mode of each version is therefore done with a sequential reading of the entire file through these blocks until it reaches the intended version, in reverse engineering style. For this reason, this type of storage becomes slower the more lines of code there are and the larger the file is. This results on also being necessary to rewrite the entire file whenever it is updated with a new version. Older versions of Baazar used this type of storage. Continuing for the Delta base model, it stores only the first version and the changes between versions, the latter being a concept entitled of deltas. In this sense, the delta model can save storage space, only having to rebuild the various versions, when requested, based on the deltas. The reconstruction time is varied according to the number of changes. The big difference between this model and Weaves is the speed and quantity stored, because while the Delta model stores only changes, the Weaves model always finds itself rebuilding every file, even with minimal changes. Still within this model one can mention Forward Deltas, which are deltas with the changes to apply to previous versions in order to have the next version. Also, contrary to what was mentioned, there are also Reverse Deltas which are deltas that record the latest full version in order to rebuild previous versions through reverse engineering. Reverse Deltas greatly improves performance because as a file advances in version history, and because as most users use the latest version, it becomes faster to do the process of accessing a version from the last to the intended version than going from the first version. Most of the most popular VCSs, such as Git, SVN, Mercurial use the Deltas-based model. Moving on to the Skip-Deltas model, which is a variant from the aforementioned homologous name model as it is a storage typology that records the changes between versions in the same way, with the particularity of instead of performing a complete reverse engineering, going through all the versions until the desired version is reached, this model analyses the target version with current version, making a calculation of which deltas or versions will actually have to go through to reach the target version. Then occurs the Delta-Skip, or the advance of deltas, being an even more efficient process in terms of time. SVN is a classic VCS that makes use of this model [8, 9]. The next aspect to mention in terms of VCS, which is also the main topic to highlight when comparing the many tools available, regards its centrality that can be a centralized or distributed model. Both of these are widely used in modern times, with the centralized model being the most common. The centralized version control system (CVCS) is based around a single, central, private server for all users, which can be seen on Fig. 1. This model allows cooperation in the same system, keeping the main copy of the files and their history, as well as the control of changes of those files when saving everything in the local repository or server, being a simple and easy to use tool. The term “centralized”, as already mentioned, derives from the fact that there is only one storage server where users can access through their local computers online. The server keeps track of all changes, while users locally keep a copy of a number of documents. Users can modify the latest version available on the server, causing these changes to be automatically shared by all other programmers. There is also the possibility to work in parallel with the local copy, synchronizing directly with the server. However, due to the fact that the correct final version of a project is in this single server, this model conditions access to the server, and only authorized users can

A Study of Version Control System in Software Development Management

145

commit modifications, while others may be allowed to read-only or not have access at all. Some problems derived from this model go through the inability to merge files on the server or even total loss of documents if the server is inaccessible or corrupt. SVN is a classic tool that uses this model.

Fig. 1. CVCS model, adapted from [10].

On the other hand, the Distributed Version Control System (DVCS) allows for the existence of a local repository for each user on their computer, and no central server is required to record all information, as it can be seen on Fig. 2. At the same time this model works in two directions, meaning that the modifications that were made, the files or the local repository can also be synchronized with the server whenever necessary, in order to share new versions with other users. Thus, it is possible to commit and collaborate with other users simultaneously, within the same project, resulting in faster tasks compared to CVCS. It should also be noted that any repository can be cloned, and there is no repository more important than another. However, for practical purposes, it is normal to choose a repository to be the “central” repository, only to create a standard hierarchy of organization for the remaining repositories. This more modern model has given rise to several tools, such as Mercurial, Git and Bazaar, which have been adopted by various open source software. Thus, the DVCS is more suitable for large projects with more independent programmers, with the advantages of this model being greater speed of operations, flexibility, the no need for remote or online access which allows working offline, the use or embedding of hosting services and the greater ease in merging and branching. While the disadvantages are about the inability to numbering and identification of versions and the fact that it does not have a central server itself which may hinder the use of backup if necessary [10, 11]. 3.1 Study Target of the VCS Tool – Apache Subversion For this project, as the SVN is the main tool currently used in the processes, a study was conducted to better understand it.

146

D. Costa et al.

Fig. 2. DVCS model adapted from [10].

Apache Subversion (SVN) came into being in 2000 by Collabnet, Inc.4, to replace the Concurrent Versions System (CVS) due to design failures that the latter featured, such as not accepting atomic commits, which are a set of modifications that may be constituted by several lines of code, considering these as a single operation. SVN became a selfhosting in August 2001, with the entire CVS team migrating to SVN. SVN has adopted a centralized architecture, in a single central server where it stores all the projects’ multimedia, allowing programmers to use SVN and have limited view of the data on their local machines. A special feature of SVN is the use of Skip-deltas. In relation to commits, they often arrive at the repository only after being examined and they are visible to all users, while the rest of the activity being largely invisible. When a commit fails its commit, the server returns to the last good state and the programmer has the ability to repair the conflict present in his file without affecting the rest. Another detail, is the fact that SVN accepts Atomic commits, thus allowing numerical assignment in the revision of the global repository. In other words, if x is the number of a revision after a commit and the commit itself, it can be declared that revision x corresponds to the state of the server after commit x. There are also 3 terms named trunk, branches and tags, which the SVN does not make any differentiation between, resulting that the only reason to an eventual attempt of differentiation is due to the agreement between programmers. That being said, the recommended structure would be to use these terms in SVN, to use trunk for the main area of development of the current work, then branches to be created from the trunk and representing the versions already approved and tested, and finally the tags which will correspond to a temporal snapshot within the trunk or branch to save the project status on a specific date or mark revisions [9, 12–14]. Due to the fact that the nature of this project was born from the necessity of finding a solution to better control versions between robotic languages, and also that SVN was the currently VCS used, it was immediately concluded the incapability on resorting to the SVN for this kind of endeavour. Following this analysis was then proceeded with the appraisal of other open source VCS such as Git and Mercurial. Despite these VCSs having a different repository model, respectively a distributed model, they still share a general set

A Study of Version Control System in Software Development Management

147

of functionalities that are somewhat similar to the SVN. These functionalities between each VCS have their pros and cons concerning the situation and the typical programming languages like C++ or python, though when talking about graphical programming or structured text programming oriented to PLC and robotic, neither options showed good capacity in supporting this kind of programming. As for example, this can be easily tested by simply inserting a graphical language, like the SIMATIC STEP 7 shown on Fig. 4, onto one of these mentioned open source VCSs and will immediately show the impossibility to either detect and change the code itself. By consequence, it also makes impossible to proceed with the control between versions. From this situation it was then concluded that none is optimal in the context of robotic/PLC. Nonetheless, two private VCS tools specifically made for this type of industry were found, the Versiondog and MDT Autosave [15]. 3.2 Optimal VCS for Programmable Logic Controllers and Robotics As mentioned before, the non-existence of an optimal VCS when programming on PLC and robotic languages is evident, which forces the expanse of the scope towards outside the classical VCSs. Therefore, and while investigating through chats and open discussions between specialists from software websites, such as https://stackoverflow. com/, and by analysing the literature, two VCS tools were always mentioned. These tools are the Versiondog and MDT Autosave [15]. After a simple analysis, followed by a meeting with Versiondog’s company, Auvesy, it was confirmed that this tool would support all the PLC and robotic languages necessary for this project. Also, a list of some of the general languages supported by this tool can be seen directly on its website [16]. Concerning the tool MDT Autosave, despite being capable as well, was not considered only by mere preference. The way Versiondog works is somewhat on the middle term between a central and distributed system, or between a SVN and Git tools for a more practical comparison. That said the Versiondog tool consists of revolves itself around a central server controlled by Auvesy which store all the data, including all versions, branches and projects. After experimenting directly, the demo provided, it was possible to have better overview and actually verify all functionalities and permitted languages. From this experiment it was noticeable the way the central server interacts with the user. By starting the tool interface, consider Fig. 3 and Fig. 4, basically when the user wants to checkout some version it just to select and the tool will immediately download it. After concluded the changes it can then upload which then it will be analysed to detect if there were in fact changes and then creating a new version in that case. It is also possible to check the user and his action on the files, as well as the date of said action. Aside from these functionalities, the most useful one is by far the capacity to literally observe the differences between versions, which is a difficult task when using the classical VCSs. There are also some more advanced tools, such as an automated control of robots and PLCs, where the central server will constantly control and maintain the code and check for differences and errors over time, while alerting the user via message or email in that scenario.

148

D. Costa et al.

Fig. 3. Versiondog interface example.

Fig. 4. Example of comparison between software versions.

4 Conclusion After all the analysis, it was concluded that working in a PLC/Robotic based industry is a modern grey area, where the use of the classic VCSs is not feasible. Even though, these VCSs, both models and tools, are a good foundation to understand the concepts of how a VCS effectively works. At the same time, by looking away from the open source tools and towards the private ones, it was possible to find specialized VCS. This project focused on the Versiondog’s Auvesy, as the company was already contacted, and the tests done with the software presented good starting results. This is great tool to start using in this area and to further study. Some future tests, not just concerning robotic and PLC, but also CAD, Microsoft office or others might be tested and analysed too, as the VCSs, Versiondog included, keep evolving everyday towards all sorts of files and languages.

A Study of Version Control System in Software Development Management

149

Acknowledgement. The project is funded by the FCT—Fundação para a Ciência e Tecnologia through the R&D Units Project Scope: UIDB/00319/2020, and EXPL/EME-SIS/1224/2021.

References 1. Ruparelia, N.B.: The history of version control. ACM SIGSOFT Softw. Eng. Notes 35(1), 5–9 (2010) 2. Lashin, M.M.: Different applications of programmable logic controller (PLC). Int. J. Comput. Sci. Eng. Inf. Technol. 4(1) (2014) 3. Burdea, G.C., Member, S.: Invited review: the synergy between virtual reality and robotics. 15(3), 400–410 (1999) 4. Alphonsus, E.R., Abdullah, M.O.: A review on the applications of programmable logic controllers (PLCs). Renew. Sustain. Energy Rev. 60, 1185–1205 (2016) 5. Bolton, W.: Mechatronics electronic control systems in mechanical and electrical engineering. Four Editi, England (2008) 6. Kilian, C.T.: Modern Control Technology: Components and Systems, 2nd ed. (2001) 7. Perez-Riverol, Y., et al.: Ten simple rules for taking advantage of git and GitHub. PLOS Comput. Biol. 12(7), e1004947 (2016). https://doi.org/10.1371/journal.pcbi.1004947 8. Chen, B., Curtmola, R.: Auditable Version Control Systems (2014) 9. Knittl-Frank, D.: Analysis and comparison of distributed version control systems. Bachelorarbeit. Univ. Appl. Sci. Up. Austria (2010) 10. Khleel, N.A.A., Nehéz, K.: Comparison of version control system tools. Multidiszcip. Tudományok 10(3), 61–69 (2020) 11. Zolkifli, N.N., Ngah, A., Deraman, A.: Version control system: a review. Procedia Comput. Sci. 135, 408–415 (2018) 12. SVN APACHE: Skip-Deltas in Subversion. [Online]. Available: http://svn.apache.org/repos/ asf/subversion/trunk/notes/skip-deltas (2002). Accessed 26 May 2021 13. Collins-Sussman, B., Fitzpatrick, B.W., Pilato, C.M.: Version Control with Subversion for Subversion 1.7 (Compiled from r4991). vol. 1, pp. 468 (2011) 14. The Nile Team: Version Control System — Wiser 0.1 Documentation. [Online]. Available: https://chiplicity.readthedocs.io/en/latest/On_Software/VersionControlSystem.html (2014). Accessed 05 Apr 2021] 15. Khudyakov, P.Y., Kisel’Nikov, A.Y., Startcev, I.M., Kovalev, A.A.: Version control system of CAD documents and PLC projects. J. Phys. Conf. Ser. 1015(4) (2018) 16. Auvesy: Version control with versiondog | AUVESY GmbH. [Online]. Available: https://auv esy.com/en/versiondog (2021). Accessed 10 Oct 2021

REGION: Relevant Entropy Graph spatIO-temporal convolutional Network for Pedestrian Trajectory Prediction Naiyao Wang1 , Yukun Wang1 , Changdong Zhou1 , Ajith Abraham2 , and Hongbo Liu1(B) 1

College of Information Science and Technology, Dalian Maritime University, Dalian, Liaoning, China {wny,lhb}@dlmu.edu.cn 2 Machine Intelligence Research Labs (MIR Labs), Auburn, WA, USA

Abstract. Modeling pedestrian interaction is an essential building block in pedestrian trajectory prediction, which raises various challenges such as the complexity of social behavior and the randomness of motion. In this paper, a new relevant entropy spatio-temporal graph convolutional network is proposed to model pedestrian interaction for pedestrian trajectory prediction, which contains regional spatiotemporal graph convolutional neural network and gated dilation causal convolutional neural network. The regional spatio-temporal graph convolutional neural network creates a matching graph structure for each time step, and calculates the weighted adjacency matrix of each graph structure through relevant entropy to obtain the sequence embedding representation of the pedestrian interaction relationship. The gated dilation causal convolutional neural network reduces the linear superposition of the hidden layer through the setting of the dilated factor, and uses the gating mechanism to filter the features. Experiments are carried out on the standard data sets ETH and UCY, higher accuracy and efficiency verify that the proposed method is effective in pedestrian interaction modeling. Keywords: Graph convolution · Relevant entropy mechanism · Trajectory prediction

1

· Gating

Introduction

As an important participant in traffic scenes, pedestrians exhibit highly random movements [1,2]. Predicting the trajectory of pedestrians is of great significance in many traffic fields such as automatic driving and monitoring systems [3,4]. Pedestrians are subjective when making route decisions, and common sense of social rules needs to be followed [5,6]. The motion subject needs to analyze other human’s actions and social behaviors to adjust their own routes. In addition, there exist interactions in the group environment, and individual behavior patterns will be implicitly affected by the surrounding environment [7–9]. Therefore, c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 150–159, 2022. https://doi.org/10.1007/978-3-030-96299-9_15

REGION

151

building a pedestrian interaction model with high interpretability and generalization is the focus of the trajectory prediction problem [10,11]. At present, data-driven methods regard pedestrian trajectory prediction as a time series data prediction problem. Although certain progress has been made, there are still some difficulties needed to be resolved, which can be summarized as the following aspects: 1) The interpretability of the pedestrian trajectory prediction problem is poor. The high-dimensional information in the scene is difficult to fit, the physical meaning is difficult to interpret, and it is not intuitive to model pedestrian scene information. 2) Pedestrian interaction is difficult to model. Pedestrian interaction is subjective. Because it changes dynamically with the transformation of time and space and is affected by potential social rules, the scope of interaction is difficult to define. 3) Pedestrian trajectory prediction is a long-sequence prediction problem, with a large amount of data calculation and a lot of interference information, which is difficult to fit the nonlinear relationship in the time dimension. In this work, we propose a new regional relevant entropy spatiotemporal graph scene modeling method, and on this basis, a pedestrian trajectory prediction model based on graph convolutional neural network is presented. The main body model consists of two main parts: the regional relevant entropy spatiotemporal graph convolutional neural network and the gated dilation causal convolutional neural network. In the regional relevant entropy spatio-temporal graph, the relevant entropy is introduced to calculate the weight of each edge and the weighted adjacency matrix indicates the strength of the mutual influence between pedestrians, then the feature of the pedestrian’s past trajectory is extracted from the matrix by means of convolution operation. By taking the obtained features as input, the gated dilation causal convolutional neural network operates on the time dimension of the embedding result, which reduces the error of the model through a gating mechanism, and finally predicts the multimodal trajectory of all pedestrians in the future. The main contributions of this paper are listed below: – We propose a new regional relevant entropy graph spatio-temporal convolutional network for modeling pedestrian groups in traffic scenes, called the REGION model. The topology of the graph is a natural way to represent the social interaction between the pedestrians in the scene. – A new gated dilation causal convolutional neural network is proposed for time series prediction. It can effectively filter useless features and prevent the problem of gradient disappearance during training. – The proposed model is trained on the standard public data set, and the obtained model is compared with other baseline methods on data and visualization.

152

2

N. Wang et al.

Related Work

Physics-based model predicts movement by simulating a set of well-defined dynamics equations. Zernetsch et al. use a physical model of cyclists containing the driving and resistance forces to predict their future position [12]. Kooij et al. address the problem of predicting the path of objects with multiple dynamic modes by the introducing of the latent variables related to pedestrian awareness [13]. Following the success of Recurrent Neural Network models for sequence prediction tasks, a Long Short-term Memory model which can learn general human movement and predict their future trajectories is proposed in the work of Alahi et al. [14]. Later, the combining of generative adversarial networks provides a new idea to solve the problem of sequence prediction, Gupta et al. propose a recurrent sequence-to-sequence model observes motion histories and predicts future behavior [15]. With the development of graph convolution technology, Mohamed et al. propose spatiotemporal graph convolutional network and temporal extrapolation network to extract features by spatiotemporal convolution on the representation of pedestrian trajectories [16]. The kind of model makes predictions by specifying motion goals and formulating strategies to achieve them. Shen et al. develops a transferable pedestrian motion prediction algorithm based on Inverse Reinforcement Learning (IRL) that infers pedestrian intentions and predicts future trajectories based on observed trajectory [17]. Further research proposes a new general framework for directly extracting a policy from data, as if it is obtained by reinforcement learning following inverse reinforcement learning [18]. By minimizing the symmetrized cross-entropy between the distribution and demonstration data, Rhinehart et al. proposes a method to forecast a vehicle’s ego-motion as a distribution over spatiotemporal paths, conditioned on features embedded in an overhead map [19]. Xie et al. proposes a sequential model that combines the CNN with the LSTM to predict surrounding vehicle trajectories [20]. Zhao et al. proposes a CNN-based model for human pedestrian trajectory prediction with the idea of motion patterns [21]. In order to model the interaction between pedestrians in the scene more reasonably while improving the efficiency of the model, this paper adopts a relevant entropy graph spatio-temporal convolutional network.

3

Methodology

The overall framework of the proposed REGION is presented in Fig. 1. First, take the position of each pedestrian in the dataset that has been marked into the model and construct the regional spatial map structure: the position information of the pedestrian is used as the nodes set in the map structure, the distance information between the two is used as the edge attribute between their nodes if the distance between pedestrians is within the threshold set in advance. The edge set in the graph structure indicates the influence between pedestrians; Secondly, for any node in the graph, calculate the influence relationship coefficient between the two directly adjacent nodes through the relevant entropy and the

REGION

153

weighted adjacency matrix of the graph structure can be formed; Then, expand the regional space map into a regional space-time map and perform convolution operations on the weighted adjacency matrix of the regional spatio-temporal graph at each time for getting the embedding representation of the graph at different times; At last, the obtained embedding result is put into the gated dilation causal convolutional neural network, which predicts on the time dimension to obtain the final trajectory prediction result.

Fig. 1. The overall framework of the proposed REGION.

Since the space-time graph changes gradually with time steps, before constructing the space-time map, this paper first constructs a set of space maps Gt to represent the position of the pedestrian in the scene at any time t in the scene. Given a topological graph Gt =< Vt , Et , Rt >, where Vt = {vti |i ∈ (1, ..., N )} ij is the vertex set, Et = {eij t |i ∈ (1, ..., N )} is the edge set and Rt = {rt |i ∈ (1, ..., N )} is the nodes relevance set. Let S ∈ {s1 , s2 , ..., sn } be the scope of pedestrian interaction at time t, ij } be the nodes relevance of vertexes vti and vtj at time t, if Rt ∈ {r1ij , r2ij , ..., rm given S, the relevant entropy of Rt is defined as Eq. (1): aij t = H(Rt |S) = −

m 

p(rtij |S = ε) log p(rtij |S = ε).

(1)

i,j=1

The calculation method of Rt ∈ {r1ij , r2ij , ..., rtij } needs to consider the distance and direction of pedestrians. We use cosine similarity to calculate it as Eq. (2): rtij =

vti · vtj ||vti ||||vtj ||

.

(2)

All the aij t constitutes a weighted adjacency matrix A, then we need to normalize the matrix A. Through this operation, the data can be made comparable and the relationship between the data can be relatively maintained. We need to

154

N. Wang et al.

multiply A ∈ {A1 , A2 , ..., At } by the degree matrix D−1 and further divide it 1 into two D− 2 , to obtain a symmetric and normalized matrix as Eq. (3): t ∗ D− 2 . At = D− 2 ∗ A 1

1

(3)

t = A + βIN , taking β = 1 (makes the characteristics of the node itself where A t = A + I, and I is the identity matrix. as important as its neighbors), we have A Then, the aggregate information around the target node is extracted through the convolution operation on the graph structure. The convolution operation definition of the regional spatiotemporal graph convolutional neural network can be obtained, as shown in Eq. (4):  P (κ(l) ) · W (v i(l) , v j(l) )). (4) v i(l+1) = σ(λ v j ∈τ

where σ is the PRelu activation function, λ is a normalization item, and P (·) represents the sampling process, which aggregates the surrounding information centered on κ. (l) represents the l layer, and W (·) represents the weight that needs to be trained in the network, τ = {v j |d(v i , v j ) ≤ D} is the set of adjacent vertices of the vertex v i . P (κ(l) ) aggregates the embedded information of neighbor nodes and its own node. We calculate P (κ(l) ) as Eq. (5): κi(l) = ψl (v i(l−1) , μl (v j(l−1) , ∀j ∈ S)).

(5)

where μl (·) is the aggregate function, and ψl (·) is the concat function. In summary, the equation of Regional spatio-temporal graph convolutional neural network is shown in Eq. (6). R(V (l) , At ) = σ(At V (l) W (l) ).

(6)

where W (l) represents the matrix of the trainable parameters of the l layer. After applying the graph convolutional neural network, the features of the graph that can be represented compactly are obtained. The embedding result obtained is expressed as V . Time series prediction requires that the prediction result at a certain time t can only be judged by the input before time t, this paper proposes a gated dilation causal convolutional neural network. The paper defines filter f : {0, 1, ..., k − 1} → R, the input sequence is V = ( v1 , v2 , ..., vt ), it is the embedding result sequence output by the spatiotemporal graph convolutional neural network. The convolution kernel with the dilation factor d at the sequence s is shown in Eq. (7). F (s) =

k−1 

f (i) · vs−di .

(7)

i=0

where k represents the size of the convolution kernel. In this paper, the dilation factor of the input layer is defined as 1, which is the ordinary convolution.

REGION

155

The dilation factor of the second hidden layer is set to 2. As the network layer increases, the expansion factor is 2 increases in exponential form. In order to get a better training effect on the nonlinear relationship in the time dimension, this module sets an output gate on the convolutional layer of the dilated causal convolution, and sets independent training parameters for it. The output of each hidden layer is controlled by the output gate. Carry out regulation. The calculation formula for each hidden layer can be expressed as Eq. (8). hl = ϕ(γ1 Xd + b) ⊗ δ(γ2 Xd + c).

(8)

where δ(γ2 Xd +c) is the output gate of the convolutional layer, through which the mapping ϕ(γ1 Xd +b) can be adjusted to make the weight is more suitable for the prediction model. Where ϕ and δ represent the activation function, respectively, γ1 and γ2 are the weights required to train the model, Xd is the feature in the expanded causal convolution, ⊗ represents the element product operation between the matrices, b and c are bias term.

4

Experiments

In this section, we will introduce experiments to verify the validity of the model. The training efficiency of this model is compared with other methods, which proves the efficiency of this method. And the prediction accuracy and visualization are compared with this model and multiple baseline methods, such as SocialLSTM [14], SocialGAN [15] and SoPhie [22]. The paper uses the average displacement error and the final displacement error to quantitatively analyze the model. All models take 8 frames as input and make predictions for the next 12 frames. The results of the prediction model are shown in Table 1 and Fig. 2. Table 1. ADE/FDE results on trajectory prediction of different methods. Datasets Linear [14] Sophie [22] S-GAN [15] S-LSTM [14] REGION eth

1.35/2.96

0.72/1.45

0.85/1.64

1.10/2.37

0.66/1.13

hotel

0.37/0.69 0.75/1.65

0.70/1.39

0.78/1.74

0.51/0.88

zara1

0.59/1.19

0.33/0.63

0.34/0.67

0.45/0.98

0.36/0.59

zara2

0.75/1.49

0.39/0.76

0.42/0.85

0.52/1.15

0.30/0.61

univ

0.83/1.60

0.52/1.22

0.75/1.50

0.67/1.42

0.47/0.83

avg

0.78/1.59

0.54/1.14

0.61/1.21

0.70/1.53

0.46/0.80

It can be seen from the Table 1 and Fig. 2 that the REGION model studied in this paper constructs a regional spatio-temporal graph and calculates the weighted adjacency matrix on the graph through relevant entropy, so that the

156

N. Wang et al.

Fig. 2. Trajectory evaluation ADE/FDE results on pedestrian trajectory prediction. (a) ADE value of the evaluation results; (b) FDE value of the evaluation results.

effective social interaction information around the target pedestrian is adopted by the network and the interference information is discarded, which has obtained better prediction accuracy on the standard data set, and can accurately predict the movement trend of pedestrians in the future. The paper uses a pre-divided test set to test the model, and provides the visualization of the trajectory prediction results of four scenarios: pedestrians from different directions merging together, pedestrians walking in opposite directions, pedestrians walking side by side in the same direction, and a single pedestrian meeting a group of pedestrians. In view of the above four scenarios, the paper compares the visualization results on trajectory prediction of the SoPhie model, the Social-GAN model and the REGION model in Fig. 3. We also show the visualization results on the multi-modal trajectory predictions of the different model, as shown in Fig. 4, which includes situations where pedestrians catching up in the same direction, individual pedestrians encountering groups in different directions, converge walking, and waiting for turning. After analyzing the evaluation results, it is proved that the REGION model proposed in this paper can predict the multiple walking routes of multiple pedestrians in the scene for a period of time in the future, and the overall prediction trend is consistent with the actual walking path of pedestrians. It can be seen from Fig. 4. SoPhie and REGION give a collision-free path prediction. However, through the visualization results, it can be intuitively observed that the accuracy of the prediction results of the REGION model proposed in this paper is slightly higher than that of the other models, and the path is more similar to the real trajectory of pedestrians.

REGION

157

Fig. 3. Trajectory prediction accuracy visualization results of different methods. The dashed circle represents the regional influence range of each pedestrian. The deeper the blue at the intersection of the two circles is, the greater the degree of influence between the pedestrians will be.

Fig. 4. Visualization of the multi-modal trajectory predictions on different methods.

5

Conclusions

The main research content of this paper is to complete the prediction of the trajectory when a group of pedestrians interact with each other under a fixed monitoring perspective. A regional graph spatio-temporal convolutional neural

158

N. Wang et al.

network algorithm is proposed. The model defines the conditions for the connections between nodes when constructing the regional spatio-temporal graph, so that the modeling of pedestrian groups is more suitable for the actual situation in society. A gated dilation causal convolutional neural network is also proposed, which can obtain good prediction results while ensuring training efficiency. In addition, the dilation of causal convolution allows the receptive field to increase simultaneously with the increase in network depth, so that the model can make accurate predictions of pedestrians’ future walking trajectory trends. The model REGION proposed in this paper generally has a good performance on the data set with moderate crowd density in the scene, but there are errors in the prediction of the data set environment with a single scene and less nonlinear relationship between pedestrians. In the follow-up work, we should adopt different modeling strategies to predict the future trajectory in different complex scenarios, entropy-based energy models will be used to more accurately model the interaction between pedestrians, and transformer-based graph convolutional networks can be introduced to better extract pedestrian features. Acknowledgements. This work is supported in part by the National Natural Science Foundation of China (Grant Nos. 61772102, 62176036) and the Liaoning Collaborative Fund (Grant No. 2020-HYLH-17).

References 1. Zhao, H., Wildes, R.P.: Where are you heading? Dynamic trajectory prediction with expert goal examples. In: Proceedings of the International Conference on Computer Vision, pp. 7629–7638 (2021) 2. Peng, Y., Zhang, G., Li, X., Zheng, L.: STIRNet: a spatial-temporal interactionaware recursive network for human trajectory prediction. In: Proceedings of the International Conference on Computer Vision, pp. 2285–2293 (2021) 3. Cai, Y., et al.: Pedestrian motion trajectory prediction in intelligent driving from far shot first-person perspective video. IEEE Trans. Intell. Transp. Syst. (2021, in press). https://ieeexplore.ieee.org/document/9340008 4. Quan, R., Zhu, L., Wu, Y., Yang, Y.: Holistic LSTM for pedestrian trajectory prediction. IEEE Trans. Image Process. 30, 3229–3239 (2021) 5. Li, Y., Liang, R., Wei, W., Wang, W., Zhou, J., Li, X.: Temporal pyramid network with spatial-temporal attention for pedestrian trajectory prediction. IEEE Trans. Netw. Sci. Eng. (2021, in press). https://ieeexplore.ieee.org/document/9373939 6. Cai, Y., et al.: Environment-attention network for vehicle trajectory prediction. IEEE Trans. Veh. Technol. 70, 11216–11227 (2021) 7. Pang, B., Zhao, T., Xie, X., Wu, Y.N.: Trajectory prediction with latent belief energy-based model. In: Proceedings of the International Conference on Computer Vision and Pattern Recognition, pp. 11814–11824 (2021) 8. Zhang, B., Yuan, C., Wang, T., Liu, H.: STENet: a hybrid spatio-temporal embedding network for human trajectory forecasting. Eng. Appl. Artif. Intell. 106, 104487 (2021) 9. Chen, G., Li, J., Lu, J., Zhou J.: Human trajectory prediction via counterfactual analysis. In Proceedings of the International Conference on Computer Vision, pp. 9824–9833 (2021)

REGION

159

10. Zhang, B., Zhang, R., Bisagno, N., Conci, N., De Natale, F.G.B., Liu, H.: Where are they going? Predicting human behaviors in crowded scenes. ACM Trans. Multimedia Comput. 17, 1–19 (2021) 11. Zhang, B., Wang, N., Zhao, Z., Abraham, A., Liu, H.: Crowd counting based on attention-guided multi-scale fusion networks. Neurocomputing 451, 12–24 (2021) 12. Zernetsch, S., Kohnen, S., Goldhammer, M., Doll, K., Sick, B.: Trajectory prediction of cyclists using a physical model and an artificial neural network. In: 2016 IEEE Intelligent Vehicles Symposium, pp. 833–838. IEEE (2016) 13. Kooij, J.F.P., Flohr, F., Pool, E.A.I., Gavrila, D.M.: Context-based path prediction for targets with switching dynamics. Int. J. Comput. Vis. 127(3), 239–262 (2019). https://doi.org/10.1007/s11263-018-1104-4 14. Alahi, A., Goel, K., Ramanathan, V., Robicquet, A., Fei-Fei, L., Savarese, S.: Social LSTM: human trajectory prediction in crowded spaces. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 961–971 (2016) 15. Gupta, A., Johnson, J., Fei-Fei, L., Savarese, S., Alahi, A.: Social GAN: socially acceptable trajectories with generative adversarial networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2255–2264 (2018) 16. Mohamed, A., Qian, K., Elhoseiny, M., Claudel, C.: Social-STGCNN: a social spatio-temporal graph convolutional neural network for human trajectory prediction. In: Proceedings of the International Conference on Computer Vision and Pattern Recognition, pp. 14424–14432 (2020) 17. Shen, M., Habibi, G., How, J.P.: Transferable pedestrian motion prediction models at intersections. In: 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 4547–4553, IEEE (2018) 18. Ho, J., Ermon, S.: Generative adversarial imitation learning. Adv. Neural Inf. Process. Syst. 29, 4565–4573 (2016) 19. Rhinehart, N., Kitani, K.M., Vernaza, P.: R2P2: a reparameterized pushforward policy for diverse, precise generative path forecasting. In: Proceedings of the European Conference on Computer Vision, pp. 772–788 (2018) 20. Xie, G., Shangguan, A., Fei, R., Ji, W., Ma, W., Hei, X.: Motion trajectory prediction based on a CNN-LSTM sequential model. Sci. China Inf. Sci. 63(11), 212207 (2020). https://doi.org/10.1007/s11432-019-2761-y 21. Zhao, D., Jean, O.: Noticing motion patterns: a temporal CNN with a novel convolution operator for human trajectory prediction. IEEE Robot. Autom. Lett. 6(2), 628–634 (2020) 22. Sadeghian, A., Kosaraju, V., Sadeghian, A., Hirose, N., Rezatofighi, H., Savarese, S.: SoPhie: an attentive GAN for predicting paths compliant to social and physical constraints. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1349–1358 (2019)

An Analysis of Multipath TCP for Improving Network Performance Virendra Dani1(B)

, Sneha Nagar2 , and Vishal Pawar2

1 Computer Science and Engineering, Shri Vaishnav Vidyapeeth Vishwavidyalaya, Indore, India

[email protected]

2 Electronics and Communication Engineering Department, Shivajirao Kadam Institute of

Technology and Management, Indore, India {snehanagar,vishalpawar}@skitm.in Abstract. MP-TCP (Multipath-Transmission Control Protocol) has the potential to greatly improve application performance by employing several routes. Multipath TCP was intended to be a backwards-compatible alternative to TCP. Accordingly, it exposes the normal socket API to applications that can’t manage how the different routes are used. This is a crucial attribute for various functions that are uninformed of the multipath behavior of the network. This, in contrast, is a restriction on application that may advantage from particular information in order to employ numerous paths in a way that best suits their requirements. As a result, hosts are frequently associated by several paths, yet TCP only allows for one path per transport connection. If these many paths could be used simultaneously, network resource use would be more efficient. This should improve the user experience by improving network stability and increasing throughput. Hence, this paper proposed modified MP-TCP which permits a particular data flow to be divides into several paths. This has key advantages in terms of reliability and connection may continue even if one of the paths fails. The Proposed work has been implemented in NS 2 with MPTCP protocol modification. Keywords: MPTCP · Congestion · Subflow · Round trip time · Simulation · NS-2 · IETF · Window size

1 Introduction Internet is a worldwide information platform made up of consistent system networks connected via wired, wireless communications [1], fiber-optic, and, among other things. The Internet’s success is due to its capacity to provide reliable end-to-end data transmission services for a wide range of applications. As a result, every network carrier makes every effort to get packets to their intended destinations [2]. Multipath transport tries to achieve some of resource pooling goals by using several disconnected (or partially disjoint) routes across a network at the same time [3]. The following are the two most important advantages of multipath transportation 1.1 Multipath TCP Protocol Multipath TCP (MPTCP), which allows to efficiently exploit several Internet paths between a pair of hosts, while presenting a single TCP connection to the application © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 160–169, 2022. https://doi.org/10.1007/978-3-030-96299-9_16

An Analysis of Multipath TCP for Improving Network Performance

161

layer. Multipath TCP allows multiple subflows to be set up for a single MPTCP session. An MPTCP session [4] starts with an initial subflow, which is similar to a regular [5, 6]. Multiple subflows can be set up for a single MPTCP session using Multipath TCP [7]. An MPTCP session begins with a primary subflow, which is identical to the one described above for a conventional TCP connection. Additional MPTCP subflows can be created after the first has been established [8]. MPTCP handles underlying TCP connections (called subflows) which convey the real data, while applications communicate via the standard socket API. As illustrated in Fig. 1, MPTCP functions as a shim layer between the socket interface and one or more TCP subflows from an architectural standpoint. It accomplishes this by utilizing TCP options to accomplish to given objectives [9]:

Fig. 1. Stack of MP-TCP

• Create MPTCP connection for the first time. • Subflows can be added to Multi path TCP connection. • Use multi path TCP connection to send data An MPTCP connection is formed by negotiating its usage using the three-way handshake using TCP options. The SYN segment’s MP CAPABLE option indicates that client supports MPTCP. A random key is also included in this option for security concerns. The server will respond with a SYN + ACK segment with MP CAPABLE option if it supports MPTCP. The server generates a random key for this option. The MP CAPABLE option confirms use of MPTCP, as well as keys to allow stateless servers, in third ACK of three-way handshake [10, 11]. 1.2 Congestion Control in Multipath TCP The loss of packets is used as an indicator of congestion in the majority of contemporary multipath TCP congestion management methods. However, it just detects congestion and does not prevent it. If the same host has a multipath connection with two or more routes that are loaded unevenly, MPTCP will favour the unloaded path and push the

162

V. Dani et al.

majority of its traffic there, reducing the load on the congested link while increasing it on the less-congested one [12]. Following are the points to be considered. • If a large enough percentage of flows are multipath, congestion will be distributed evenly across groups of connections, resulting in “resource pools,” or links that function as if they were a single, higher-capacity link shared by all flows. • In the rise phase of congestion management, resource pooling is applied; in this case, Multipath TCP will enable less-crowded subflows to expand proportionally more than congested ones. • Finally, the overall increase of Multipath TCP is dynamically selected across all of its subflows.

2 Related Study Various academics have looked at how Multipath TCP should handle interfaces and subflows that are accessible. Honglin Li et al. [13] presented a delay-based congestion management method. The suggested method consists mostly of two stages. To reduce the delay difference between distinct pathways, a restricted optimization problem is created first. Second, depending on the obtained rate distribution vector and round-trip duration, the authors modify the congestion windows to disperse traffic across each path. Shiva Raj Pokhrel et al. [14] created a complete technique capable of measuring performance of long-lived MPTCP flows with combined Wi-Fi and Cellular network access, while taking into consideration the differences between the two types of networks. Lucas Chaufournier et al. [15] presented a large-scale empirical investigation of MPTCP’s efficacy and practicality in data centre and cloud applications under various network circumstances. Chauvenne R et al. [16] proposed comprehensive measurements and analyses of a large number of experiments. The authors compare the performance of MultiPath TCP against that of TCP using a new metric that they create, and then watch how these performances change when other parameters are changed. Tongguang Zhang et al. [17] studied the problem of data delivery in MANETs in order to improve the quality of service (QoS) and quality of experience (QoE) that users get.

3 Proposed System As previously stated, there are certain difficulties with Multipath TCP that have an impact on the network’s overall performance. Load balancing, retransmission mechanisms, and packet reordering are the primary variables that influence Multipath TCP. This part explains detail work of this paper. 3.1 Problem Scenario The Internet Engineering Task Force defined Multipath TCP (MPTCP) (IETF). It is backward compatible with single-path TCP and enables multi-homing. Main goal of this

An Analysis of Multipath TCP for Improving Network Performance

163

work is to increase performance by aggregating bandwidth over many accessible routes. However, reordering is a problem connected with MPTCP. Because communication between a typical pair of end nodes takes multiple pathways with varied bandwidth and latency characteristics, the traffic between them follows different paths. Packets arriving out of order must wait in the receive buffer before being sent to a specific application in the correct sequence. When out-of-order data exceeds the receive buffer, buffer is stopped, resulting in substantial performance degradation. 3.2 Objectives The specific objectives of this revision are as follows: • • • •

Examine the elements that influence MPTCP’s execution Implement the original MPTCP to evaluate its performance. Modify the elements that impact the original MPTCP’s performance. Compare the performance of original MPTCP with modified MPTCP.

3.3 Methodology We will schedule the data in a manner such that packet arrives in order at the receiver. This will enhance the performance of the existing MPTCP. • We use 2 nodes (in Fig. 2) to construct the Multipath Network Topology with two subflow and three subflow. For example:

Fig. 2. An example of topology subflow

• In the above diagram, we use two nodes and two subflow to construct the Multipath Network Topology. Likewise, we have to construct the same network for two nodes and three subflows. • Selection of source and destination • Simulation with two subflows The total throughput, congestion window, average waiting time, total number of retransmissions will have to analyze for each scenario. Finally generate graph for the performance of MPTCP and the proposed algorithm with MPTCP for the performance metrics:

164

• • • •

V. Dani et al.

Number of Retransmissions Average waiting time per segment Congestion window of each subflow Global throughput by combining the individual subflow’s throughput.

From that result, we prove that our proposed i.e. modified TCP (MPTCP) will achieve efficient result when compared with existing. Similarly we will Simulate for 2 different subflows also and compare the results.

4 Simulation and Results This section provides the simulation depiction of the black-hole attack detection where packets are forward to the legitimate nodes using different number of nodes. 4.1 Simulation Arrangement The suggested study is performed using the NS2 [18] network simulator version 2. Furthermore, the following configuration is available for performance evaluation and simulation. This section depicts a simulation of MPTCP (Table 1). Table 1. Semulation setup definition Simulation properties

Values

Number of interface

1

Link speed

10 MBPS

Queue type

Two ray ground

Agent

FTP

Traffic model

Multipath TCP

Simulation time

20.0 s

Dimension

1000 × 1000

Channel type

Wireless channel

A From the above simulation table, we define different option and their values for implementing MPTCP according to provide efficient simulation. 4.2 Result Discussion Following are the graph of the result obtained from the execution of the proposed and traditional MPTCP for 2 subflow and 3 sublow listed below- (Figs. 3, 4, 5 and 6)

An Analysis of Multipath TCP for Improving Network Performance

For 2 Subflow 50.00012 50.0001 50.00008 50.00006 50.00004 50.00002 50 49.99998 49.99996 49.99994

Modified MP -TCP Old MP-TCP

Window Size

Fig. 3. Comparision of method for window size

0.000269 0.0002685 0.000268 0.0002675

Modified MP-TCP

0.000267

Old MP-TCP

0.0002665 0.000266 0.0002655 Sum of Retransmission

Fig. 4. Comparision of method for no. of retransmission

0.005 0.004 0.003

Modified MP-TCP

0.002

Old MP-TCP

0.001 0 Delay of Link

Fig. 5. Comparision of method for time delay of link

165

166

V. Dani et al. 75.5 75 74.5

Modified Mp-TCP

74

Old MP-TCP

73.5 73 Throughput

Fig. 6. Comparision of method for network throughput

For 3 Subflows (Figs. 7, 8, 9 and 10)

60 50 40 Modified MP-TCP

30

Old MP-TCP

20 10 0 Window Size

Fig. 7. Comparision of method for window size

0.002 0.0015 Modified MP-TCP

0.001

Old MP-TCP 0.0005 0 Sum of Retransmission

Fig. 8. Comparision of method for no. of retransmission

An Analysis of Multipath TCP for Improving Network Performance

167

0.0035 0.003 0.0025 0.002

Modified MP-TCP

0.0015

Old MP-TCP

0.001 0.0005 0 Delay of Link

Fig. 9. Comparision of method for time delay of link

75.5 75 74.5

Modified MPTCP

74

Old MP-TCP

73.5 73 72.5 Throughput

Fig. 10. Comparision of method for network throughput

In the above all scenario of result, we demonstrate the execution of 2 and 3 subflow of the Modified MP-TCP against compare to base MP-TCP protocol. By obtaining result maximum utilization of the network and reducing dropping packet that our result showing better performance compared to traditional protocol. We configure network to for simulating different parameter based on their functionalities. Therefore network congestion of the window where maximum traffic is going to splitting different network route.

5 Conclusion The most important update to Transmission Control Protocol in the last 20 years is multipath TCP, which allows current TCP applications to gain improved performance with reliability in current network environment. While simulations are valuable for evaluating a protocol’s large-scale characteristics, fundamental modifications like MPTCP need rigorous assessments of its performance in real-time scenarios. This work is present

168

V. Dani et al.

multipath TCP (MP-TCP) method which can improve performance of TCP in dispersal wired/wireless scenarios. The standard single path based TCP works badly in this situation caused by high packet losses and connection failures along the way. By utilizing multiple paths from sources to destinations and duplicating TCP data packets on these multiple paths, the MPTCP can significantly decrease packet loss rate from end to end; this study uses this implementation to explore performance of MPTCP in a variety of real-time circumstances, including datacenters, mobile communications, and multi-homed networks. Proposed work is implemented using simulation i.e. NS2.

References 1. Dani, V., Bhati, N., Bhati, D.: EECT: energy efficient clustering technique using node probability in Ad-Hoc network. Innovations in Bio-Inspired Computing and Applications. IBICA 2020. Advances in Intelligent Systems and Computing, vol. 1372. Springer, pp. 187–195 (2021) 2. Jacobsson, K.: Dynamic modeling of Internet congestion control (Doctoral dissertation, KTH) (2008) 3. Wischik, D., Handley, M., Braun, M.B.: The resource pooling principle. ACM SIGCOMM Comput. Commun. Rev. 38(5), 47–52 (2008) 4. Peng, Q., Walid, A., Low, S.H.: Multipath TCP algorithms: theory and design. ACM SIGMETRICS Perform. Eval. Rev. 41(1), 305–316 (2013) 5. Ford, A., Raiciu, C., Handley, M., Bonaventure, O.: TCP extensions for multipath operation with multiple addresses (2013) 6. Djahel, Ford, A., Raiciu, C., Handley, M., Bonaventure, O.: TCP Extensions for multipath operation with multiple addresses, draft-ietf-mptcp-multiaddressed-09. Internetdraft, IETF (March 2012) 7. Barré, S., Paasch, C., Bonaventure, O.: Multipath TCP: from theory to practice. In: International conference on research in networking, pp. 444–457. Springer, Berlin, Heidelberg (2011) 8. Yarnagula, H.K., Anandi, R., Tamarapalli, V.: Objective QoE assessment of dash adaptation algorithms over multipath TCP. In: 2019 11th International Conference on Communication Systems & Networks (COMSNETS), pp. 461–464. IEEE (2019) 9. Sathyanarayana, S.D., Lee, J., Lee, J., Grunwald, D., Ha, S.: Exploiting client inference in multipath TCP over multiple cellular networks. IEEE Commun. Mag. 59(4), 58–64 (2021) 10. Paasch, C., Khalili, R., Bonaventure, O.: On the benefits of applying experimental design to improve multipath TCP. In: Proceedings of the 9th ACM Conference on Emerging Networking Experiments and Technologies, pp. 393–398 (2013) 11. Szilágyi, S., Bordán, I.: Throughput performance measurement of the MPT-GRE multipath technology in emulated WAN environment. In: Proceedings of the 1st Conference on Information Technology and Data Science: CITDS, pp. 187–195 (2020) 12. Gonzalez, R., Pradilla, J., Esteve, M., Palau, C.E.: Hybrid delay-based congestion control for multipath tcp. In: 2016 18th Mediterranean Electrotechnical Conference (MELECON), pp. 1–6. IEEE (2016) 13. Li, H., Wang, Y., Sun, R., Guo, S., Wang, H.: Delay-based congestion control for multipath TCP in heterogeneous wireless networks. In: 2019 IEEE Wireless Communications and Networking Conference Workshop (WCNCW), pp. 1–6. IEEE (2019) 14. Pokhrel, S.R., Mandjes, M.: Improving multipath TCP performance over WiFi and cellular networks: an analytical approach. IEEE Trans. Mob. Comput. 18(11), 2562–2576 (2018)

An Analysis of Multipath TCP for Improving Network Performance

169

15. Chaufournier, L., Ali-Eldin, A., Sharma, P., Shenoy, P., Towsley, D.: Performance evaluation of multi-path tcp for data center and cloud workloads. In: Proceedings of the 2019 ACM/SPEC International Conference on Performance Engineering, pp. 13–24 (2019) 16. Chauvenne, R., Libioulle, T., Bonaventure, O.: MultiPath TCP connections: analysis and models (Doctoral dissertation, PhD thesis) (2017) 17. Zhang, T., Zhao, S., Cheng, B.: Multipath routing and MPTCP-based data delivery over manets. IEEE Access 8, 32652–32673 (2020) 18. Rehmani, M.H., Saleem, Y.: Network simulator NS-2. In: Encyclopedia of Information Science and Technology, 3rd edn, (pp. 6249–6258). IGI Global (2015)

Improving 3D Plankton Image Classification with C3D2 Architecture and Context Metadata Nassima Benammar1(B) , Haithem Kahil1 , Anas Titah1,2 , Facundo M. Calcagno3 , Amna Abidi1 , and Mouna Ben Mabrouk1 1

Altran Research, Capgemini Engineering, Paris, France [email protected] 2 ENSEEIHT, Toulouse, France 3 Servier Monde, Suresnes, France

Abstract. Studying the variations of the submarine environment at the plankton level can significantly contribute to the preservation of the environment. In situ plankton imaging systems have known an important evolution giving large scale plankton data for organism classification and analysis. Automated classifiers based on Convolutional Neural network are identified as highly efficient methods for image classification but require careful configuration especially for 3D images. In this paper, we propose a CNN architecture for 3D image classification to classify 155 classes of plankton from TARA Oceans dataset in four levels of hierarchical classes. We experiment and compare our proposal denoted C3D2 with competitive CNNs already performed on the case of plankton recognition such as DenseNet and SparseConvNet. Furthermore, we design several methods to incorporate context metadata on CNN architectures in order to boost the performance of the classification model. Finally, we show that C3D2 is more precise than other models. We also show the impact of incorporating context metadata into CNN architecture on different levels of classes. Keywords: Deep learning classification

1

· CNN · Plankton recognition · 3D image

Introduction

Oceans and seas houses billions of various and complex organisms called planktons that are fundamental for ocean ecosystems being the basis of the food chain. [14,22]. Anormal changes at plankton level is a concrete indicator of climatic events. Recently, advanced imaging methods such as In Situ Ichthyoplankton Imaging System (ISIIS) [6] provide a huge number of digital images for scientific studies especially for plankton classification. Several plankton image datasets are now available such as WHOI [24], ZooScan [6] and National data science Bowl [2,3]. Accurate classification is not easily achieved in the case of plankton classification because of the amount of classes and the complexity of their c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 170–182, 2022. https://doi.org/10.1007/978-3-030-96299-9_17

Improving 3D Plankton Image Classification

171

shapes. According to their taxonomic classification, in [18], authors found out manually more than 85000 classes of plankton. Manual classification is not relevant because of the non-reasonable time cost. In [22], authors present the main difficulties of this task. For example, the hand labeling by [16] took 2,880 person hour to identify around 70000 frames with few experts. Consequently, there is a real need of automated classification. Fortunately, it is possible to overcome classification issue with the emergence of deep neural networks and the insurance of their high efficiency in image classifications [7], in particular Convolutional Neural Networks (CNNs). For plankton recognition, several classification studies using CNNs [4,20] obtained high performances using plankton 2D image classification for WHOI and National data science Bowl datasets. In this paper, we use TARA dataset from the Oceanomics project1 . The European consortium Oceanomics combines a group of partners aiming to promote a better knowledge and use of the Ocean plankton. They collected samples and Eco morphogenetic data from different plankton organisms in more than 150 locations around the world. The main ambition of this project is to reach a new taxonomic, metabolic and ecosystemic comprehension of the maritime evolution. Capgemini engineering, formerly Altran, is a member of the consortium particularly involved in the image classification task. This project is an extension of a project developed by TARA Oceans Foundation2 using more recent deep learning techniques. In [1], an automated imaging system has been developed to detect types of plankton from 3D-images generated by a technique called environmental high content fluorescence microscopy (e-HCFM). A collection of 480 numeric 2D/3D features is then generated. In [15], a fusion of CNN models produced 95.8% accuracy for WHOI Plankton (22 classes). However, it is difficult to reach such performance on TARA dataset because the planktons are pictured in 3D images. Secondly, TARA dataset is composed of 104586 3D images with 155 imbalanced classes while the existing contributions on plankton recognition consider up to 33 classes [15] with sufficient amount of data per class. Fortunately, TARA dataset is also composed of metadata that can boost the classification precision. In this work, we show how we enhanced TARA 3D image classification using a new 3D CNN architecture denoted C3D2 which is an extension to [27] with data augmentation, the use of dropout and batch normalization layers. The C3D2 is trained on TARA dataset and compared to the above mentioned CNN architectures. Secondly, inspired by [4] and [20], we present three approaches in order to incorporate context metadata into the CNN architecture. The main objective of these approaches is to enhance the classification precision especially for 155 classes. This paper is organized as follows: various works about image classification especially for plankton recognition are presented in Sect. 2. Section 3 details TARA 3D images and the additional metadata of the dataset. The main 1 2

http://oceanomics.eu/fr. https://oceans.taraexpeditions.org/en/.

172

N. Benammar et al.

contributions of this paper are presented in Sect. 4. We first present the C3D2 architecture and two competitive methods that are Spatially Sparse ConvNet and DenseNet, because of their high performance in plankton recognition [15,17]. We also present the three methods that incorporate metadata into a CNN architecture. Section 5 discusses the experimentation results followed by the conclusion in Sect. 6.

2

Related Work

With the emergence of Convolutional Neural Network (CNN), such as LeNet [13], improved by AlexNet [12] and its followers ZFNET [28], GoogLeNet [26], VGGNet [23], ResNet [9], and more recently DenseNet [10], convolutional neural networks have become the most accepted technique in image classification due to their high performance in the field. In [15], authors tackled the plankton recognition issue on the three datasets used by [29]3 : National Data science Bowl dataset by [2] available on Kaggle, WHOI Plankton and Zooplankton. They trained several CNN models and obtained the precision with DenseNet (up to 94%). In the same paper, authors experimented the fine tuning and the fusion of multiple CNNs which increased the precision to 95%. In [17], 75000 images from [2] have been classified with 84% of precision using a CNN model based on VGG16 called Spatially Sparse ConvNet. Its particularity is to use Fractional Max Pooling. However these methods have been designed for 2D images. Several studies about 3D recognition in case of video, object or action recognition propose 3D CNN Models. LonchaNet [5] trains three pretrained GoogLeNet models on 3 slices of a 3D object. This solution reduced the complexity since it takes 2D projections but cannot be applied in our use case because of the high complexity of TARA images. Fortunately, the idea of looking for filters to help image recognition has not only been developed for 2D images. In, [27], the proposed model denoted C3D has the ability to input videos and output, with a great accuracy, a generic spatiotemporal feature. As 3D images can have the same matrix representation as videos, we took these model ideas as draft and added Batch Normalization [11] and Dropout [25] to control over-fitting. Furthermore, the use of Data Augmentation makes the model more robust through simple techniques, such as cropping, rotating, and flipping input images [21]. In, [25], a 3D Data Augmentation pipeline is developed for 3D human pose estimation in the wild. In order to compare our proposal to existing architectures such as DenseNet [10], we used 3D convolutional layer instead of 2D layers to suit 2D CNN models to our use case. Furthermore, several works proposed different approaches in order to boost a CNN based classifier with context metadata. In [4,20], different methods have been proposed to incorporate metadata features into CNN layers. The accuracy has been increased up to 12.2 points [4] for 27 classes. Consequently, as a second contribution, we explore the incorporation of context metadata and design different approaches in order to increase the precision. 3

https://github.com/zhenglab/PlanktonMKL/tree/master/Dataset.

Improving 3D Plankton Image Classification

3

173

TARA Oceans Dataset

A set of 104586 samples have been generated using an automated imaging system developed by [1] using a technique, called Environmental High Content Fluorescence Microscopy (e-HCFM), to picture planktons in 3D objects and generates a set of 480 numeric features. They are classified according to 4 hierarchical classes: from 2 classes to 155 classes. In this paper, we are interested in levels 2 (15 classes), 3 (33 classes) and 4 (155 classes). 3.1

Image Data Analysis

Each image is represented by a Graphics Interchange Format (GIF) composed of 20 frames that are slices on the original object. The animated image in thus a variation in the depth and not in time. Each frame is composed of six different sub-images in a tabular form of 3 × 2 as illustrated in Fig. 1. They are generated from different filters. For example, the first two columns are fluorescent signals such as Blue channel chlorophyll and cyan purple channel Alexa546.

Fig. 1. Input image Example. Representation of two Plankton images, of first hierarchy “Living”, “Bacillariophyta” as second hierarchy, with third hierarchy “Coscinodiscophyceae” and final hierarchy “Centric 14”.

All the images are from different sizes depending on the dimension of the plankton. Furthermore, Fig. 1 shows the complexity of the use case because several planktons, belonging to the same 4th level class, can have different shapes. The 3D shape of the images and the 6 sub-images raise some interrogations about the effectiveness of the filters and the frames in the representation of the plankton. We analyzed the total dataset in order to select the filters and the frames to be considered in the training. The reduction of the image size into the most significant part of it produces better performance and makes the training faster. We computed the amount of white pixels in each sub-image and each frame and depicted the statistic results with boxplots (Fig. 2). When the boxplot is small with a median greater that 90%, it means that a major part of the image is white and thus useless as for filters 1 (Fig. 2.(a)) and 3 (Fig. 2.(c)) starting from frame 13. We transformed the images into gray scale and switched into 255 (white) every light color pixel having a value greater than 240 in order to eliminate light gray spots. After the transformation, the boxplots of all filters

174

N. Benammar et al.

(Boxplots e, f, g and h in Fig. 2) show that the essential part of the image is in average in the 8 first frames.

Fig. 2. Boxplots representing the distribution of pixels across the 20 frames for four filters.

3.2

Metadata Analysis

The metadata are generated thanks to the e-HCFM technique. A set of 480 features, including bio-volumes, intensity distribution, shape descriptors and texture features are associated to each image. These features, being incorporated into CNN layers, can improve the classification. However, we need to analyze them with feature engineering tools. The correlation matrix in Fig. 3, while illegible, shows that an important number of features are correlated as we can distinguish dark blue points out of the diagonal. By selecting only one feature among a group a highly correlated features, we obtain 270 features. We also applied an ANOVA (analysis of variance) on the 270 features in order to reduce them any further. We found out that 25% of the numeric features have a F-score value more than 700 which means that these features are strongly correlated and may be sufficient to increase the precision. We confirm this assumption by the curves in Fig. 4 where we trained Random forest and XGBoost models using 32 features (with F-score greater than 900). For the 2nd level, the precision is about 94%, respectively 93%, for XGBoost, respectively Random Forest, while SVM produced 88%. For 155 classes, we obtained 56% of precision with Random Forest but more than 80% for approximately 60 classes.

4

Automated Classifications

In this section, we detail our proposal C3D2 model and present the most connected models to our case study. In addition, we present our approaches to incorporate context metadata into a CNN architecture.

Improving 3D Plankton Image Classification

175

Fig. 3. Correlation matrix of metadata.

4.1

CNN Models

CNNs forward images as three-dimensional arrays (N × M × C) where (N × M) are the size of the image and C the number of colors channels. In our case, we have one more dimension that corresponds to the number of frames (the depth of the GIF image) and color channels correspond to the sub-images. For the depth of the image, we have carefully chosen the ten first frames following our image analysis. We also left apart the sixth sub-image because it represents a concatenation of the 5 other sub-images in addition to a border detection. As depicted in Fig. 5, the network in our case has a 4D input of size (5, 10, 200, 200) and the size of the output of the network depends on the level of the hierarchy (from 2 to 155 classes). C3D2 Model. According to [8], 3D Convolutional models are more suitable for spatiotemporal feature learning compared to 2D ConvNets and a homogeneous architecture with small 3 × 3 × 3 convolution kernels in all layers is among the best-performing architectures for 3D ConvNets. Following these findings and using their C3D model (based on VGG16 [23]) as draft, we propose a new version of C3D2 to address TARA Oceans image recognition problem. Our model called C3D2 has 5 convolutional groups of one or two convolutions. We also use Leaky ReLU as activation function, Batch-Norm layers (3D and 1D) and Dropout layers to control over-fitting. Figure 6 gives the number of convolutions at each group starting with 64.

176

N. Benammar et al.

Fig. 4. Precision curves for increasing number of classes associated to the level 4 (from 10 to 155 classes) using Random forest (left) and XGBoost (right).

Fig. 5. Image data transformation into input tensor to the CNN model.

Related Work Models. To the best of our knowledge, the best models trained on 2D plankton datasets cited previously are DenseNet [10,15] and SparseConvnet [17]. For DenseNet, the average precision for 22 classes on 6600 WHOI samples was 94.9% and for 38 classes of the data science bowl dataset (on 14300 samples), the precision was 91.4%. In the case of Sparse ConvNet 42000 images with 108 classes have been grouped into 37 taxonomically and functionally groups. The average precision was 90.7%. In our case study, we have adapted the networks to 3D images using 3D Convolutional layers and batch normalization. Sparse ConvNet [17] is composed of 13 convolutional layers and 12 fractional max pool layers with a scaling factor of √12 . The network The n-th convolutional produces an output image with 32 * n color-channels. DenseNet [10] is an evolution of ResNet [9] where the layers are densely connected in a feed forward fashion which means that for L layers, the number of connections is L(L − 1)/2, while in standard CNN models, the number of connections is L − 1. The network is composed of DenseBlocks (a group of multiple 2D convolutional layers) in which the inputs have the same size. There is a transition block between DenseBlocks. It is composed of a convolutional

Improving 3D Plankton Image Classification

177

Fig. 6. C3D2 network architecture.

layer, a batch norm, ReLU and average pooling. In this paper, we have applied the DenseNet 121 with 4 DenseBlocks where 6, 12, 24 and 16 are the respective numbers of convolutional layers in the 4 DenseBlocks. 4.2

Metadata Interaction

Related work showed that the incorporation of metadata at a specific level in a neural network can improve the precision of the classifier. Since our study aims to reach a high precision in classifying TARA oceans dataset, we need to explore the incorporation of context metadata in our CNN model. The 32 extracted features can compensate the lack of sample per class for several classes and overcome the complexity of the images as well. We propose 3 methods, directly inspired from [4,20], where C3D2 is the backbone network. Metadata Neural Network - Pixel Neural Network (MNN-PNN) (Fig. 7): The input of size (5, 10, 200, 200) is forwarded through all the layers of C3D2 producing an output of size 2048. At this level, we concatenate 2048 features and 128 coming from a fully connected layer relied to the 32 features vector.

Fig. 7. MNN-PNN architecture

Metadata Neural Network - Pixel Neural Network + (MNN-PNN+) (Fig. 8): in this method, we add more interaction of the extracted features by employing the fully connected layers before the concatenation function. Metadata and Pixels Neural Network (MPNN) (Fig. 9): On the contrary of the other methods, we concatenate the extracted features (about 65k) from the convolutional group of C3D2 with the 32 vector. After that, we apply the two fully connected layers as for the C3D2 model.

178

N. Benammar et al.

Fig. 8. MNN-PNN+ architecture

Fig. 9. MPNN architecture

4.3

Network Training

The Oceanomics project provided more than 100k samples of 3D plankton images manually classified into 155 classes. Since the lineage of each plankton is mentioned, the hierarchical classes have been automatically generated in order to simplify the classification. However, the dataset is imbalanced especially for the level 4 of classes (155) where 90 classes have less than 100 samples and 4 classes have more than 20000 samples. It is difficult to achieve good accuracy for 155 classes. This is why we used a subset of 10000 samples generated by a pseudo balanced algorithm. With this subset, the precision is still challenging because most of the classes have less than 100 samples. In order to overcome this problem, we used the data augmentation with random rotates, flips, crops and brighter gray color. Indeed, the generation of subtle variations of the original images can improve the generalization of the classifier. We trained for 40 epochs with 36 of batch size and a learning rate of 0.0005 after a tuning on all the models. We have chosen SGD as a classifier and cross entropy as loss function. We use one NVIDIA GPU with 4 CUDA workers and 32 Go of memory. For the 4th level, the training took from 25 h for C3D2 to 70 h for DenseNet.

5

Experimental Analysis

We first trained the three models on the 10k dataset over 40 epochs for the 2nd and the 4th hierarchical levels of classes and compare the contribution of 3 subimages (1, 3 and 5) and 5 sub-images (the first 5 filters). We present only the precision metric giving that its behavior is the same as the accuracy in this case with an average gap of 0.02. After that, we present the results of the training of the three above presented methods in order to discuss the performance of the incorporation of metadata into CNN architecture. Since the C3D2 is faster and more precise, we experimented the three methods only on C3D2. We have

Improving 3D Plankton Image Classification

179

three training tests for each method, corresponding to each of the three levels of classification. Table 1. Classification results of 3D CNN models trained on 10K samples of TARA oceans dataset for hierarchical levels 2 and 4 using 3 and 5 sub-images. Model

Level 2 (15 classes) Level 4 (155 classes) 3 filters 5 filters 3 filters 5 filters

SparseConvNet [17] 67%

69%

-

40%

DenseNet121 [10]

71%

71%

45%

38%

C3D2

78%

81%

55%

56%

Before comparing any methods or competitive models, we trained the model C3D2 on the total dataset of 104586 samples over 155 classes and obtained 83% of accuracy and 37% of precision. The weak precision is due to the imbalanced data with a huge difference in terms of number of samples per class. The training of 10k samples, generated using a pseudo balanced algorithm, in addition to image processing presented in Fig. 2, allowed us to enhance the precision up to 56% with C3D2 (Table 1). When the samples among 155 classes are grouped in 33 or 15 classes, the precision is way better. This result confirms the findings of [17] that assert that hierarchical classification improves the precision for unbalanced data: 108 classes have been grouped into 37 classes producing a precision of 90%. The best results we obtained is 81% with C3D2 on the 2nd level of classification with 5 filters. It is relatively an important result because of the complexity of 3D plankton images in our context. One interesting result is that the competitive models don’t behave the same performance as for other 2D plankton datasets because of the significant difference between them. The contribution of the subimages is relevant here when we compare the results of 3 filters and 5 filters. One little exception is noticed for the DenseNet where the precision for 3 filters is greater. This is probably due to the fact that the network is too dense with 5 filters. Regarding the context metadata incorporated into the C3D2, we notice that we couldn’t improve the 81% of precision for the 2nd level of classification. One interesting finding is that metadata have no real impact on the higher levels of classification. The principal reason is that several classes from 3rd or 4th level can be associated to a 2nd level class. As shown in the metadata analysis, the features are very correlated to the 4th level classes which means that several images belonging to different classes from level 4 and to the same 2nd level class can have different distribution of numeric features. In this case, the network cannot be led to the right weights especially with the lack of samples in this case. Fortunately, the 155 classes are enough specific to the plankton types to be directly impacted by the incorporation of metadata. The precision has been increased by 15% using the MPNN (Table 2). We also notice that MPNN method is better than MNN-PNN and MNNPNN+ for 155 classes. This is due to the main difference: MPNN concatenate

180

N. Benammar et al.

Table 2. Classification results using the three methods of incorporation of the context metadata into C3D2. 15 classes 33 classes 155 classes MNN-PNN

74.59%

69.09%

63.48%

MNN-PNN+ 74.56%

69.8%

65.2%

MPNN

69.63%

65.48%

75.33%

the metadata with the extracted features before any fully connected layer on the contrary of MNN-PNN methods that first introduce a fully connected layers before merging the outputs. Nevertheless, the results of MNN-PNN+ are close to the MPNN. One reason can be the number of fully connected employed compared to MNN-PNN. MNN-PNN methods are inspired from [4,20] where they produced significant improvements. The difference with our case is the difference of the output size of the neural networks in addition to the difference between the datasets.

6

Conclusion

In this paper, we studied the automated classification of 3D plankton images using CNN architectures. We proposed C3D2 based on C3D by adding dropout and batch normalization. We trained C3D2 and 2 additional models DenseNet121 and Sparse ConvNet, on a pseudo balanced dataset of TARA Oceans images and obtained the best results with C3D2 models up to 81% of precision for 15 classes. However, for 155 classes, the obtained precision does not exceed 56%. This weak performance was expected because of the highly imbalanced data. Nevertheless, the pseudo balanced algorithm improved the precision which was only 37%. Moreover, we explored the training of a CNN architecture with the interaction of numeric features, from context metadata, with extracted features from the pixels. We proposed three methods that incorporate the numeric features in different ways: MNN-PNN and MNN-PNN+ extract features from pixels and metadata and employ fully connected layers before the fusion while MPNN approach merges the extracted features before fully connected layers. The precision for the 2nd level has not been improved the precision obtained with C3D2 (81%). This is due to the important difference between numeric features for samples belonging to the same 2nd level class. Nevertheless, the challenging level 4 has been significantly improved from 56% to 65% because of the highly correlated metadata to the 155 classes. We aim to achieve better performance in terms of precision with the fusion of multiple CNN models as studied in [15] and [19]. Furthermore, we also foresee the experimentation of fine tuning approaches as proposed in [15] using large scale external datasets such as WHOI plankton [24]. Finetuning consists on pretraining a model with a large scale domain dataset in order to get familiar with

Improving 3D Plankton Image Classification

181

submarine images before a second round training with the target imbalanced dataset which can efficiently improve the precision for TARA dataset.

References 1. Colin, S., et al.: Quantitative 3D-imaging for cell biology and ecology of environmental microbial eukaryotes. Elife 6, e26066 (2017) 2. Cowen, R.K., Sponaugle, S., Robinson, K.L., Luo, J.: PlanktonSet 1.0: Plankton imagery data collected from F.G. Walton smith in straits of Florida from 3 June 2014 to 6 June 2014 and used in the 2015 national data science bowl (NODC Accession 0127422) (2015) 3. Cowen, R.K., Guigand, C.M.: In situ ichthyoplankton imaging system (ISIIS): system design and preliminary results. Limnol. Oceanogr. 6, 126–132 (2008) 4. Ellen, J.S., Graff, C.A., Ohman, M.D.: Improving plankton image classification using context metadata. Limnol. Oceanogr. 17, 439–461 (2019) 5. Gomez-Donoso, F., Garcia-Garcia, A., Garcia-Rodriguez, J., Orts-Escolano, S., Cazorla, M.: LonchaNet: a sliced-based CNN architecture for real-time 3D object recognition. In: 2017 International Joint Conference on Neural Networks (IJCNN). IEEE (2017) 6. Gorsky, G., et al.: Digital zooplankton image analysis using the ZooScan integrated system. J. Plankton Res. 32(3), 285–303 (2010) 7. Guo, Y., Liu, Y., Oerlemans, A., Lao, S., Wu, S., Lew, M.S.: Deep learning for visual understanding: a review. Neurocomputing 187, 27–48 (2016) 8. Hara, K., Kataoka, H., Satoh, Y.: Can spatiotemporal 3D CNNS retrace the history of 2D CNNS and ImageNet? In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018) 9. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2016) 10. Huang, G., Liu, Z., Van Der Maaten, L., Weinberger, K.Q.: Densely connected convolutional networks. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2017) 11. Ioffe, S., Szegedy, C.: Batch normalization: accelerating deep network training by reducing internal covariate shift. In: International Conference on Machine Learning, PMLR (2015) 12. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional neural networks. Adv. Neural Inf. Process. Syst. 25, 1097–1105 (2012) 13. LeCun, Y., Bottou, L., Bengio, Y., Haffner, P.: Gradient-based learning applied to document recognition. In: Proceedings of the IEEE (1998) 14. Lumini, A., Nanni, L.: Ocean ecosystems Plankton classification. In: Hassaballah, M., Hosny, K. (eds.) Recent Advances in Computer Vision. Studies in Computational Intelligence, vol. 804, pp. 261-280. Springer, Cham (2019). https://doi.org/ 10.1007/978-3-030-03000-1 11 15. Lumini, A., Nanni, L., Maguolo, G.: Deep learning for plankton and coral classification. Appl. Comput. Inf. (2020) 16. Luo, J.Y., et al.: Environmental drivers of the fine-scale distribution of a gelatinous zooplankton community across a mesoscale front. Mar. Ecol. Prog. Ser. 510, 129– 149 (2014)

182

N. Benammar et al.

17. Luo, J.Y., et al.: Automated plankton image analysis using convolutional neural networks. Limnol. Oceanogr. 16(12), 814–827 (2018) 18. McClatchie, S., et al.: Resolution of fine biological structure including small narcomedusae across a front in the southern California bight. J. Geophys. Res. Oceans (2012) 19. Nanni, L., Costa, Y.M., Aguiar, R.L., Mangolin, R.B., Brahnam, S., Silla, C.N.: Ensemble of convolutional neural networks to improve animal audio classification. EURASIP J. Audio Speech Music Process. 2020, 8 (2020). https://doi.org/10. 1186/s13636-020-00175-3 20. Nunnari, F., Bhuvaneshwara, C., Ezema, A.O., Sonntag, D.: A study on the fusion of pixels and patient metadata in CNN-based classification of skin lesion images. In: Holzinger, A., Kieseberg, P., Tjoa, A., Weippl, E. (eds.) CD-MAKE 2020. LNCS, vol. 12279, pp. 191–208. Springer, Cham (2020). https://doi.org/10.1007/978-3030-57321-8 11 21. Perez, L., Wang, J.: The effectiveness of data augmentation in image classification using deep learning. arXiv preprint arXiv:1712.04621 (2017) 22. Robinson, K.L., Luo, J.Y., Sponaugle, S., Guigand, C., Cowen, R.K.: A tale of two crowds: public engagement in plankton classification. Front. Mar. Sci 4, 82 (2017) 23. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 24. Sosik, H.M., Olson, R.J.: Automated taxonomic classification of phytoplankton sampled with imaging-in-flow cytometry. Limnol. Oceanogr. Methods 6, 204–216 (2007) 25. Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15, 1929–1958 (2014) 26. Szegedy, C., et al.: Going deeper with convolutions. In: IEEE Conference on Computer Vision and Pattern Recognition (2015) 27. Tran, D., Bourdev, L., Fergus, R., Torresani, L., Paluri, M.: Learning spatiotemporal features with 3D convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision (2015) 28. Zeiler, M.D., Fergus, R.: Visualizing and understanding convolutional networks. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014. LNCS, vol. 8689, pp. 818–833. Springer, Cham (2014). https://doi.org/10.1007/978-3-31910590-1 53 29. Zheng, H., Wang, R., Yu, Z., Wang, N., Gu, Z., Zheng, B.: Automatic plankton image classification combining multiple view features via multiple kernel learning. BMC Bioinf. 18, 570 (2017). https://doi.org/10.1186/s12859-017-1954-8

Livestock Application: Naïve Bayes for Diseases Forecast in a Bovine Production Application Use of Low Code Aline Neto, Susana Nicola(B) , Joaquim Moreira, and Bruno Fonte R. Dr. António Bernardino de Almeida 431, 4200-072 Porto, Portugal [email protected], {sca,jaq}@isep.ipp.pt, [email protected]

Abstract. The Brazilian agriculture expansion, the increase in exportation and the search to guarantee food safety is increasingly the investment urgency on technologies focused on this activity. Data Mining and Machine Learning could be a vehicle to facilitate the fight against the spread of infective severity. In this paper, we applied the Naïve Bayes algorithm in a web mobile application (App) developed in the Outsystems platform using low code, to predict diseases in beef cattle production. It can store information to examine the health and ensure well-being of livestock. The project was conducted using Cross Industry Standard Process for Data mining (CRIPS-DM) methodology, composed by the phases: business understanding, data understanding, data preparation, modeling, and evaluation. The paper presents a case study in an extensive beef cattle farm in Brazil, that nowadays, is facing a common problem in its own agriculture, where there are increasing cases of bovine mortality due to several diseases. As a result, qualitative research to assessing the App usability was done with seven producers. But two of them didn’t answer because didn’t have computer access. They said the application is relevant, versatile, easy, pleasant to use and to understand, also stating that use of artificial intelligence is interesting. Keywords: Livestock · Data mining · Naive Bayes · Food security

1 Introduction Brazilian agriculture is expanding, and it is the essential economic activity for the growth of the country. In 2017, it occupied a territory of 351,289,816 hectares with 4,996,287 markets in this area. Among which, 50.49% worked in the cattle herd raising a total of 172,719,164 animals. Large-sized animals being responsible for 70% of this production (IBGE 2019a). In the third quarter of 2019, around 394,000 tons of beef were exported, with a turnover of US$1,627,547 million. This economic activity has a high loss rate, as in the third quarter of 2019, 8,493,975 animals had been slaughtered due to sanitary issues in government inspections, which, as a result, it had a significant contribution to the high bovine mortality rate in the sector (IBGE 2019b). In rural Brazil, the São José farm is an example of the Brazilian agriculture context. They have extensive beef cattle production in a natural pasture with 55 hectares of © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 183–192, 2022. https://doi.org/10.1007/978-3-030-96299-9_18

184

A. Neto et al.

extension. The annual mortality rate is 10% effective and equivalent to 3,500 euros of damage, about 67% are from unknown causes. They do not use information systems (IS) as the region faces limitations due to the precariousness of the telecommunications infrastructure. Producers are unaware of the advantages of adopting a technological system and are not able to predict, identify and control the early spread of infectious disease in animals. Such challenges formulated the following research question: How to reduce the mortality rate, expenses, financial losses and identify early clinical cases of contagious diseases in a beef cattle? Gottschall et al. (2010) observed that the Brazilian herd is marking by low longevity due to high mortality. It is possible to implement preventives measures to reduce the rate and make the production system more profitable. Gates (2021) annunciated that much of the world’s population suffers from poverty and depends on agriculture to survive. In addition, he said when short farms are not productive, the economy is impaired generate more subhuman life, more starvation, and pain. Therefore, making agriculture more invulnerable and resistant adaptive to innovation becomes more and more necessary and urgent. Currently, livestock is being widely criticized for the lack of sustainability of its productive activity, increasing the pressure for technological innovation adoption to act on this deficiency (Linden et al. 2020). Some authors consider Data Mining as synonymous to search for knowledge of a data set, that involve an application of algorithms and techniques on data to explore valuable and implicit knowledge information (Goldscmidt and Passos 2005). According to Ramos Filho (2012) Data Mining uses techniques to process data and recognize meaningful patterns to provide knowledge, inducing intuitions in specialists and help them make current decisions. The aim of this paper is to contribute to a sustainable development of a new agriculture era. Using the Data Mining innovation to extract information from a long dataset and, using a Machine Learning algorithm that consists of forecast models, it will be possible to reduce cattle mortality ensuring the profitability of farming activity. Using Low-Code technology, and to predict diseases and further infections, an application, was developed. Making disease forecasts precedes the veterinarian arrival, enabling the animal isolation quickly minimizing the potential for contagion among the other animals in the herd. It will assist the producers in reducing bovine mortality and increasing the guarantee of food safety of the meat offered to the final consumer.

2 CRIPS-DM Methods for Data Mining Schröer et al. (2020) use the CRIPS-DM method for Data Mining projects according to six phases. The phase of Business Understanding, the expert analyzes the situation, requirements, business objectives are evaluated, also goals and project planning are determined. Data Understanding, is the phase where the data are collected, analyzed, and described. It is possible to identify errors to correct in the next phase by building a more qualified dataset. Phase 3, Data Preparation, it is related with the selection of attributes, as well as, the treatment and the construction, cleaning, and the integration of the dataset. In the modeling phase, the modeling techniques are selected, the parameters are specified, and the test design is defined. In addition, the data is processed by the algorithm, as well as the explanation and the description of results. Evaluation: at this stage, the results

Livestock Application: Naïve Bayes for Diseases

185

are interpreted and compared to the initial business objectives. Afterward, a review of the entire project is carrying out, and new actions are defined to obtain improvements. The last phase, Deployment, is when the reports and conclusions are writing with future strategies. Which is includes project monitoring and maintenance plan. 2.1 Machine Learning Based on Wedel and Marx (2022) Machine Learning has the ability to obtain value information to the expert perform relevant interpretations. Which is sectioned in reinforcement, unsupervised and supervised learning. The supervised, which was used in this project, present labels and the result classification is generated during the learning of the relation between an unknown input and output data. The machine learning algorithm Naïve Bayes was used as a method to extracted knowledge during the data mining project conducted by the CRIPS-DM methodology. The algorithm carries out the forecast and classifications of sick animals during the inspection activity by the production staff prior to the veterinarian arrival. Nevertheless, because of the small historical dataset available by the farms was not possible analyze the accuracy of the algorithm’s forecasts using the evaluated metrics, making a topic for future works. 2.1.1 Naïve Bayes Naive Bayes is a probabilistic supervised learning algorithm, which has labels and is based on the Bayes Theorem. According to the Eq. (1), A and B are two events or attributes, and P(A|B) is the probability of A given B. It does not need a huge amount of data; it is effective for the current reality and complex cases. Because the uncertainties do not affect the calculation of probability either the final decision result (Ferreira 2012). P(A|B) =

P(B|A)P(A) P(B)

(1)

The algorithm requires little storage space, moreover, it has fast processing, low computational complexity, is ideal for binary values, is naturally robust to missing values and noise (Osisanwo et al. 2017).

3 Development: Implementation of CRISP-DM 3.1 Business Understanding In the case study, São José farm’s, main economic activity is the extensive unconfined beef cattle raising. This farm does not work with animal reproduction cattle, they are purchased young and undergo a fattening process until reaching the ideal weight for sale. Meanwhile, the production is facing difficulties, a very significant mortality rate of 10% per year. This rate is responsible for an annual loss of approximately 3500 euros. In addition, they cannot precisely identify the death causes, about 67% of the cases are unknown. They do not register or analyze data neither use computer systems. As a result, they aim to achieve the strategic objectives of 40% mortality rate reduction

186

A. Neto et al.

and rise the identification of 70% of the death causes. Then, they can reduce 15% of medical expenses and decrease the financial loss of deaths. In addition, to achieving tactical goals facilitating the planning of preventive actions for animal integrity and early diagnosis of diseases and infestation risks. Improving the reliability and quality of meat supplied also transforms the production more sustainable and profitable. The Naive Bayes algorithm is applied to carry out sick animals’ classifications as an automatic screening system anticipated. Its goal is to alert the producer in case of infection risks and help to animal’s isolation and treatment on time. Then avoid the aggravation of clinical cases or epidemiological spread among the herd. The algorithm optimizes the effectiveness of veterinary consultations and reducing medical expenses, facilitating the fight against mortality. 3.2 Data Understanding This section contains the description and analysis of the data collected available by the farm. The data was available through Excel sheets, physical files, as well as, through interviews and meetings. In this phase it will be possible to identify errors presented in the dataset and the corrections needed to improve the quality. The data set contains 118 animals that lived on the farm between the last quarter of 2019 until the first half of 2021. According to the dataset available, 56% of the cattle are female, 21% male, and 23% are not specified. The bovine population is composed of four species: 33% of cows, adult, and breeding females over two years old; 13% of heifers, young adult females between 12 and 24 months of age that did not reproduce; 13% of steers, young adult males between 12 and 24 months of age that did not procreate and 41% are calves up to 6 months of age. The Fig. 1 shows percentage charts of herd gender and species.

Fig. 1. Characteristic of the herd

During the study, the veterinarian recorded three diagnoses of three calves with bovine parasitic sadness, a contagious disease among the herd due to contact with parasites. Then the attributes were defined as characteristic symptoms presented by the animal, that were processed by the Naive Bayes algorithm. The attributes were: Is the animal prostrate? Is it having diarrhea? Is it having pale ocular mucosa? Is it presents a water belly? The farmer believed that identification and isolation, in these cases, were too late, with transmission occurring between the two animals. As a result, 800 historical

Livestock Application: Naïve Bayes for Diseases

187

records were provided by the veterinarian, corresponding to 20 days of inspections on 40 animals. The inspections forms had records of 3 animals diagnosed and 37 heaths. The Table 1 shows the inspections collected data. Table 1. Data collected Prostrate animal? Yes

16

Diarrhea? 8

Pele eye mucosa? 14

Belly of water? 7

Sick animal? 20

No

784

792

786

793

780

Total

800

800

800

800

800

3.2.1 Problems Identified in the Dataset After collecting the information, some issues have been observed. Due to the nonadoption of information systems, they stored a small amount of historical data information, with errors in data standardization, and with no historical data recorded. The problems were alleviated during the data preparation phase, making the reliability and quality of information more satisfactory and accurate. 3.2.2 Solutions for the Dataset Problems To improve the quality of the dataset, the necessity to type all information available on paper by the farm arose. Then, there was the need to standardize some data such as the animals ID, which presented repetition and lack of logic and sequencing in the original dataset. Also, it was necessary to select relevant data, clean and correct wrong and insignificant information, and fill in some missing values with the farmers help. To prevent the same mistakes derived from using the App by the users, it was extremely necessary to apply some data validation methods in the tool’s forms, such as, the application of mandatory type and alert messages to avoid storing empty and wrong type of data. This enabled the user to save forms with missing information and wrong values. 3.3 Data Preparation During the data preparation phase, was created a dataset, the access link menus, measurements errors prevention such as user empowerment and data validation. Then the data were prepared and corrected, formatted, inserted, and integrated. The Fig. 2 shows the entities diagram and their relationships. After that, some data validation measures were standardized the values, minimizing errors during data recording. For example, the mandatory filling alert messages to avoid storing empty messages. The Fig. 3 illustrated the inspection form and the algorithm calculation table.

188

A. Neto et al.

Fig. 2. Diagram of the entities on the Outsystems platform and data integration relationships

Fig. 3. Inspection form built in the application for storing record data over the production days. The calculation algorithm table

3.4 Modeling During the modeling phase, the algorithm of machine learning Naïve Bayes was chosen to be the Data Mining technique. After the choice decision, the technique had been implemented on the calculation table inspections form to make disease forecast, on the Outsystems development platform. 3.4.1 Modeling Technique Choices The modeling technique choices needed to meet business and dataset requirements. The first requirement evaluated was that it should extract knowledge from inspection table dataset and generate disease predictions. The machine learning algorithm would be ideal. As the data in the table were binary and had labels the algorithm had to be supervised, with good performance to process binary data. Also, it must be applicable to Outsystems platform. The Table 2 below was used as a theoretical foundation to compare and choose the ideal algorithm. The authors Osisanwo et al. (2017) compared 13 performance factors of the follow six algorithm, Decision Trees and Neural Networks, and, Naïve Bayes, and K-Nearest Neighbors, and Support Vector Machine, and RuleLearner’s. The Naïve Bayes stands out first because it is based on the Bayes theorem

Livestock Application: Naïve Bayes for Diseases

189

composed only of equations and is easy to be implemented in Outsystems. Afterwards, it can generate sickness forecast among the inspections data set because it performs well with binary data as mentioned above, it is a requirement of the dataset (Fig. 4).

Fig. 4. Comparing learning algorithms (**** stars represent the best and * star the worst performance) Source: (Osisanwo et al. 2017)

According to Osisanwo et al. (2017) The algorithm Naïve Bayes fit all the requirements needed when compared to others: - it is a probabilistic model which minimizes the impact of missing values and noise, and redundant attributes, being a differential given all the reported data collected problems described in subtopic 3.2.1 above; - presents low complexity in handling parameters and fast processing and learning; - speed to classify, and require little storage space; - low complexity in knowledge extraction and explanation of classification results that it would help increase application performance and the speed of the forecast response to the user. 3.4.2 Algorithm Construction The algorithm construction starts with the frequency count of each class of the four attributes had performed. Subsequently, the probabilities of the events had been calculated. And the Naive Bayes formula presented above had applied for the positive rating and the negative. Then the application performs the comparison between both results.

190

A. Neto et al.

Fig. 5. Algorithm construction and Outsystems implementation diagrams

The one with the highest value corresponds to the forecast response. After that, the platform feeds back the database, storing the new classification as historical data for future predictions. The steps of the algorithm construction and the Outsystems implementation diagram have detailed in the following Fig. 5. 3.5 Evaluation Four farmers and veterinarians of the case study farm answered an evaluation form with questions about the application’s usability. The evaluated questionary was sent to six productors of the region but 2 of then didn’t have access to computer and could not test the App and answer the questionary. As a result of the answers, a positive evaluation had been obtained. They said that the application is easy and to use and to understanding, and nice to browse, relevant to the livestock industry, versatile because could use on mobile phone also the use of Artificial Intelligence interesting. The questions and their answer had summarized in Table 2 below.

Livestock Application: Naïve Bayes for Diseases

191

Table 2. User feedback on the application useability Unsatisfactory

Satisfactory

Fine

Good

How easy is to record/edit, understand, and query data?

1

2

How relevant are the features for your productive activity?

1

2

Excellent

How easy is it to use and interpret the results?

2

1

How pleasant is the navigation and visual part of the App?

1

2

How important is it to use management systems aimed at livestock?

2

1

What do you think about the use of Intelligence Artificial?

1

How relevant is the versatility of using in mobile devices and computers? Total

3

2 2

1

13

5

4 Conclusion Farmers are increasingly looking for solutions to deal with daily problems that generate high losses, such as the high mortality rate of herds. Aiming to provide a solution, it was developed an application considering a technological innovation. It is an integrated service that offers in the same solution a database available for export, with data transforming functionalities that provide relevant information to users, such as the innovation of the artificial intelligence algorithm resource to predict diseases. It was able to answer the research question, most of the financial losses were related to medical expenses and with the high mortality rate. The goal of the application was to manage and to develop a method to effectively assist farmers in reducing these costs. During the evaluated phase the farmers reported that adoption of management systems at livestock production, as well as the use of artificial intelligence were important. In addition, the tool was evaluated as relevant for the cattle productive activity, it is easy to use and interpret, and pleasant to navigation. The added value of the application is the versatility of web and mobile access. As future work, historical information from other farms with the same economic activity can be collected and used to finalize the algorithm validation. Another solution is to include in the future works a preliminary phase to collect and record data, then introduce a phase to make training and tests.

192

A. Neto et al.

References Ferreira, M. Da C. Da S.: Classificação hirerárquica da atividade económica das empresas a partir de texto da web [dissertação de mestrado]. Repositório aberto da universidade do porto (2012) Gates, B.: A warmer world will hurt this group more than any other. https://www.gatesnotes.com/ energy/helping-the-worlds-poorest-adapt-to-climate-change?wt.mc_id=20210318100000_hta acd_bg-li_&wt.tsrc=bgli (2021) Goldscmidt, R., Passos, E.: Data mining: um guia prático. Editora campus. https://doi.org/10. 21529/RESI.2006.0501011 (2005) Ibge: Censo agropecuário 2017: resultados definitivos. Censo agropecuário 8, 1–105 (2019a). https://biblioteca.ibge.gov.br/visualizacao/periodicos/3096/agro_2017_resultados_d efinitivos.pdf Ibge: Indicadores ibge: estatística da produção pecuária. (2019b). https://biblioteca.ibge.gov.br/ visualizacao/periodicos/2380/epp_2019_3tri.pdf van der Linden, A., de Olde, E.M., Mostert, P.F., de Boer, I.J.M.: A review of European models to assess the sustainability performance of livestock production systems. Agric. Syst. 182(July 2019), 102842 (2020). https://doi.org/10.1016/j.agsy.2020.102842 Osisanwo, F.Y., Akinsola, J.E.T., Awodele, O., Hinmikaiye, J.O., Olakanmi, O., Akinjobi, J.: Supervised machine learning algorithms: classification and comparison. Int. J. Comput. Trends Technol. 48(3), 128–138 (2017). https://doi.org/10.14445/22312803/IJCTT-V48P126 Ramos, R.C.: Harvard - um sistema para extracção automática de conhecimento [tese de doutoramento, repositório aberto da universidade do porto]. https://repositorio-aberto.up.pt/handle/ 10216/72664 (2012) Gottschall, C.S., Canellas, L.C., de Almeida, M.R., Magero, J., Bittencourt, H.R.: Principais causas de mortalidade na recria e terminação de bovinos de corte. Rev. Acad.: Ciência Anim. 8(3), 327 (2010). https://doi.org/10.7213/cienciaanimal.v8i3.10916 Schröer, C., Kruse, F., Marx, J., Kruse, F., Marx, J.: Sciencedirect sciencedirect a systematic literature review a systematic literature review on applying process model on applying crisp-dm process model. (2019), 526–534 (2020). https://doi.org/10.1016/j.procs.2021.01.199 Wedel, F., Marx, S.: Application of machine learning methods on real bridge monitoring data. Eng. Struct. 250, 113365 (2022). https://doi.org/10.1016/j.engstruct.2021.113365

Development of a Workstation Assessment APP, Integrating Performance with the Worker’s Health and Well-Being Pedro C. Ribeiro1 , Marlene F. Brito1 , and Ana L. Ramos2(B) 1 Institute of Engineering, Polytechnic Institute of Porto - ISEP/IPP, Porto, Portugal

{1161503,mab}@isep.ipp.pt

2 GOVCOPP Research Unit/Department of Economics, Management, Industrial Engineering

and Tourism (DEGEIT), University of Aveiro, Aveiro, Portugal [email protected]

Abstract. The purpose of this paper is the development of a mobile application inspired by a workstation assessment tool that integrates performance with the worker’s health and well-being. This development consists of an improvement of the tool, adapting it to a more current context, promoting its convenience and effectiveness. The tool is called ErgoSafeCI and consists of ten items that are considered essential to have a lean, safe, and ergonomic workplace. This assessment instrument supports the implementation of continuous improvement projects, specifically the implementation of the Lean philosophy, making it possible to monitor throughout the entire process and helping to decide on the focus of improvement in that state. Is outlined the emergence and applicability of lean philosophy, as well as methods and tools based on it. Keywords: Lean manufacturing · Lean-based audits · Lean assessment tools · Ergonomics · Safety · Mobile application

1 Introduction Nowadays, companies are under enormous pressure to be competitive in the markets where they operate. The conditions in these markets challenge companies to strengthen and maintain their capabilities to stand out and be competitive [1]. To confront these challenges, many companies update their traditional management style and adopt methods that lead to improvements in cost, quality, productivity, and operational performance, such as lean manufacturing [2]. Lean is a management philosophy that emerged after World War II, stemming from Toyota Production System created by Toyota. In the 1950s, Eiji Toyoda and Taiichi Ohno were able to create the system called TPS, based on the Jidoka and JIT philosophies with the Ford assembly line [3, 4]. According to Womack and Jones, lean production is “lean” because it uses fewer resources compared to mass production. The authors specify that all the human effort, production space, investment in tools and hours spent to develop a new product is only © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 193–204, 2022. https://doi.org/10.1007/978-3-030-96299-9_19

194

P. C. Ribeiro et al.

needed in half [3]. In 1996, the same authors defined five lean principles to improve the functioning of a company, eliminating waste: specify value; identify the value stream; flow; pull; and pursue perfection. The objective of these principles is to do more with less. Furthermore, they believe lean ideas are the single most powerful available tool for creating value and reducing waste in any organization [5]. Some companies looking to increase their productivity levels focus on the process of continuous improvement using the lean philosophy. However, is needed a methodology aimed at the correct application of these concepts without neglecting the human factor [6]. An ergonomic intervention is able to identify ergonomic risk factors during work and can be used as a tool to reduce waste associated with unnecessary movement. Movements that generate waste in ergonomics, such as stretching, bending and inappropriate postures, can hurt the safety and health of workers as well as reduce productivity and efficiency [7]. Ergonomics has the potential to decrease lead time and increase yield, removing the waste of non-productive movements and activities [8]. In conclusion, a lean transformation considering ergonomic principles can increase productivity, reduce workplace accidents, and improve the design and layout of the workstation [9]. Lean implementation should be seen as a long-term process, without an end state. A company that implements the lean philosophy must be able to identify the current level of leanness, being continuously monitored, to determine the future improvement process. In addition to knowing “where to start” and “how to proceed”, the company must know the tools available at its range [10]. There is a huge set of lean tools, however, choosing the right tools requires a lot of knowledge and experience in lean implementation [11]. Various organizations focused on the implementation of only hard lean tools and techniques and neglected soft lean practices (human-related practices) [12]. In the current paper, the assessment tool ErgoSafeCI that aims to improve performance, ergonomics, and safety conditions in an integrated way is transformed into a mobile application. The difficulty for companies is to identify which are the most critical areas in the current state and prioritize them before improvement interventions. With this tool, it is possible to carry out audits integrating performance with the worker’s health and well-being. The main objective of transforming the tool into a mobile application, apart from making it user friendly, is to make the workstation audit process more efficient, convenient, and easier. In addition to the implemented improvements related to the user experience, an audit history will be implemented to enable the user to consult all data referring to audits already carried out, without the need to save extra files. The saved audits will remain in the account of the user who performed them, and they can be accessed quickly. This feature is very useful, as this instrument aims to be a long-term self-assessment model.

Development of a Workstation Assessment APP

195

2 Literature Review 2.1 Assessment Audits Based on Lean Tools The expansion of lean implementations in companies provided the development of several mechanisms and methodologies capable of assessing to understand the effectiveness of the lean implementation [13]. However, most of the existing lean tools focus on “how to become leaner” rather than knowing “how lean it is”. According to Wan and Chen, the three main categories that concern the level of leanness are lean assessment tools, lean metrics, and value stream mapping [14]. In 1996, Karlsson and Ahlstrom created a lean assessment tool consisting of nine indicators that must be assessed: multifunctional teams, elimination of waste, decentralization, continuous improvement, vertical information systems, integration of functions and pull of materials [15]. In 2001, Sanchez and Perez developed a checklist of thirty-six lean indicators divided into six groups to assess changes related to lean [16]. In 2002, Goodson created one of the best known and most useful plant assessment tools, which aims to assess whether a plant is truly lean in just 30 min, called “Rapid Plant Assessment” [17]. In 2006, Srinivasaraghavan and Allada proposed a model that assesses the difference between the current state of the system and the benchmarking performance, so the result provides a measure of the level of leanness but is highly dependent on the quality of the benchmark [18]. In 2008, Wan and Chen proposed a methodology to quantify the level of leanness of productive systems based on an ideal leanness reference obtained from historical data [14]. Later, they developed an adaptive assessment of lean that shows a way to guide the entire lean implementation process effectively [11]. In 2011, Saurin et al. noted that existing models were primarily designed to assess the level of lean implementation across the plant, not specific units of the production system. Therefore, they introduced a framework to assess the use of lean production practices in these specific units, such as workstations or assembly lines [19]. 2.2 Ergonomics, Health and Safety “Ergonomics is no longer just a buzzword; it is going to be around for a long time because it makes good business sense” [20]. Fernandez also argues that the goal of ergonomics is “to adapt the task to the individual and not the individual to the task”, for this reason, companies should make ergonomic changes before the occurrence of major accidents at work instead of making these changes after the occurrence of these accidents [20]. The inclusion of ergonomics in the continuous improvement process is very important since the traditional interventions of lean systems, when trying to minimize resources and increase productivity, tend to neglect the limitations, and needs of the human factor. However, companies do not realize that if ergonomic principles are integrated and implemented simultaneously with lean systems there is potential to further improve productivity gains [21].

196

P. C. Ribeiro et al.

Tortorella et al. claim that the lean philosophy presents the human element as a fundamental factor for the sustainability of continuous improvement. On the basis of a lean ideology, ergonomics removes barriers to quality, improving productivity and human performance [22]. Yazdani et al. believe that organizations should present ergonomics and the prevention of musculoskeletal disorders as significant components of the management practices used in the company [23]. According to Wilson, the lean implementation team must consider ergonomics and safety as core values of the lean process, along with reducing waste and creating value [24]. Integration of Human Factors in Lean Assessment Audits In 2014, Wong et al. developed a lean index to assess an organization level of leanness in sustaining lean transformation considering the relationship between technology, systems, and humans [25]. In 2016, Jarebrant et al. built a tool that aims to improve ergonomic conditions and performance at the same time, this application is called Ergonomic Value Stream Mapping (ErgoVSM). The implementation of this tool is an effort to recognize the importance of assessing health risks in each job [26]. In 2017, Gonçalves and Salonitis stated that the evaluation of workstation design should focus on lean and ergonomic aspects. Lean assessment reduces workstation waste, and an ergonomic assessment protects worker safety and comfort. This relationship is essential for long-term success. Therefore, these authors developed an assessment model and a tool to assess each requirement that they consider fundamental to the workstation design [27]. All tools mentioned, as well as safety and ergonomics checklists/assessment tools, were analyzed in detail and served as input in the construction of the ErgoSafeCI tool. The difference between these tools and the ErgoSafeCI is in the evaluation of workstations through the combination of key dimensions as continuous improvement, productivity, health and safety, quality, visual management, materials flow, ergonomics, and work organization [1].

3 Modelling To clarify and facilitate the development of the mobile application, some models were created that explains the software’s operating and behaviour characteristics. This modelling became very useful to avoid programming errors and increase productivity because when the development phase started, the solution format was already known. First, a UML use case diagram was designed, represented in Fig. 1, to demonstrate the functionality of the system.

Development of a Workstation Assessment APP

197

Fig. 1. UML use case diagram

The basic features are: register a user in the application; log in as a registered user who, in turn, will be verified; perform audits; and consult audits. Next, a UML class diagram was designed (Fig. 2) that represents the structure and relation of the classes that should be created. An audit will consist of the results of each section, the average of those results, the objective, the information about the workstation and the domains. To finish this phase, a mock-up was developed to design the application’s interfaces and assess their usability and test their design.

198

P. C. Ribeiro et al.

Fig. 2. UML class diagram

4 Computational Implementation The latest version of the ErgoSafeCI contains eighty-eight questions divided into ten items: performance/efficiency indicators, continuous improvement, health and safety at work, standards and visual management, process and operations, material flow, zero defects, physical ergonomics, organizational and cognitive ergonomics, and discipline.

Development of a Workstation Assessment APP

199

Once the evaluation phases of the tool, planning the improvements to be implemented and modelling were completed, the development of the mobile application began using the Android Studio software and the Kotlin programming language. Initially, the layouts for the pages were created, identical to those idealized in the mock-up, and then the code necessary for the ErgoSafeCI tool to work. Figure 3 represent the first interfaces that the user interacts with when opening the mobile application.

Fig. 3. Login, register and menu page layouts

The user can log in or register a new account if he doesn’t have one or even if he wants to use another account to perform audits since the history of audits is associated with each account. Only in this way is it possible to advance to the next menu, which allows the user to start a new audit, consult audits already carried out with the same account or log out. When starting a new audit, the user must indicate some information about the workstation and then start the checklist. To answer, it is only necessary to select the desired answer, and in Sect. 10 that corresponds to the discipline, one of the allowed values must be indicated (Fig. 4). The questions should be answered in these forms: yes, no, and not applicable (NA).

200

P. C. Ribeiro et al.

Fig. 4. Checklist layout

Each page of the checklist corresponds to a previously defined section, and they work in the same way, so the code developed is similar to all of them. At any time during the completion of the checklist, the user is required to respond to all fields to proceed to the next phase. A temporary warning message will appear on the screen if a parameter is not filled in or a question is answered. When finished, the scores for each section, the average score, the objective, and the priority domains to work are presented. Scores can be analyzed in text accompanied by a positive, neutral, or negative sign or else in a radial graph (Fig. 5). The domains are sorted and the first ten are presented, considered as priorities out of a total of thirty-nine, as they improve the most critical sections of the assessment. In addition, a glossary is made available that helps in understanding the domains to work on, if necessary (Fig. 6). Finally, the biggest advantage of using the mobile application compared to the Microsoft Excel version of the tool, the audit history, as shown in Figs. 7 and 8.

Development of a Workstation Assessment APP

Fig. 5. Results layout

Fig. 6. Glossary layout

201

202

P. C. Ribeiro et al.

Fig. 7. Audit list layout

Fig. 8. Audit history layout

Development of a Workstation Assessment APP

203

5 Conclusion For lean manufacturing to become a successful production methodology, a human factororiented approach is essential. For this, it is necessary to use decision support tools, which represent a great contribution to the design of lean manufacturing systems [28]. Several investigators have tried to develop methods and procedures to quantify how lean a process is, called the level of leanness [14]. Therefore, these methods and procedures often correspond to audits, that in addition to being used as a form of metering, should assist the lean implementation process at all stages, providing an excellent way to determine whether previous suggestions have been put into practice and whether improvements have been achieved [29]. In short, the main objective was to develop a mobile application based on the assessment tool ErgoSafeCI. Moreover, the proposed framework enables an easy and userfriendly way to realize audits, and stores all the audits carried out so that, at a more advanced stage, it is possible to make comparisons and determine their effectiveness. In terms of future work, in addition to optimization the software and possible improvements to the tool, the ideal would be to test the application in a real context and in different sectors to verify its usefulness. However, in other studies this tool has already been implemented in a real context, but with another version. Acknowledgement. This work was financially supported by the research unit on Governance, Competitiveness and Public Policy (UIDB/04058/2020) + (UIDP/04058/2020), funded by national funds through FCT - Fundação para a Ciência e a Tecnologia.

References 1. Brito, M.F., Ramos, A.L., Carneiro, P., Gonçalves, M.A.: A continuous improvement assessment tool, considering lean, safety and ergonomics. Int. J. Lean Six Sigma 11(5), 879–902 (2020) 2. Sharma, V., Dixit, A.R., Qadri, M.A.: Modeling Lean implementation for manufacturing sector. J. Model. Manag. 11(2), 405–426 (2016) 3. Womack, J.P., Jones, D.T., Roos, D.: The Machine that Changed the World. Free Press, Simon & Schuster, New York, NY (1990) 4. Dekier, L.: The origins and evolution of lean management system. J. Int. Stud. 5(1), 46–51 (2012) 5. Womack, J.P., Jones, D.T.: Lean Thinking: Banish Waste and Create Wealth in Your Corporation. Free Press, Simon & Schuster, 1996, New York, NY. Second Edition (2003) 6. Naranjo-Flores, A.A., Ramírez-Cárdenas, E.: Human factors and ergonomics for lean manufacturing applications. In: García-Alcaraz, J.L., Maldonado-Macías, A.A., Cortes-Robles, G. (eds.) lean Manufacturing in the Developing World, pp. 281–299. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-04951-9_13 7. Yusuff, R.M., Abdullah, N.S.: Ergonomics as a lean manufacturing tool for improvements in a manufacturing company. Proceedings of the International Conference on Industrial Engineering and Operations Management 8–10, 581–588 (2016) 8. Galante, J.J.: “lean ergonomics”, Technical Paper - Society of Manufacturing Engineers (2014)

204

P. C. Ribeiro et al.

9. Aqlan, F., Lam, S.S., Testani, M., Ramakrishnan, S.: Ergonomic risk reduction to enhance lean transformation. IIE Annual Conference and Expo 2013, 989–997 (2013) 10. Liker, J.K.: Becoming lean: Inside Stories of US Manufacturers. Productivity Press, New York, NY (1997) 11. Wan, H., Chen, F.F.: Decision support for lean practitioners: A web-based adaptive assessment approach. Comput. Ind. 60(4), 277–283 (2009) 12. Liker, J., Rother, M.: Why lean programs fail. In: Lean Enterprise Institute, pp. 1–5 (2011) 13. Narayanamurthy, G., Gurumurthy, A.: Leanness assessment: a literature review. Int. J. Oper. Prod. Manag. 36(10), (2016) 14. Wan, H., Chen, F.F.: A leanness measure of manufacturing systems for quantifying impacts of lean initiatives. Int. J. Prod. Res. 46(23), 6567–6584 (2008) 15. Karlsson, C., Åhlström, P.: Assessing changes towards lean production. Int. J. Oper. Prod. Manag. 16(2), 24–41 (1996) 16. Sánchez, M., Pérez, M.P.: Lean indicators and manufacturing strategies. Int. J. Oper. Prod. Manag. 21(11), 1433–1452 (2001) 17. Goodson, R.E.: Read a plant-fast. Int. J. Prod. Res. 80(5), 105–113 (2002) 18. Srinivasaraghavan, J., Allada, V.: Application of mahalanobis distance as a lean assessment metric. Int. J. Adv. Manuf. Technol. 29, 1159–1168 (2006) 19. Tarcisio Abreu Saurin, G.A.M., Ribeiro, J.L.D.: Quantifying benefits of conversion to lean manufacturing with discrete event simulation: a case study. Int. J. Prod. Res. 49(11), 3211– 3230 (2011) 20. Fernandez, J.E.: Ergonomics in the workplace. Facilities 13(4), 20–227 (1995) 21. Nunes, I.L.: Integration of ergonomics and lean six sigma: a model proposal. Procedia Manuf. 3, 890–897 (2015) 22. Tortorella, G.L., Vergara, L.G.L., Ferreira, E.P.: Lean manufacturing implementation: an assessment method with regards to socio-technical and ergonomics practices adoption ergonomics in the workplace. Int. J. Adv. Manuf. Technol. 89, 3407–3418 (2017) 23. Yazdani, A., et al.: Integration of musculoskeletal disorders prevention into management systems: a qualitative study of key informants’ perspectives. Saf. Sci. 104, 110–118 (2018) 24. Wilson, R.: Guarding the line. Ind. Eng. 37(4), 46–49 (2005) 25. Wong, W.P., Ignatius, J., Soh, K.L.: What is the leanness level of your organization in lean implementation? An integrated lean index using ANP approach. Prod. Plan. Control: The Manag. Oper. 25(4), 273–287 (2014) 26. Jarebrant, C., Winkel, J., Hanse, J.J., Mathiassen, S.E., Ojmertz, B.: ErgoVSM: a tool for integrating value stream mapping and ergonomics in manufacturing. Hum. Factors Ergon. Manuf. Serv. Ind. 22(6), 191–204 (2016) 27. Gonçalves, M.T., Salonitis, K.: Lean assessment tool for workstation design of assembly lines. Procedia CIRP 60, 386–391 (2017) 28. Nunes, I.L., Machado, V.C.: Merging ergonomic principles into lean manufacturing. In: Conference Proceedings of IIE Annual Conference and Expo 2007 – Industrial Engineering’s Critical Role in a Flat World, pp. 836–841 (2007) 29. Bhasin, S.: Measuring the leanness of an organization. Int. J. Lean Six Sigma 2(1), 55–74 (2011)

Deep Learning for Big Data Filipe Correia1(B) , Ana Madureira1,2 , and Jorge Bernardino3,4 1 ISEP/IPP, Porto, Portugal

{1150524,amd}@isep.ipp.pt

2 ISRC - Interdisciplinary Studies Research Center, Porto, Portugal 3 Polytechnic of Coimbra, ISEC, Coimbra, Portugal

[email protected] 4 CISUC - Centre of Informatics and Systems of University of Coimbra, Coimbra, Portugal

Abstract. We live in a world where data is becoming increasingly valuable and increasingly abundant in volume. All companies produce data from sales, sensors, and various other sources. The main challenges are how can we extract insights from such a rich data environment and if Deep Learning is capable of circumventing Big Data’s challenges. To reach a conclusion, Social Network data is used as a case study for predicting sentiment changes in the Stock Market. The main objective of this paper is to develop a computational study and analyze its performance. Deep Learning was able to handle some challenges of Big Data, allowing results to be obtained and compared with real world situations. The outputs contribute to understand Deep Learning’s usage with Big Data and how it acts in Sentiment Analysis. Keywords: Deep Learning · Big Data · Stock data · Financial markets · Social Networks

1 Introduction There is a growing academic effort to find ways to take advantage of the vast information capabilities provided by Big Data using Machine Learning (ML) to gather insights. Big Data is defined as: “significantly large datasets beyond the ability of traditional database software to store, capture, manage or analyze” [1]. Deep Learning (DL) is an approach to ML and allows system improvement through data insights [2, 3]. ML algorithms have seldom been as challenged such as by Big Data in obtaining knowledge. Big Data offers massive volumes of data and information for ML algorithms to work upon, allowing the extraction of patterns or building analytical models. DL is appointed as one of the possible ways to tackle some issues presented by Big Data, mainly feature engineering, non-linearity, data heterogeneity, uncertain, dirty, and noisy data [4, 5]. Recent advances in the field of neural networks have led to the creation of new deep structured architectures and related algorithms that makes them attractive for complex classification problems. In this paper, we propose that Deep Neural Networks (DNN) could be important in extracting value from Big Data. Our objective is investigating, developing, and implementing a system of automatic classification, based on DL, which © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 205–215, 2022. https://doi.org/10.1007/978-3-030-96299-9_20

206

F. Correia et al.

allows, in each scenario, to validate its adequacy and efficiency. The use case for which this system will be built is the classification of Sentiment of Social Network users regarding the values of Stock Markets. The rest of this paper is organized as follows. Section 2 and Section 3 present a brief state-of-the-art that cover similar results reported by other authors on Big Data and DL, respectively. Section 4 describes the methodology that was implemented for the proposed approach. Section 5 presents the computational study and the performance assessment. Finally, Sect. 6 presents the conclusions and some ideas for future work.

2 Big Data Big Data has been characterized by “7V’s” [5, 6], each with an explanation on what separates the concept of Big Data from that of traditional datasets: Volume (quantity or amount of data, which is usually higher than in traditional datasets); Variety (In type, nature, format, etc. This data can be comprised of images, videos, phrases, numbers, etc. and come from different sources); Velocity (the speed of data generation), Value (the insights one can take from it and the impact these may cause);Veracity (trustworthiness and quality of captured data); Volatility (related to Data retention, which means it can be destroyed) and Validity (the relation between Data and its correctness to the intended usage). Each of these concepts carries with it challenges for the analysis of Big Data. Table 1 illustrates some of these challenges. Table 1. Big Data Aspects and Challenges [6] Aspect

Challenges

Volume

Processing Performance; Curse of Modularity; Class Imbalance; Curse of Dimensionality; Feature Engineering; Non-Linearity; Bonferroni’s Principle; Variance and Bias

Velocity

Data availability; Real-Time Process/Streaming; Concept Drift; Independent and Identically Distributed Random Variables

Variety

Data Locality; Data Heterogeneity; Dirty and Noisy Data;

Veracity

Data Provenance; Data Uncertainty; Dirty and Noisy Data

Value

Context dependent

Volatility

Expensive Long-Time Storage; Unrecoverable Data

Validity

Lack of Data relationship; Context dependent

New technologies emerged for the different stages of Big Data, ranging from data extraction to data analytics. Storage techniques like Distributed File Systems (DFS) are designed to store very large data sets reliably. By distributing storage and computation across many servers, the resource can grow with demand while remaining economical at every size [7–9].

Deep Learning for Big Data

207

Techniques like parallel processing are essential to manage a massive volume of data in a timely manner. MapReduce and other MapReduce-like models [10] such as Spark are such techniques, which are capable of parallel processing. Spark [11] targets mainly machine learning and interactive querying workloads. It is built to take advantage of resilient distributed datasets (RDD) to achieve performance improvements over classic MapReduce [12]. Big Data is a candidate for complex new processes of storage, access, and processing. The effort to make Big Data more “accessible” is associated to the possible value that allows entities to make more informed decisions on their future.

3 Deep Learning Deep Neural Networks (DNN) are more complex versions of Artificial Neural Networks (ANN). DNNs are composed of multi-layer interconnected nodes with more than one level of hidden layer nodes. Having multiple hidden layers allows for parameter learning and classification, all in the same network. Convolutional Neural Networks (CNN) are mostly used for image and video processing and are a subtype of DNNs, which have been inspired by the visual cortex present in animals, that breaks inputs in smaller parts (e.g.: an image with 32 pixels being analyzed 5 pixels at a time) as a way to keep the network at a manageable node size [13]. CNNs are hierarchical architectures and are good for extracting position invariant features [14]. CNNs are not largely associated to NLP by themselves, but when used with positional encoders they have proved to be useful for NLP classification. Recursive Neural Networks (RvNN) are DNNs whose architecture allows the same set of weights to be applied recursively within a structural setting. RvNNs have a specific skewed tree structure, which allows RvNNs to perform well in NLP [15]. Recurrent Neural Networks (RNN) are a subtype of RvNNs with a specific structure. NLP is dependent on the order of words or sentences, so it is useful to have a “memory” of previous elements when processing new ones due to backward dependencies (e.g.: semantic word meaning may depend on the words before or after). RNNs accomplish this by combining the outputs of two layers, allowing phrase analysis in forward and backward directions (bidirectional Recurrent NNs) [16]. Long Short-Term Memory (LSTM) are a subtype of RNN which exchange single recursive nodes for several individual neurons (interconnected nodes), connected in a manner designed to retain, forget or expose important information [15]. GRU is a subcategory of LSTM, which does not contain a separate memory cell [17]. An LSTM cell controls the exposure of its memory to the other cells in the network, while the GRU always exposes all of it. An aspect which was reported to improve the performance of DL on text data was to use an attention mechanism. These mechanisms allow a NN to focus on certain aspects of the input, filtering out noise [18]. These different NN techniques have been used in different works with some success. The use of CNNs, LSTM and GRU by themselves provided good results with Social Network data [19]. There is a possibility to create hybrid models which combine CNNs with both GRUs and LSTMs. In this paper, we intend to use a CNN for sentiment analysis and LSTM for numerical data analysis as well.

208

F. Correia et al.

4 Methodology 4.1 Overview This analysis consisted of three phases. The first phase consisted of selecting an existing dataset, which had already been processed and used in other works. This data was used to train seven different algorithms, one for each NN layer type, as well as hybrid versions of them (CNN, LSTM, GRU, CNN-LSTM, CNN – GRU, CNN – BiGRU and CNN – LSTM with Stock Indicators), for Sentiment Analysis related to Stock Value changes. The second phase consisted of collecting data from different Social Networks and storing it in a DFS warehouse. This data was then used to test the models which were obtained in the training phase. The third phase consisted of two trading simulations based on the two best models in the test phase. To validate if there was any value in the information which was used for test and training, there is a baseline simulation where 1000$ are invested in the first day of the simulation and never removed, following the “will” of the market. 4.2 Materials and Tools For the training phase, we have used a dataset, which consists of three different tables. A table containing Dow Jones Industrial Average (DJIA) information, with Date, Open, High, Low, Close, Volume and Adjusted Close for the DJIA index, with 1990 lines. Another containing Reddit news headlines, containing only Date and News columns, with 74377 lines. The combination of these tables resulted in dataset with 1990 rows, each containing the date, the sentiment (1 for positive, 0 for negative) of the day and up to 25 different news headlines, extracted from reddit, for each date. This data was transformed into a structure which maps each news headline to a single row, so there are various rows for a single day, each with a single news headline. For the second phase, we have used the subsets shown in Table 2. Three different sources of data were collected. Stocks were combined with both Reddit and Twitter data. From Twitter were extracted tweets mentioning four different Stock Indexes (AAPL, BTC, DJIA, and TSLA). From Reddit were extracted Post titles from three different subreddits, which discuss stock values. For Twitter data, each subset was related to the Stock Index it represented. For Reddit data, since the collected sentences are related to general Stock Market Sentiment, a general Stock Index was used (DJIA). The difference in open and closing prices for stock data was calculated for each day. When the result was a negative value, the sentiment for all sentences of a certain day were labeled as negative. When the result was positive, the sentiment was labeled as positive. All algorithms except CNN – LSTM with Stock Indicators (CNN – LSTM SI) were trained and tested using sentence data as input, since this was the required input. The difference for this algorithm was the usage of numerical indicators, which contained the open, close, high, and low values of stock indexes. The different data sources give a perspective on how the different algorithms behave when presented with a new issue. Having multiple sources and multiple algorithms to test will give a clearer representation on how and if DL can overcome the characteristic of Variety.

Deep Learning for Big Data

209

Table 2. Collected Dataset Description Subset

Number of sentences

Number of days

Origin

AAPL

21936

12

Twitter

BTC

14600

18

Twitter

DJIA

12640

12

Twitter

TSLA

22800

12

Twitter

Stocks

21132

25

Reddit

Stock market

21203

28

Reddit

Wallstreetbets

20270

15

Reddit

All Reddit data

62605

28

Reddit

4.3 Development Each of the eight different algorithms were created and trained using the same technique of nested Cross Validation (nCV). This approach consists of a cross-validation cycle [21] inside another cross-validation cycle, where the inner cycle selects the model’s best hyperparameters, as shown in Fig. 1. These two loops work by dividing data into training and validation data, which the Tensorflow framework uses on its feedback and feedforward loops, providing a better fit during training. After the inner loop is finished, the resulting model is then used to classify the test data, which allows us to calculate the accuracy and loss of the model.

Fig. 1. Nested cross validation methodology

The algorithm CNN – LSTM with Stock Indicators used an adapted nCV methodology due to the nature of its data necessities. The input of this algorithm consists of four distinct multi-level matrices. The first contained sentence data for the day in analysis. The second and the third contained an aggregation of data for the previous week

210

F. Correia et al.

and month of sentence data respectively. The last contained the stock value indicators. This meant the algorithm had to be treated as a time series classifier. For this behaviour, the training was done in an incremental manner, where each iteration contained (where possible to keep in memory) all the previous iteration data. Apart from the CNN algorithm, all others were implemented to use embedding. This allowed a vocabulary to be established with the n most popular words in a dataset, mapping each word to a number, and with the help of GloVe, convert each of those numbers into arrays. The arrays consisted of relationships between words defined by numbers. This conversion was done in runtime, removing the need to keep the huge data arrays in memory. The CNN algorithm used a complex input which consisted of various embedding techniques, not only GloVe, where the position of the word in a sentence, the meaning and other factors are used to embed. This results in a higher memory requirement. Using as many different implementations gives a wider view on how different DL techniques influence the results of analysis of Big Data. It’s possible to see that some techniques have advantages and disadvantages, both in terms of training time, accuracy and other metrics.

5 Model Evaluation ML has a set of core performance indicators which can be used for a model’s selfevaluation. These can be extracted from a confusion matrix. A confusion matrix is a way to record and extract meaning from predictions and true values. False positives (FP) are values predicted to be positive but are negative. False negatives (FN) are values predicted to be negative but are positive. True positives (TP) are predicted and positive. True negatives (TN) are predicted and negative. These metrics allow the extraction of indicators, such as accuracy, precision, recall (or sensitivity), fallout, specificity, F1 Score, Matthews Correlation Coefficient (MCC), and many others [20]. Since the case study of this paper concerns stock data, for each trained model, an investment simulation was added. After finding out the two best performing models in the test data, these are used to simulate an investment strategy. The investment strategy consists of starting with 1000 dollars which are invested in the days when the model predicts a positive sentiment. When there is money invested, if the models predict a negative sentiment for a certain day, all the money is extracted. When a positive sentiment is predicted, all the money is invested, despite there being a loss or an earning compared to the initial investment. 5.1 Training Phase The training phase yielded the accuracy and loss of the different models as shown in Table 3. The values collected were done so by averaging each of the accuracies on test data from the nested cross-validation method of training. In terms of accuracy, CNN – LSTM SI appears to stand out in relation to the other models, likely due to the mix of NLP and numerical stock value data. CNN – LSTM SI is capable of a greater accuracy but at the same time takes much longer to train.

Deep Learning for Big Data

211

In terms of loss, CNN – LSTM (the one without Stock Index data) shows the worst results. This means it fails considerably more predictions than the other algorithms. All the other algorithm’s metrics seem to have low difference. The next section focuses on models which resulted from the training phase. The tests section is about exposing these models to new data. Data from different sources and with different contexts. Table 3. Training Indicators Algorithm

Accuracy

Loss

Train time

CNN

0.520

0.125

47 h

LSTM

0.644

0.073

14 h

GRU

0.534

0.025

22 h

CNN–GRU

0.632

0.105

16 h

CNN–BiGRU

0.591

0.093

17 h

CNN–LSTM

0.534

0.233

22 h

CNN–LSTM SI

0.726

0.125

20 days

5.2 Test Phase The test phase outputted further indicators. Table 4 contains the average of each extracted indicator, obtained by averaging the results for the different datasets. CNN has obtained the highest average accuracy result and CNN – LSTM SI the lowest, the opposite of the results obtained in the training phase. The results obtained in the CNN show that it is highly biased towards positive sentiment. Specificity is very low, possibly representing a low proportion of negatives that are correctly identified as such. This is supported by the results in both precision and sensitivity, the proportion of positive identifications which was correct and or the proportion of actual positives which was correctly identified. Fallout also shows a relation to this, since fallout is the ratio between the number of negative events wrongly categorized as positive (false positives) and the total number of actual negative events. The analysis of the results should not fall only on accuracy. The best accuracy results, discarding the biased CNN, are presented from the CNN – LSTM model. Despite being the lowest average scorer, CNN – LSTM SI had the best results when classifying the AAPL dataset, reaching a 69% accuracy, as well as 64% on the TSLA dataset, which can be a result of the use of numerical data for the stock indicators. It was also the model with the highest variance in the classification indicators. When using the MCC formula, it ends up as the one with the highest correlation coefficient while CNN ended up with the lowest, much as the training results. Figure 2 shows the distribution of the metrics in Table 4. Sensitivity and fallout indicate the training dataset may have been slightly unbalanced, containing more positive than negative values. This is corroborated by the specificity, which is lower, indicating a lower number of negative labels. Some outliers are attributed to the CNN algorithm,

212

F. Correia et al. Table 4. Test Indicators

Algorithm

Accuracy Specificity F1score MCC

CNN

0.575

0.011

0.727

−0.001 0.577

0.99

0.989

LSTM

0.507

0.396

0.579

−0.003 0.573

0.6

0.604

GRU

0.506

0.449

0.551

0.57

0.551

CNN–LSTM

0.525

0.409

0.597

0.601

0.591

CNN–LSTMSI 0.486

0.676

0.366

0.03

0.489

0.316

0.324

CNN–GRU

0.521

0.302

0.625

0.007 0.579

0.704

0.698

CNN–BiGRU

0.518

0.331

0.611

−0.004 0.574

0.665

0.669

0.02

Precision Sensitivity Fallout

0.574

−0.022 0.596

which is highly biased, outputting mostly positive outcomes. This is due to overfitting which occurred during training. CNN-LSTM SI also ended up having outliers related to the mean and average of each indicator. These outliers show a reverse trend of more negative than positive predictions, being a conservative predictor. Given the results, which were obtained in the accuracy metric favor a biased model whereas MCC penalizes it (supported by the other metrics) the algorithm selected to run the investment simulation in the next section was based on the MCC indicator where CNN – LSTM SI had the best result.

Fig. 2. Test Indicator distribution

The selected datasets for the simulation were the ones with the best performance for the selected model. These were also the best results from all the samples. The first dataset is the AAPL dataset, which had the second best MCC and the best accuracy. The second dataset is the one with the best MCC result and second-best accuracy, the TSLA

Deep Learning for Big Data

213

dataset. The next section will focus on simulating an investment strategy using the two mentioned dataset’s results. 5.3 Simulation Phase This section covers the steps taken to simulate an investment strategy based on the model CNN – LSTM SI’ prediction results on the AAPL dataset and the TSLA dataset. The simulation followed these rules: The starting balance is 1000$; If the predicted sentiment is 1, the full balance is invested; If the predicted sentiment is 0, the full balance is removed from the investment and placed into savings; Each time the money is invested it was completely, leaving savings empty; In the end, the profit or loss is obtained by calculating the difference of balance and the initial investment; The Return on investment (ROI) is the profit or loss divided by the initial investment; To Provide a baseline, 1000$ are invested in the first day and never removed. Table 5 contains a summarized version of what happened in this simulation. Table 5. Summary of the simulation Simulation

Baseline

Baseline ROI

Model strategy

Model strategy ROI

TSLA

1091.49$

9.1%

1043.95$

4.4%

AAPL

970.09$

−3.0%

979.04$

−2.1%

In both cases, the algorithm followed the trend for the financial index for the analyzed period, between 13/09/2021 and 28/09/2021. The AAPL index fell 3%, while the TSLA increased 9.1% as shown by the simulation baseline. The AAPL ROI was higher than the baseline but still ended in a loss. Meanwhile the TSLA ROI was smaller than the baseline and ended in profit. These results seem to indicate that the model ended up smoothing the risk associated to the investment.

6 Conclusions and Future Work We conclude that some DL algorithms require data structures which are too complex and too big to store in RAM. A way to avoid this problem is to break data in smaller batches and train the algorithms in incremental intervals. The CNN – LSTM SI ended up being significantly worse in terms of training time than the others as it used this method. This is a point which could influence the selection of the algorithm to use, as well as the method used to train it. In this case, this method was used as the algorithm behaves as a time series classifier. In the Test and Simulation phases, Reddit and Twitter are not exactly moderated by specialists in the subject in analysis. This means that the sentiments may have no correlation to the stock values which were analyzed. The data seems to have some Value however, as it returned a profit in one of the simulations and smoothed the loss in the other.

214

F. Correia et al.

As future work, sentiment analysis from professional stock traders may help improve classification. It’s likely to be Valuable data with higher Veracity associated. Another change would be to use different data contexts for training. These models have only been trained on the context of DJIA stock data with Reddit news headlines. DJIA does not have high value changes, so it’s stock value usually increases. In terms of Big Data technologies, future work could focus on the usage of multiple nodes for DFS as well as Spark, as well as other tools which were not used in this paper.

References 1. Wang, W.Y.C., Wang, Y.: Analytics in the era of big data: the digital transformations and value creation in industrial marketing. Ind. Mark. Manag. Elsevier 12–15 (2020) 2. Lecun, Y., Bengio, Y., Hinton, G.: Deep learning [Internet]. Nature. Nature Publishing Group. http://colah.github.io/ (2015). Accessed Jan 25 2021, pp. 436–444 3. Ahad, M.A., Tripathi, G., Agarwal, P.: Learning analytics for IoE based educational model using deep learning techniques: architecture, challenges and applications. Smart Learn Environ 2018 51 [Internet]. SpringerOpen. https://slejournal.springeropen.com/articles/https:// doi.org/10.1186/s40561-018-0057-y (2018). Accessed Nov 13 2021, vol. 5, pp. 1–16 4. L’Heureux, A., Grolinger, K., Elyamany, H.F., Capretz, M.A.M.: Machine learning with Big Data: Challenges and approaches. IEEE Access. Institute of Electrical and Electronics Engineers Inc. 5, 7776–7797 (2017) 5. Zhou, L., Pan, S., Wang, J., Vasilakos, A.V.: Machine learning on big data: opportunities and challenges. Neurocomputing 237, 350–361 (2017) 6. Khan, M., Uddin, M.F., Gupta, N.: Seven V’s of Big Data understanding Big Data to extract value. In: Proc 2014 Zo 1 Conf Am Soc Eng Educ - “Engineering Educ Ind Involv Interdiscip Trends”, ASEE Zo 1 2014. IEEE Computer Society (2014) 7. Shvachko, K., Kuang, H., Radia, S., Chansler, R.: The Hadoop distributed file system. In: 2010 IEEE 26th Symp. Mass Storage Syst. Technol. MSST2010. IEEE Computer Society (2010) 8. Ahad, M.A., Biswas, R.: Comparing and analyzing the characteristics of Hadoop, Cassandra and Quantcast file systems for handling big data. Indian J. Sci. Technol. [Internet]. The Indian Society of Education and Environment. https://indjst.org/articles/comparing-and-analyzingthe-characteristics-of-hadoop-cassandra-and-quantcast-file-systems-for-handling-big-data (2017). Accessed Nov 13 2021, vol. 10, pp. 1–6 9. Ahad, M.A., Biswas, R.: Request-based, secured and energy-efficient (RBSEE) architecture for handling IoT big data: doi: 101177/0165551518787699 [Internet]. SAGE Publications, Sage UK, London, England. https://journals.sagepub.com/doi/https://doi.org/10.1177/ 0165551518787699 (2018) Accessed Nov 13 2021, vol. 45, pp. 227–238 10. Ji, C., Li, Y., Qiu, W., Awada, U., Li, K.: Big data processing in cloud computing environments. In: Proc 2012 Int Symp Pervasive Syst Algorithms, Networks, I-SPAN 2012, pp. 17–23 (2012) 11. Inoubli, W., Aridhi, S., Mezni, H., Maddouri, M., Mephu Nguifo, E.: An experimental survey on big data frameworks. Futur. Gener. Comput. Syst. 86, 546–564 (2018) 12. Kumar Vavilapalli, V., Murthy, A.C., Douglas, C., Agarwal, S., Konar, M., Evans, R. et al.: Apache Hadoop YARN: Yet Another Resource Negotiator. doi: https://doi.org/10.1145/252 3616.2523633 (2013). Accessed Feb 3 2021, vol. 13, pp. 1–3 13. Albawi, S., Mohammed, T.A., Al-Zawi, S.: Understanding of a convolutional neural network. In: Proc. 2017 Int. Conf. Eng. Technol. ICET 2017, pp. 1–6. Institute of Electrical and Electronics Engineers Inc. (2018)

Deep Learning for Big Data

215

14. Yin, W., Kann, K., Yu, M., Schütze, H.: Comparative study of CNN and RNN for natural language processing. http://arxiv.org/abs/1702.01923 (2017). Accessed Feb 27 2021 15. Irsoy, O., Cardie, C.: Deep recursive neural networks for compositionality in language. Adv. Neural. Inf. Process. Syst. 2096–2104 (2014) 16. Otter, D.W., Medina, J.R., Kalita, J.K.: A survey of the usages of deep learning in natural language processing. arXiv. arXiv (2018) 17. Yang, Z., Yang, D., Dyer, C., He, X., Smola, A., Hovy, E.: Hierarchical Attention Networks for Document Classification (2016) 18. Zhou, X., Wan, X., Xiao, J.: Attention-based LSTM network for cross-lingual sentiment classification. In: EMNLP 2016 – Conf. Empir. Methods Nat. Lang. Process Proc., pp. 247– 256 (2016) 19. Basiri, M.E., Nemati, S., Abdar, M., Cambria, E., Acharya, U.R.: ABCDM: An Attentionbased Bidirectional CNN-RNN Deep Model for sentiment analysis. Futur. Gener. Comput. Syst. 115, 279–294 (2021) 20. Handelman, G.S., Kok, H.K., Chandra, R.V., Razavi, A.H., Huang, S., Brooks, M., et al.: Peering Into the Black Box of Artificial Intelligence: Evaluation Metrics of Machine Learning Methods. Am. J. Roentgenol. [Internet]. American Roentgen Ray Society 212, pp. 38–43. https://www.ajronline.org/ (2019). Accessed Feb 28 2021 https://doi.org/10.2214/AJR.18. 20224

A Review on MOEA and Metaheuristics for Feature-Selection Duarte Coelho1,4 , Ana Madureira2 , Ivo Pereira1,2,3(B) , and Ramiro Gon¸calves4 1 E-goi, Matosinhos, Portugal {dcoelho,ipereira}@e-goi.com 2 Interdisciplinary Studies Research Center (ISRC), ISEP/IPP, Porto, Portugal [email protected] 3 University Fernando Pessoa (UFP), Porto, Portugal [email protected] 4 Universidade de Tr´ as-os-Montes e Alto Douro (UTAD), Vila Real, Portugal [email protected]

Abstract. In the areas of machine-learning/big data, feature selection is normally regarded as a very important problem to be solved, as it directly impacts both data analysis and model creation. The problem of optimizing the selected features of a given dataset is not always trivial, however, throughout the years various ways to counter this optimization problem have been presented. This work presents how feature-selection fits in the larger context of multi-objective problems as well as a review of how both multi-objective evolutionary algorithms and metaheuristics are being used in order to solve feature selection problems. Keywords: Big data · Feature selection · Multi-objective Evolutionary algorithms · Machine-learning

1

·

Introduction

Data is at the core of machine learning. However, when collecting data, there are cases when too many features may end up being stored in order to solve a given problem. Features may be redundant or even irrelevant for the problem, meaning they only add noise to the dataset [1]. In order to deal with this problem, feature selection is used to find the best subset of features that transmit most of the important information contained in the initial dataset. This step is a regular occurrence in any machine/deep learning pipeline (as seen in Fig. 1). By removing data related errors/anomalies, various benefits can be ripped, such as better model interpretability, shorter training times, and reduced overfitting risk [2]. Additionally, depending on the problem, these algorithms may be able to reduce the operational and risk costs (i.e. in clinical trials, or even cases related to field sampling) [1]. c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 216–225, 2022. https://doi.org/10.1007/978-3-030-96299-9_21

A Review on MOEA and Metaheuristics for Feature-Selection

217

Fig. 1. Example of a regular machine-learning pipeline (adapted from [3])

The literature for feature selection proposes many methods. Generally, two main types can be referred, wrapper and filter techniques [1]. In the filter approach, features are selected based on a performance metric independently of the classifier being used. In wrapper approaches, a classifier is used to test each feature subset, which means they are classifier-dependent [1]. They distinguish themselves from one another via their speed and efficiency: “filter methods are usually faster than wrapper methods since they have lower computational cost, however, wrapper methods have usually better performance than the filter methods since they select more representative features” [2]. The main problem tackled by the different methods regards the process of finding the best possible local solution, since searching all possible subset spaces for a considerable feature number becomes an impossible operation [1,2]. When the idea of feature-selection first arose, it was mostly defined as a single objective optimization problem. Since single metrics like accuracy are the only objective to be optimized. Nowadays, multi-objective oriented approaches have been taken regarding this theme. In addition to simple metrics such as accuracy, this class of problems includes multiple objectives. Such as generalization capability for supervised classifiers, and bias counter-balancing, with a variable amount of features, for unsupervised approaches [4]. One type of solution that as been shown to be successfully used to solve this type of multi-objective problem has been meta-heuristics. Both classical approaches as well as more recent ones such as nature inspired methods have shown good results in different scenarios [5]. In the following section the various notions needed to understand the problem will be presented. Firstly the problem will be described, followed by presenting important notions about each relevant area, then detailing some literature review on those fields and, lastly, discussing limitations as well as future work.

218

2

D. Coelho et al.

Multi-Objective Evolutionary Algorithms

Let us first understand how Multi-Objective Evolutionary Algorithms (MOEAs) may be defined so then we can build upon that knowledge. 2.1

Definition

In the context of multi-objective optimization, the principle of finding the most optimal solution cannot be applied to one objective alone, since the various other objectives are equally important. Meaning multi-objective optimization can be expressed by two goals [6]: convergence and diversity. Convergence seeks tofind a (finite) set of solutions which lies on the Pareto-optimal1 front. On the other hand, Diversity tries to find a set of solutions which is diverse enough to represent the entire range of the Pareto-optimal front. MOEA try to establish their identity by following both the principles stated above [6]. Figure 2 schematically shows the principles followed in an MultiObjective Evolutionary Procedures (MOEPs). Since MOEPs are heuristic based, they may not guarantee finding exact Pareto-optimal points solutions. However, MOEPs have operators to constantly improve the evolving non-dominated points similar to how most natural and artificial evolving systems continuously improve their solutions. The main difference and advantage of using a MOEA compared to a posteriori Multiple Criteria Decision-making Method (MCDM) is that multiple trade-off solutions can be found in a single run of an MOEA, whereas most a posteriori MCDM methodologies would require multiple independent runs [6].

Fig. 2. Schematic of a two-step multi-criteria optimization and decision-making procedure (adapted from [6])

2.2

Types of Algorithms

Now that a definition for MOEAs has been presented, it is possible to introduce some of the broader notions used in multiple objective evolutionary approaches. In order to present this broader scope the following section serves as a short review of various frameworks used to develop MOEAs. 1

Pareto optimality can be roughly defined as a state at which resources in a given system are optimized such that one dimension cannot improve without a second worsening. “The main idea (...) is that a society is enjoying maximum ophelimity when no one can be made better off without making someone else worse off” [7].

A Review on MOEA and Metaheuristics for Feature-Selection

219

Domination Based Algorithms. This group of algorithms uses the dominance relation in the fitness assignment process, thus following suggestions presented in [8]. As the dominance relation itself does not preserve the diversity in the population, another techniques, such as niching, are needed to obtain a good spread of solutions. These algorithms have been very popular since mid 1990s [8]. An especially relevant algorithm in this category is refered as NSGA-II. Its main advantage is its effectiveness which even makes it competitive with the modern algorithms, when optimizing for a low number of objectives [9]. Decomposition Based Algorithms. A recent framework is that of MultiObjective Evolutionary Algorithms based on Decomposition (MOEA/D) [10]. It is based on conventional aggregation approaches in which an MOP is decomposed into a number of Scalar Objective optimization Problems (SOPs). The objective of each SOP, also called a sub-problem, is a (linearly or nonlinearly) weighted aggregation of the individual objectives. Neighborhood relations among these sub-problems are defined based on the distances between their aggregation weight vectors. Each sub-problem is optimized in the MOEA/D by using information mainly from its neighboring sub-problems. A major advantage of MOEA/Ds is that a scalar objective local search can be used in each sub-problem in a natural way since its task is to optimize single objective [10]. Preference Based Algorithms. Due to the conflicts among the objectives in MOPs, the total number of Pareto optimal solutions might be very large or even infinite. However, the decision maker (DM) may be only interested in preferred solutions instead of all. To find the preferred solutions, preference information is needed to guide the search towards the region of interest to the DM. Based on the role of the DM in the process, multi-objective optimization methods can be classified into priori methods, posteriori methods, and interactive methods [11]. In a priori method, preference information is given by the DM before the solution process. An MOP can be converted into an SOP. Then, a scalar objective solver is applied to find the desired Pareto optimal solution. A posteriori method uses the DM’s preference information after the search process. A well distributed approximation is first obtained. Then, the DM selects the most preferred solutions based on its preferences. In an interactive method, the intermediate search results are presented to the DM to investigate; then the DM can understand the problem better and provide more preference information for guiding the search [11]. The earliest attempts on MOEAs based on the DM’s preference were made by Fonseca and Fleming [12] and Tanino et al. [13] in 1993. Indicator Based Algorithms. Indicator-based MOEAs use an indicator to guide the search, particularly to perform solution selection. Zitzler and K¨ unzli [14] first suggested a general Indicator-Based Evolutionary Algorithm (IBEA). This approach uses an arbitrary indicator to compare a pair of candidate solutions. In comparison to other MOEAs, the IBEA only compares pairs of individuals instead of entire approximation sets. In [15], Basseur and Zitzler proposed an indicator-based model for handling uncertainty, in which each solution

220

D. Coelho et al.

is assigned a probability in the objective space. In an uncertain environment, some methods for computing expected indicator values are discussed, and several variants of their -indicator-based model are suggested and empirically investigated. Recently, Bader and Zitzler [16] further investigated the robustness of hypervolume-based multi-objective search methods. Three existing approaches for handling robustness in the area of evolutionary computing, modifying objective functions, adding objectives, and adding robustness constraints, are integrated into a multi-objective hypervolume-based search. An extension of the hypervolume indicator is also proposed for robust multi-objective optimization. Hybrid Algorithms. In MOEAs, there are many techniques which have different characteristics and advantages. Hybridizing these techniques is thus a natural choice to utilize their advantages when dealing with complicated MOPs. What techniques to use and how to hybridize them are two major problems to solve when designing a hybrid MOEA. Some recent work could thus be categorized: – Hybridizing different search methods: A general idea is to combine global search and local search methods, known as the memetic approach [17]. Another widely used idea is to combine the search operators of different algorithms. Particle Swarm Optimization (PSO) and Evolutionary Algorithms (EA) are hybridized in [18]. – Hybridizing search and updating methods: This strategy hybridizes different components from different algorithms. For example, in [18], the PSO’s operator is inserted into an EA’s main loop. – Hybridizing different methods in different search phases: in the strategies above, hybrid methods are used in each generation. It is also possible to partition a search process into different phases and to use different search strategies in these phases. For example, in [19], to emphasize dominated solutions, balance dominated and non-dominated solutions, and focus on non-dominated solutions, the search is partitioned into three phases. NSGA-II and a local incremental search algorithm are used to achieve the goals.

3

MOEA Based Feature Selection

This section will present different approaches used to deal with the problem of feature selection. Huang [20] proposed a multi-objective feature selection approach applied to churn prediction in the telecommunication service field, based on the optimization approach NSGA-II. The main idea is to modify the approach NSGA-II to select local feature subsets of various sizes, and then use the method of searching non-dominated solutions to select the global non-dominated feature subsets. Gaspar-Cunha [21] introduced an optimization methodology based on the use of MOEAs in order to deal with problems of feature selection in the context of cardiac diagnosis. For that purpose a Support Vector Machines (SVM) classifier was adopted. The aim being to select the best features and optimize the classifier

A Review on MOEA and Metaheuristics for Feature-Selection

221

parameters simultaneously while minimizing the number of features necessary and maximize the accuracy of the classifier and/or minimize the errors obtained. That is a A reduced Pareto set genetic algorithm (elitist) (RPSGAe) was adopted in by using an SVM to reduce the size of the Pareto optimal set. The obtained results were favorable to the approach. Xue [22] suggested the adaption of NSGAII and SPEA2, in order to create two different filter based feature selection frameworks. Four multi-objective feature selection methods were then developed by applying mutual information and entropy as two different filter evaluation criteria in each of the two proposed frameworks. The results reached by the authors show that the proposed multiobjective algorithms can automatically evolve a set of non-dominated solutions that include a smaller number of features and achieve better classification performance than using all features. Additionally, NSGAII seems to achieves similar performance to SPEA2 for the datasets that consist of a small number of features and slightly better performance when the number of features is larger. Yong [23] the author proposed the use of an adapted multi-objective feature selection algorithm in order to deal with unreliable data. It accomplishes this by taking an effective multi-objective feature selection algorithm based on barebones particle swarm optimization and incorporating two new operators. One is a reinforced memory strategy, which is designed to overcome the degradation phenomenon of particles. Another is a hybrid mutation, which is designed to improve the search ability of the proposed algorithm. Comparison results suggest that the proposed algorithm is highly competitive for the proposed context.

4

Metaheuristics Based Feature Selection

Metaheuristics were first brought forward to define heuristic2 frameworks that could be applied to a large set of different problems (mainly optimization). A metaheuristic can be taken as a generic algorithm framework which may be applied to various optimization problems with a relative low amount of effort [24]. Regarding hybrid metaheuristics, these can mostly be divided in two categories: those which combine parts of other metaheuristics, and those that combine parts of other techniques. Since most recent advances in the area rely on the hybridization of metaheuristics, most of the works presented bellow will reflect hybrid metaheuristics [25–28]. L. Mousin et al. [25] presented an approach to feature selection based on a local-search metaheuristic. The authors consider the Feature Selection problem for classification as a combinatorial optimization one. They re-implement a tabu search algorithm firstly developed to solve a railway network problem, and then propose a learning mechanism in order to increase its performance. This learning mechanism works as a map that records the estimation of quality 2

Involving or serving as an aid to learning, discovery, or problem-solving by experimental and especially trial-and-error methods (in Merriam-Webster Dictionary).

222

D. Coelho et al.

of each combination of features, which are computed from the quality of solutions where those combinations appear. This accelerates future iterations and is related to the pheromones concept of ant colony optimization. Afterwards, various experiments are performed in order to measure its efficiency. According to a data-mining perspective the authors solution ends up being better performing then the base algorithm, which may be explained by the small number of features selected by the proposed algorithm. M. Mafarja et al. [26] suggested another approach based on a different local-search metaheuristic. Their proposed approach combined a simulated annealing (SA) algorithm with the global search capabilities of a whale optimization algorithm (WO). Two different hybrid models were created, in one SA was used as a local search operator around the selected search agents in WO. By contrast, in the second one, SA was used to search the neighborhood of the best found solution after each iteration of WO. Both approaches performance was measured and compared. Two criteria were reported to evaluate each approach: classification accuracy, average selection size. It was found that the second approach, which used SA to intensify the neighboring region of the best solution found in each iteration of WO and tournament selection to select the search agents, showed the best performance among all proposed models. M. Mafarja et al. [27] proposed two implementations of variants of an hybrid ant lion optimizer (ALO) for feature selection plus two different hillclimbing algorithms. One of these hill-climbing algorithms was quick reduct. Quick reduct is a set-based filter method for feature selection that simulates the forward generation method where the algorithm starts from an empty set and only features that improve a fitness value are added. The other hillclimbing method is an algorithm for reduction of knowledge with computing core (CEBARKCC), it works by finding the core features and adding them to the feature subset. Both implementations were tested over various datasets and the approach combining ALO and quick reduct showed best results in terms of accuracy while the one combining ALO and CEBARKCC performed better regarding minimal reducts. Additionally, the authors claim that both approaches performed better than other hybridized ALO methods in most case studies. Qasem et al. [28] put forward a binary version of an hybrid metaheuristic based on grey wolf optimization (GWO) and particle swarm optimization (PSO). The authors argue that this is necessary since feature optimization is inherently a binary problem. They proceed to evaluate the proposed approach. In order to find the best solutions, the wrapper-based method K-nearest neighbors classifier with Euclidean separation metric is used. A set of evaluation measures over eighteen datasets were used to assess the proposed method. The results show that the proposed binary hybrid approach significantly outperformed the binary GWO, the binary PSO, the binary genetic algorithm, and the whale optimization algorithm with simulated annealing when using various performance measures including accuracy while selecting the best optimal features.

A Review on MOEA and Metaheuristics for Feature-Selection

5

223

Discussion

Nowadays, most work being produced in the meta-heuristics research field is concerned with the development of hybrid approaches. The greatest problem with these, is not one of performance or limitations but one of comprehension and lack of information. What Christian Blum [29] stated in 2010 remain mostly true: “the process of designing and implementing hybrid metaheuristics is rather complicated and involves knowledge about a broad spectrum of algorithmic techniques, programming and data structures, as well as algorithm engineering and statistics”. This remains true up to today. This is mostly due to the great amount of metaheuristics that exist that could contribute with as a positive feature for a given hybrid approach. Not to mention it is even possible to make combinations of hybrid approaches. In a way, this problem too, is one inherently related to combinatorial optimization. When in the realm of hybrid approaches the solutions obtained start to rely more on the general planning and structure of the framework used then on the base algorithms being used. This fact should not dissuade researchers from continuing working on the area, because due to the great amount of metaheuristics there is always some extra component that might show positive contributions if applied to an hybrid approach. Similarly, the research and application in evolutionary multi-objective optimization over the in recent years has resulted in a number of efficient algorithms. MOEAs are now regularly applied to different problems in most areas of science, engineering, and commerce using in many cases metaheuristic based approaches. One area that currently seems especially enticing for researchers are collaborative EMO-MCDM (Evolutionary Multi-objective Optimization - Multiple-Criteria Decision-Making) algorithms for achieving a complete multi-objective optimization task to find a set of trade-off solutions and finally arriving at a single preferred solution. Another direction taken by researchers is to address guaranteed convergence and diversity of EMO algorithms through their hybridization with mathematical and numerical optimization techniques similar to the current trend with metaheuristics [6]. When it comes to the MOEAs known as the bleeding-edge, these are the ones based on hybridization approaches, such as the ones used in [18]. Even then more regular approaches such as NSGA and its variants, or algorithms as common as PSO can be found being applied to the problematic of feature-selection. It is possible to conclude that feature-selection continues to be recognized as a challenging multi-objective problem to solve. However, the creation of new metaheuristics and the merging or inter-operation of existing ones, through hybridization, also keeps producing better and more accurate solutions to face it.

6

Conclusion

Feature selection is an essential part of any machine-learning pipeline. This stage is important since the number of features of a model as well as their quality have been proven to affect model performance [30]. Feature selection has been one of

224

D. Coelho et al.

the multi-objective problems long tackled by metaheuristics, since it is essentially a combinatorial optimization problem. Nowadays, most work in meta-heuristics is produced via hybridization of metaheuristics, between themselves, and other various techniques. That is to say, it refers creation of new and more effective approaches by combining older metaheuristics. This being the case it is natural that various hybrid algorithms have tried to tackle the long existing problem that is feature selection. As we have seen throughout this work, results presented by solutions to this problem have been very promising and in the future even better performances seem all but assured. Acknowledgments. This article is a result of the project “Cria¸ca ˜o de um N´ ucleo de I&D para a gera¸ca ˜o de novo conhecimento nas ´ areas de Inteligˆencia Artificial, Machine Learning, Intelligent Marketing e One-2-One Marketing”, supported by Operational Programme for Competitiveness and Internationalization (COMPETE 2020), under the PORTUGAL 2020 Partnership Agreement, through the European Regional Development Fund (ERDF).

References 1. Talbi, E.-G., et al.: Hybrid Metaheuristics. Springer, Heidelberg (2013). https:// doi.org/10.1007/978-3-642-30671-6 ¨ 2. Zorarpacı, E., Ozel, S.A.: A hybrid approach of differential evolution and artificial bee colony for feature selection. Expert Syst. Appl. 62, 91–103 (2016) 3. Olson, R.: TPOT (tree-based pipeline optimization tool) (2017) 4. Sheth, P., Patil, S.: A review on feature selection problem solving using multiobjective evolutionary optimization algorithms. Int. J. Eng. Appl. Sci. Technol. 2(9), 42–54 (2018) 5. Diao, R., Shen, Q.: Nature inspired feature selection meta-heuristics. Artif. Intell. Rev. 44(3), 311–340 (2015). https://doi.org/10.1007/s10462-015-9428-8 6. Deb, K.: Multi-objective evolutionary algorithms. In: Kacprzyk, J., Pedrycz, W. (eds.) Springer Handbook of Computational Intelligence. Springer Handbooks, pp. 995–1015. Springer, Heidelberg (2015). https://doi.org/10.1007/978-3-662-435052 49 7. Luc, D.T.: Pareto optimality. In: Chinchuluun, A., Pardalos, P.M., Migdalas, A., Pitsoulis, L. (eds.) Pareto Optimality, Game Theory and Equilibria. Springer Optimization and Its Applications, vol. 17, pp. 481–515. Springer, New York (2008). https://doi.org/10.1007/978-0-387-77247-9 18 8. Goldenberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learning, 13th edn. Addison-Wesley Professional, Reading (1988) 9. Ishibuchi, H., Tsukamoto, N., Nojima, Y.: Evolutionary many-objective optimization: a short review. In: 2008 IEEE Congress on Evolutionary Computation (IEEE World Congress on Computational Intelligence), pp. 2419–2426. IEEE (2008) 10. Zhang, Q., Li, H.: MOEA/D: a multiobjective evolutionary algorithm based on decomposition. IEEE Trans. Evol. Comput. 11(6), 712–731 (2007) 11. Miettinen, K.: Nonlinear Multiobjective Optimization, vol. 12. Springer Science & Business Media (2012). https://doi.org/10.1007/978-1-4615-5563-6 12. Fonseca, C.M., Fleming, P.J., et al.: Genetic algorithms for multiobjective optimization: formulation discussion and generalization. In: ICGA, vol. 93, pp. 416– 423. Citeseer (1993)

A Review on MOEA and Metaheuristics for Feature-Selection

225

13. Tanino, T., Tanaka, M., Hojo, C.: An interactive multicriteria decision making method by using a genetic algorithm. In: 2nd International Conference on Systems Science and Systems Engineering (1993) 14. Zitzler, E., K¨ unzli, S.: Indicator-based selection in multiobjective search. In: Yao, X., et al. (eds.) Parallel Problem Solving from Nature - PPSN VIII. PPSN 2004. Lecture Notes in Computer Science, vol. 3242, pp. 832–842. Springer, Heidelberg (2004). https://doi.org/10.1007/978-3-540-30217-9 84 15. Basseur, M., Zitzler, E.: Handling uncertainty in indicator-based multiobjective optimization. Int. J. Comput. Intell. Res. 2(3), 255–272 (2006) 16. Bader, J., Zitzler, E.: Robustness in hypervolume-based multiobjective search. Computer Engineering and Networks Laboratory (TIK), ETH Zurich, TIK Report, vol. 317 (2010) 17. Lara, A., Sanchez, G., Coello, C.A.C., Schutze, O.: HCS: a new local search strategy for memetic multiobjective evolutionary algorithms. IEEE Trans. Evol. Comput. 14(1), 112–132 (2009) 18. Elhossini, A., Areibi, S., Dony, R.: Strength pareto particle swarm optimization and hybrid EA-PSO for multi-objective optimization. Evol. Comput. 18(1), 127–156 (2010) 19. Yang, D., Jiao, L., Gong, M.: Adaptive multi-objective optimization based on nondominated solutions. Comput. Intell. 25(2), 84–108 (2009) 20. Huang, B., Buckley, B., Kechadi, T.-M.: Multi-objective feature selection by using NSGA-II for customer churn prediction in telecommunications. Expert Syst. Appl. 37(5), 3638–3646 (2010) 21. Gaspar-Cunha, A.: Feature selection using multi-objective evolutionary algorithms: application to cardiac SPECT diagnosis. In: Rocha, M.P., Riverola, F.F., Shatkay, H., Corchado, J.M. (eds.) Advances in Bioinformatics. AISC, vol. 74, pp.85–92. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-13214-8 11 22. Xue, B., Cervante, L., Shang, L., Browne, W.N., Zhang, M.: Multi-objective evolutionary algorithms for filter based feature selection in classification. Int. J. Artif. Intell. Tools 22(04), 1350024 (2013) 23. Yong, Z., Dun-wei, G., Wan-qiu, Z.: Feature selection of unreliable data using an improved multi-objective PSO algorithm. Neurocomputing 171, 1281–1290 (2016) 24. Blum, C., Roli, A.: Metaheuristics in combinatorial optimization: overview and conceptual comparison. ACM Comput. Surv. (CSUR) 35(3), 268–308 (2003) 25. Mousin, L., Jourdan, L., Kessaci Marmion, M.E., Dhaenens, C.: Feature selection using Tabu search with learning memory: learning Tabu search. In: Festa, P., Sellmann, M., Vanschoren, J. (eds.) LION 2016. LNCS, vol. 10079, pp. 141–156. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-50349-3 10 26. Mafarja, M.M., Mirjalili, S.: Hybrid whale optimization algorithm with simulated annealing for feature selection. Neurocomputing 260, 302–312 (2017) 27. Mafarja, M.M., Mirjalili, S.: Hybrid binary ant lion optimizer with rough set and approximate entropy reducts for feature selection. Soft Comput. 23(15), 6249–6265 (2019) 28. Al-Tashi, Q., Kadir, S.J.A., Rais, H.M., Mirjalili, S., Alhussian, H.: Binary optimization using hybrid grey wolf optimization for feature selection. IEEE Access 7, 39496–39508 (2019) 29. Blum, C., Puchinger, J., Raidl, G., Roli, A., et al.: A brief survey on hybrid metaheuristics. In: Proceedings of BIOMA, pp. 3–18 (2010) 30. Sugumaran, V., Ramachandran, K.: Effect of number of features on classification of roller bearing faults using SVM and PSVM. Expert Syst. Appl. 38(4), 4088–4096 (2011)

Detection and Classification of Age-Related Macular Degeneration Using Integration of DenseNet169 and Convolutional Neural Network F. Ajesh1,2(B)

and Ajith Abraham1,3

1 Machine Intelligence Research Labs (MIR Labs) Scientific Network for Innovation and

Research Excellence, Auburn, WA 98071, USA [email protected] 2 Department of Computer Science and Engineering, Sree Buddha College of Engineering, Alappuzha, Kerala, India 3 Center for Artificial Intelligence, Innopolis University, Innopolis, Russia

Abstract. Age-related macular degeneration (AMD) is a leading cause of vision loss and blindness around the world. With an increase in age, the number of people impacted by the disease is observed to be growing. Knowledge about the occurrence of AMD should be used to develop appropriate eye care for these people. So, in this paper, we present an AMD detection and classification using DenseNetCNN. Data is collected from various repositories such as AREDS, Optretina and STARE. These are initially pre-processed using the histogram equalization technique. Then it is passed to feature extraction technique where GLCM comes in hand for extracting required features and finally passed to quintessential process which is the classification where DenseNet169 + CNN comes in play for effective classification. We have evaluated our model under accuracy, sensitivity, specificity performance measure and is compared with other pre-trained models like VGG16, ResNet50, GoogleNet, MobileNet and Inception V3 in which our model outperforms other state-of-art models with 98% of accuracy. Keywords: Age-related macular degeneration · Ant colony optimisation · Classification · Convolutional neural network

1 Introduction In Western countries, age-related macular degeneration (AMD) is the primary reason for vision loss and irreversible blindness of the aged [1]. It encompasses a wide range of macula problems. Although the initial stage of AMD is silent, examination of the retina can reveal tiny lesions known as drusen. The appearance of hemorrhages (wet AMD) or the formation of regional atrophy are both signs of progression of the disease, as is a growth in the size or quantity of drusen (late dry AMD). A Study [2] developed a clinical categorization for AMD. It is divided into four categories: non-AMD, mild, moderate, © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 226–238, 2022. https://doi.org/10.1007/978-3-030-96299-9_22

Detection and Classification of Age-Related Macular Degeneration Using Integration

227

Fig. 1. a) Vision of normal people. b) Vision of person with AMD

and progressive AMD [3]. Figure 1 (a) depicts the vision of normal people and 1(b), the vision of a person with AMD. SD-OCT (Spectral-Domain Optical Coherence Tomography) is widely used equipment that explains specific AMD findings such as drusen, intra-retinal fluid (IRF), subretinal fluid (SRF), sub-retinal hyper-reflective substance, which includes hemorrhage and retinal colourant epithelium detachment, among others [12]. Exudative modifications (intra- and sub-retinal fluid and hemorrhage) are the most common reasons for many doctors to start anti-VEGF medication and monitor its effectiveness [13]. For the big clinical experiment discussed earlier [14], zero tolerance was used. As a result, the volume of OCT data that needs to be analyzed is growing faster than clinical capacity [15]. Machine-learning advances have found an answer for interpreting vast volumes of medical picture data resulting from repeated patient treatment and follow-up observations [16]. Convolutional neural networks (CNNs) have made tremendous progress in the processing of clinical images [18]. CNNs have already been used in ophthalmology for the categorization of diabetic retinopathy on fundus images, visual field evaluation of patients, and assessment of paediatric nuclear cataracts, among other applications [19]. Computerized detection of AMD characteristics in OCT and fundus pictures, anti-VEGF drug counselling, and disease progression tracking have all been done with neural networks [19]. Although several researchers have categorized the exudative element of AMD for computerized segmentation [21], we are unaware of any publications that categorize exudative alterations with AMD in deep learning models excluding segmentation. This paper focuses on an accurate and efficient model for detection and classification of AMD in which key points are as mentioned below: 1. 2. 3. 4.

The detection of AMD by Deep Learning Datasets used are AREDS, Optretina and STARE. DenseNet169 + CNN is used for Classification. An accuracy of 98% is achieved.

228

F. Ajesh and A. Abraham

The remaining sections organized as follows: Sect. 2 depicts the works related to this paper that is been so far done by research while Sect. 3 illustrates the methodologies of the proposed model. Section 4 is focused on the simulation analysis of proposed model and finally the paper concludes with Sect. 5.

2 Literature Review Using the Optretina database and CNN approach, Zapta et al. [17] used AI to detect retinal fundus images, quality verification, laterality analysis, macular degeneration, and potential glaucoma. They had a 94.7% accuracy rate. A Convolutional Neural Network is much slower due to procedures such as a max pool [4]. The training technique takes a long time if the computer does not have a powerful GPU and a ConvNet requires a huge database to analyze and train the neural network. Motozawa et al. [26] used the STARE database and the CNN approach to develop OCT-based DL (Deep Learning) models for distinguishing normal and AMD, as well as exudative and non-exudative AMD alterations. They had a 96% accuracy rate. Computational power (depends on network design and data amount) and Sophisticated architecture are two drawbacks of their job (Not every time). With the database AREDS and CNN, Keenan et al. [6] developed a DL system for automatic detection of geographic atrophy from color fundus pictures. They achieved an accuracy of 92.7%. The disadvantage of their work is that it does not work well with high dimensions and requires more computational resources. Grassman [20] used the CORRA database and the RNN approach to construct a DL system to predict Age-Related Eye Disease (AREDS) research intensity scale for AMD using color fundus photography. They had an accuracy percentage of 92.7%. The drawbacks of their work include the fact that neural networks are black boxes, meaning we don’t know whether each independent variable influences the dependent variables. Traditional CPU training is both technologically incompetent and time-intensive, and neural network training data is crucial [23]. As a result, over-fitting and generalization become a worry. The method is more dependent on training data and can be customized to meet your specific requirements.

3 Methodology Here the author tries to propose a new methodology for the detection and classification of AMD using a hybrid structure of Densenet 169 and CNN [5]. Figure 2 depicts the diagram of the methodology of the proposed model. Here, initially, we collect datasets from repositories like AREDS, Optretina and STARE which is followed by proposing this dataset as it contains noises and other anomalies. For that, we use histogram equalization processing technique. Further, extracting features using GLCM is done. Finally, passes to the main stage which is classification [25]. Here we use the DENSENET169 neural network for accurately classifying images into disease/no-disease or dry/wet ARMD.

Detection and Classification of Age-Related Macular Degeneration Using Integration

229

Fig. 2. Methodology of the proposed model

3.1 Dataset The repositories used are AREDS, Optretina and STARE. • The AREDS is a multi-centre, panel study of AMD and cataracts. The AREDS study included a clinical investigation of large quantity mineral and vitamin supplements for AMD [7] and a clinical investigation of large quantity vitamin supplements for cataracts, as well as natural history data. Participants in the AREDS study had to be between the ages of 55 and 80 at the time of enrollment, and they had to be in healthy conditions which would otherwise make a longitudinal observation or drug administration difficult or impossible. 4,757 participants were divided into four AMD categories based on fundus images assessed by a centralised reading centre, finest visual acuity, and ophthalmologic examinations. • Optretina has been doing telemedicine screenings at optical centres since 2013, and in workplace offices and private firms since 2017. The Optretina tagged contents, which contain 306,302 retinal pictures, were used in this study. NMC (colour fundus and red-free) and optical coherence tomography are used to create the images (OCT) [8]. Images from various types and brands of cameras were included in the dataset. • The STARE dataset (Structured Analysis of the Retina) is a retinal vascular segmentation dataset. It includes 20 colour fundus photos that are all the same size (700605).

230

F. Ajesh and A. Abraham

3.2 Pre-processing Once the data is collected, preprocessing should be done to remove any noises or anomalies. For the cleansing purpose, we used histogram equalization. Histogram equalization is a contrast correction approach in image processing that uses the image’s histogram. For that, we need to convert the given image into a greyscale version. In greyscale conversion, we normalize the non-uniformities and then enhances the contrast of the image. Then this converted image is given for histogram equalization process where it enhances image based on the intensity values and thereby receiving a pre-processed output which will then be passed on for extraction process [9]. 3.3 Feature Extraction In the proposed model, the GLCM technique is used to derive information from the image. The proportion of co-occurring values at a given interval is represented by a grey level co-occurrence matrix (GLCM), which is a matrix formed over an image [22]. In a GLCM matrix, the quantity of grey levels, G, in an image equals the number of rows and columns. As a result, using statistical characteristics is one of the earliest methodologies proposed in the image analysis literature. Haralick [14] recommended using a co-occurrence matrix. It evaluates the connection between the two adjacent pixels, the first of which is referred to as a reference pixel and the second as a neighbour pixel. It’s also known as a grey tone spatial dependence matrix [10]. It is used to identify the texture of an image by tabulating the intensity values of an image and the frequency with which they appear. The five characteristics we extracted are homogeneity, correlation, co-occurrence, energy, and entropy. There are two-pixel values included in GLCM, standard pixel and neighbor pixel. The neighbor pixel is picked as the pixel to the right of each reference pixel. The GLCM matrix is a square matrix with Ni as the no: of grey levels. To get element [j,k], divide the total number of such comparisons by the number of times the pixel value of j is next to the pixel value of k. The features extracted from the GLCM are contrast, correlation, energy, homogeneity and entropy. Table 1 depicts the corresponding equations of the features extracted. 3.4 Feature Selection and Feature Optimization Feature selection is a technique for choosing a collection of extracted features or creating factors with the greatest classification results. This method prevents model overfitting by removing potentially irrelevant or misleading information. In other terms, it identifies key characteristics which can be utilized to distinguish healthy from unhealthy images. Ant colony optimization (ACO) is a successful approach that is useful in the solution of NP-hard combination optimization problems and it has been widely applied in GWAS. The basic idea of ACO is to express the feasible solutions of optimization problems with ant paths and use overall paths of the ant group to constitute the solution space of optimization problems. Ants on relatively short paths tend to release more pheromone. With the passage of time, pheromone concentration that accumulates on the short paths gradually increases and more and more ants choose the paths. Eventually, all the ants

Detection and Classification of Age-Related Macular Degeneration Using Integration

231

Table 1. Statistical features of the features used in feature extraction Extracted features

Equation

Mean

 m=1   n=1  1 M = m×n f (x, y)

Standard deviation

   m=1   n=1  SD(σ ) =  1 (f (x, y) − M )2

x=0 y=0

Entropy

m×n

E=−

m=1  n=1 

x=0 y=0

f (x, y)2 f (x, y)2

x=0 y=0

Contrast

  (f (x,y)−M )3 1 Sk (X ) = mn SD3   (f (x,y)−M )4 1 Kurt (x) = mn SD4 m−1 n−1 Con = k=0 y=0 (x − y)2 f (x + y)

Correlation

Corr =

Coarseness

1 Cness = 2m+n

Skewness Kurtosis

m−1 n−1 x=0

y=0

(x,y)f (x,y)−Mx My σx σy

m−1 n−1 k=0

y=0 f (x, y)

will gather on the optimal path under positive feedback, which exactly corresponds to the optimal solution of the optimization problem. In this study, a genetic optimization strategy formed using crossover and mutation operators in the genetic algorithm was combined with an ant colony optimization shortest-path method to achieve feature selection. Holland created the Genetic Algorithm (GA), which is a method of addressing an optimization issue that operates similarly to biological evolution. In a genetic algorithm, a starting population of solutions (similar to chromosomes) is selected and subjected to iterative change. Every member in the current population has a fitness value assigned to them. The fitness value is determined by training the prediction model with the training data set and then calculating the selection error. The lower the value of selection error, the lower the fitness. As a result, those with a higher fitness value will be chosen to create the next population. For a predetermined number of generations, the algorithm continues until the best answer throughout the evolution process does not shift to a better value. This predefined value can be 20% or 30% of the generation number determined by the best solution so far. For example, at generation 50, the algorithm achieves a value of 200, which does not change for 15 generations (30% of 50), causing the process to stop. Our primary objective was to distinguish between dry and moist ARMD [11], as well as ARMD without symptoms. To obtain a P-value, variance analysis was used to choose the top-ranking energy and entropy parameters. For categorization, the top ten statistically significant (P 0.05) variables (1 energy, 3 entropy, 6 other nonlinear) were chosen. The optimization is

232

F. Ajesh and A. Abraham

Fig. 3. Flowchart of ant colony optimization

done by Ant Colony Optimization. The flowchart of Ant Colony Optimization is depicted in Fig. 3. 3.5 Classification This is the most crucial process of the entire system in which we use several ML or DL techniques to predict the desired output. So, in this paper, for diagnosing the AMD anddisease/no-disease or dry/wet ARMD, we use DenseNet169 + CNN. It is a Fully Convolutional Neural Network (FCN) that is a simultaneously trained learning network with convolutional filters in place of fully linked layers as judgement layers. By linking the output neurons of wholly attached levels to all input neurons, this alteration on the top level’s aids in the reduction of data connected to a place caused by fully linked areas. DenseNet169 + CNN’s defining feature is its ability to reuse features extracted and boost characteristic dispersion by establishing a straight link among each layer and every subsequent layer. Dense, transition down, and transition up are the three essential blocks of the DenseNet169 + CNN network. A batch normalization step follows a ReLU as an input signal, a 3 × 3 convolution layer, and a dropout layer with a 0.2 decreasing rate in the Dense block (DB). A batch normalization layer, a ReLU as an input signal, a 3 × 3 convolution layer, a dropout layer with a 0.2 dropping rate, and a 2 × 2 Max pooling layer make up a transition down (TD) block. Three transposed convolution layers make up a transition Up (TU) block. It’s worth noting that batch normalization and dropout

Detection and Classification of Age-Related Macular Degeneration Using Integration

233

can both help to reduce overfitting. The network may need one, both, or none of these, based on the extent of overfitting. We discovered that integrating both dropout and batch normalization in our network improves performance on this issue. The network’s architecture includes a 3 × 3 convolution layer on the input, five dense blocks with 4, 5, 7, 10, and 12 layers each, a transition down component, one dense block with 15 layers in the final layer of the down-sampling path (bottleneck), five transitions up blocks with dense blocks of 12, 10, 7, 5, and 4 layers, and a 1 × 1 convoluted layer followed by a non-linearity indicated by the Softmax function. RMSprop [24], a stochastic gradient descent optimization technique, is utilized for training the network with a rate of learning of 10–3 in 120 epochs and a 30 epoch early-stop condition. The images are enhanced with vertical flips and irregular cropping to artificially increase the number of images. Uniform distribution was used to initialize the network’s weights, and the loss function was cross-entropy. Testing of stages can begin once the model has been developed by utilizing the trained model to section the images in the test set. Figure 4 depicts the architecture of DenseNet169 + CNN.

Fig. 4. Architecture of DenseNet169 + CNN

4 Experimental Results The proposed DenseNet169 + CNN model is compared with the existing cut-edge models such as VGG16, ResNet50, GoogleNet, MobileNet and Inception V3 and their performance is measured using several measures such as Accuracy, Specificity, Sensitivity, Precision, F-measure and confusion matrix. .

234

F. Ajesh and A. Abraham Table 2. Metrics of performance measures

Metrics of performance measure

Mathematical description

1. Sensitivity, TPR

TP = TP+FN

2. Specificity, S

TN = FP+TN

3. Precision

TP = TP+FP

4. Accuracy

TP+TN = TP+FN +TP+TN

5. F Score

2TP = 2TP+FN +FP

Here, True positive (TP): You predicted a positive and it turned out to be correct; True Negative (TN): The negative you predicted is correct; False Positive (FP): You predicted something positive but it wasn’t true, and False Negative (FN): You predicted something negative but it wasn’t true. Table 2 shows the metrics of performance measures. Table 3 illustrates the evaluation results of DenseNet169 + CNN with other models in which CNN gives better results. Figure 5 gives the graphical representation of the confusion matrix. Table 3. Confusion matrix Classification models

TP

FN

FP

TN

VGG16

142

20

12

59

GoogleNet

136

28

21

43

InceptionV3

120

38

38

30

ResNet50

132

29

27

41

MobileNet

149

7

5

70

DenseNet169 + CNN (Proposed)

160

1

1

30

Table 4 represents the comparison of the proposed model with other models based on performance for images 1, 2 and 3 respectively. Figures 6, 7 and 8 depict a graph of overall performance analysis of images 1, 2 and 3. Table 5 depicts comparison of proposed model with state-of-art works.

Detection and Classification of Age-Related Macular Degeneration Using Integration

235

Fig. 5. Graphical representation of confusion matrix

Table 4. Performance analysis of images 1,2 and 3. Metrics

VGG16 GoogleNet inceptionV3 ResNet50 MobileNet Densenet169 + CNN (Proposed)

Image Accuracy 91 1 Sensitivity 91.78

75.1

69

82

95

98

81.59

78.08

86.6

95.77

98.5

Specificity 88.8

60

44.44

68

93.33

96.5

Precision

82.6

79.17

89.04

97.14

98.5

82.09

78.62

87.84

96.45

98

79

64

74

93

98

89

78.01

88.11

93.42

97.95

94.7

F-measure 93.71 Image Accuracy 90 2 Sensitivity 84.1 Specificity 88

64.21

54.24

69.21

94.11

96.5

Precision

92.12

74.54

89.20

88.24

98

98.5

F-measure 93.71

82.09

78.62

87.84

96.45

98

Image Accuracy 89 3 Sensitivity 94.21

76.21

72

85

96.02

98

78.59

67.08

87.56

93.70

98.15

Specificity 92.12

64

55.54

69.12

89.12

96

Precision

84.17

72.46

69.19

90.04

96.14

98

F-measure 84.71

79.09

81

86.94

96

97.95

236

F. Ajesh and A. Abraham

Fig. 6. Overall performance analysis of image 1

Fig. 7. Overall performance analysis of image 2

Fig. 8. Overall performance analysis of image 3

Detection and Classification of Age-Related Macular Degeneration Using Integration

237

Table 5. Comparison of proposed model with state of art works Author

Database

Method

Accuracy (%)

Zepta et al. [17]

Optretina

AI

94.7

Motozoa et al. [26]

Stare

CNN

96

Keenan et al. [6]

Areds

DL

92.7

Grassman et al. [20]

Areds

DL

92.7

Proposed model

Areds, Optretina and Stare

DenseNet169 + CNN

98

5 Conclusions This research introduces a new DenseNet169 + CNN model for AMD detection and classification. We have suggested a new methodology with a few steps, which makes the model more cut edge one among the present ones. This methodology used the DenseNet169 + CNN more specific and sensitive compared to other models. Here, we used DenseNet169 + CNN for better evaluation and it clearly classified the given image into disease/non-disease or wet/dry AMD. This model is compared with other present models and we obtained a better result of 98% accuracy. Also, this paper is much useful for other researchers in helping them to bring other hybrid models for even better evaluation of AMD. Acknowledgement. This research has been financially supported by The Analytical Center for the Government of the Russian Federation (Agreement No. 70–2021-00143 dd. 01.11.2021, IGK 000000D730321P5Q0002).

References 1. Stark, K., Olden, M., Brandl, C.: The german AugUR study: study protocol of a prospective study to investigate chronic diseases in the elderly. BMC Geriatr (2015) 2. Prenner, J.L., Halperin, L.S., Rycroft, C., Hogue, S., Williams Liu, Z., Seibert, R.: Disease burden in the treatment of age-related macular degeneration: findings from a time-and-motion study. Am. J. Ophthalmol. 160(4), 725–731e1 (2015) 3. National Eye Institute: Facts about age-related macular degeneration. Available: https://nei. nih.gov/health/maculardegen/armd_facts (2015). Accessed 28 Jul 2017 4. Koh, J.E.W., Ng, E.Y.K., Bhandary, S.V., Laude, A., Acharya, U.R.: Automated detection of retinal health using PHOG and SURF features extracted from fundus images. Appl. Intell. 48(5), 1379–1393 (2017). https://doi.org/10.1007/s10489-017-1048-3 5. Steinberg, J., Uibel, S., Berndt, T., Müller, D., Quarcoo, D., Groneberg, D.A.: Zentralblatt für Arbeitsmedizin, Arbeitsschutz und Ergonomie 61(8), 270–286 (2011). https://doi.org/10. 1007/BF03345002 6. Keenan, T.D., et al.: Progression of Geo- ´ graphic atrophy in age-related macular degeneration: AREDS2. Ophthalmology 125, 1913–1928 (2018) 7. Tan, J.H., Acharya, U.R., Bhandary, S.V., Chua, K.C., Sivaprasad, S.: Segmentation of optic disc, fovea and retinal vasculature using a single convolutional neural network. J. Comput. Sci. 20, 70–79 (2017)

238

F. Ajesh and A. Abraham

8. He, K., Zhang, X., Ren, S., Sun, J.: Delving deep into rectifiers: surpassing human-level performance on imagenet classification. In: Proceedings of the IEEE International Conference on Computer Vision, vol. 11–18–Dece, pp. 1026–1034 (2016) 9. Köse, C., Sevik, ¸ U., Gençalioˇglu, O.: Automatic segmentation of age-related macular degeneration in retinal fundus images. Comput. Biol. Med. 38(5), 611–619 (2008) 10. Köse, C., Sevik, ¸ U., Gençalio˘glu, O., ˙Ikiba¸s, C., Kayıkıçıo˘glu, T.: A statistical segmentation method for measuring age-related macular degeneration in retinal fundus images. J. Med. Syst. 34(1), 1–13 (2010) 11. Ferris, F.L.: Clinical classification of age-related macular degeneration. Ophthalmology 120(4), 844–851 (2013) 12. Kuhn, M.: Building predictive models in R using the caret package. J Stat Softw. (2008) 13. Aiello, S., Eckstrand, E., Fu, A.: Fast scalable R with H20. In: Grün, B., et al. (eds.) Foundation for Open Access Statistics. ISSN (2015) 14. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet Classification with Deep Convolutional Neural Networks. Curran Associates Inc (2012) 15. Szegedy, C., Liu, W., Yangqing, J.: Going deeper with convolutions. In: 2015 IEEE Conf. Comput. Vis. Pattern Recognit. IEEE, Computer Society Conference on Computer Vision and Pattern Recognition. IEEE Computer Society Press, ISSN: 1063–6919 16. Gulshan, V., Peng, L., Coram, M.: Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA (2016) 17. Zapata, M.A., Royo-Fibla, D., Font, O.: Artificial intelligence to identify retinal fundus images, quality validation, laterality evaluation, macular degeneration, and suspected glaucoma. Clin Ophthalmol. 14, 419–429 (2020) 18. Szegedy, C., Vanhoucke, V., Ioffe, S.: Rethinking the Inception Architecture for Computer Vision. Computing Research Repository (CoRR). abs/1512.0. Available at: https://arxiv.org/ corr/home (2015) 19. Fritsche, L.G., Igl, W., Bailey, J.N.C.: A large genome-wide association study of age-related macular degeneration highlights contributions of rare and common variants. Nat. Genet. (2016) 20. Grassmann, F., Mengelkamp, J., Brandl, C.: A deep learning algorithm for prediction of age-related eye disease study severity scale for age-related macular degeneration from color fundus photography. Ophthalmology 125, 1410–1420 (2018) 21. Grassmann, F., Ach, T., Brandl, C.: What does genetics tell us about age-related macular degeneration?. Annu. Rev. Vis. Sci. (2015) 22. Swartz, R., Loewenstein, A.: Early detection of age-related macular degeneration, Int. J. Retin. Vitr. Article Number 20, December 01 2015 23. Swaroop, A., Branham, K.E., Chen, W., Abecasis, G.: Genetic susceptibility to age-related macular degeneration: a paradigm for dissecting complex disease traits. Hum. Mol. Genet. (2007) 24. Holz, F.G., Bindewald-Wittich, A., Fleckenstein, M.: Progression of geographic atrophy and impact of fundus autofluorescence patterns in age-related macular degeneration. Am. J. Ophthalmol. (2007) 25. Ratnapriya, R., Chew, E.Y.: Age-related macular degeneration-clinical review and genetics update. Clin. Genet. 84, 160–166 (2013) 26. Motozawa, N., et al.: Optical coherence tomography-based deep-learning models for classifying normal and age-related macular degeneration and exudative and non-exudative age-related macular degeneration changes. Ophthalmol Therapy 8(4), 527–539 (2019). https://doi.org/ 10.1007/s40123-019-00207-y

An Augmented Lagrangian Artificial Bee Colony with Deterministic Variable Selection for Constrained Optimization Marco Antˆ onio Florenzano Mollinetti1(B) , Bernardo Bentes Gatto2 , and Ot´ avio Noura Teixeira3 1

2

Plimes Inc., Tsukuba, Ibaraki, Japan [email protected] National Institute of Advanced Industrial Science and Technology (AIST), Tokyo, Ibaraki, Japan [email protected] 3 Natural Computing Laboratory (NCL), UFPA, Tucuru´ı, PA, Brazil [email protected]

Abstract. Nonlinear constrained optimization problems with nonlinear constraints are common in real-life application models. A viable option to handle such problems is metaheuristics that use proper penalty methods to bound solutions to the feasible space delimited by the constraints. Most penalty methods not only hinder the diversity of solutions but fail to exploit the feasible boundary of constraints from within the infeasible region. In light of this, we propose two methods to be incorporated into derivative-free algorithms for constrained optimization: a deterministic decision variable procedure based on previous works on multimodality; and a penalty method based on the augmented Lagrangian. We limit the study of the effects of our approach to the use of the Artificial Bee Colony algorithm (ABC) and several of its variants due to its simplicity and modular implementation. We validate our hypothesis by means of a numerical experiment using seven distinct nonlinear constrained optimization instances comparing the canonical ABC and some variants made for constrained optimization against their counterparts with the proposed deterministic selection and penalty method. Results suggest a positive outcome in relation to the integration of both methods to the ABC, opening up new avenues of possibilities for our proposed methods to be incorporated into other derivative-free algorithms.

Keywords: Constrained optimization Bee Colony

1

· Penalty methods · Artificial

Introduction

Problems featuring equality and inequality constraints better approximate reallife scenarios; moreover, instances of such where the constraints are nonlinear c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 239–250, 2022. https://doi.org/10.1007/978-3-030-96299-9_23

240

M. A. F. Mollinetti et al.

are very difficult to derive a differentiable approximation [19]. Derivative-free algorithms, such as metaheuristics, are appropriate techniques to tackle these problems firsthand because they do not require derivatives to guide their solutions. However, they are ill-equipped for constrained optimization without the inclusion of any form of penalty function. Simple barrier methods to restrict the solution set within the bounds of the feasible space may be seen as a simple and straightforward quick fix. However, they are unable to take advantage of any property of the constraints or the infeasible space to improve the quality of solutions, possibly leading the solutions to local optima to the feasible side of the boundary of the search space, thus reducing the diversity of solutions. From the assertions above, we argue that the development of metaheuristics that are able to maintain a healthy balance of exploitation and exploration while giving certain freedom for solutions to be in the infeasible region would result in algorithms that not only could converge faster to promising local attractors, but also maintain a degree of exploration to keep solutions from stagnating at unpromising local optima. In this paper, our proposal is twofold. First, an augmented Lagrangian penalty method for derivative-free algorithms that use solution sets, along with a penalty barrier method is employed to handle equality and inequality constraints, where the Lagrange multipliers and barrier multipliers are updated at every iteration. Second, a revision to the Adaptive Decision Variable Matrix (A-DVM) method from Mollinetti et al. [16] for constrained optimization problems together with specific rules to the ABC algorithms to choose decision variables according to the degree of satisfiability of the constraints. A numerical experiment is carried out to validate our proposition. We choose seven nonlinear optimization problems from the field of engineering design due to their difficulty and popularity in the literature. In this first moment, we limit the scope of our experimentation to test the effects of our approach solely to Artificial Bee Colony (ABC) algorithms from Karaboga [12] due to the modular nature of the algorithm, where effects can be easily seen and explained. The canonical ABC and several ABC variants suited to constrained optimization problems are tested against their counterparts with the A-DVM and the Lagrangian penalty method integrated into their implementation.

2

Artificial Bee Colony

Artificial Bee Colony (ABC) is a Swarm Intelligence (SI) heuristic developed by Karaboga [12] based on the mathematical model of the foraging and information sharing behavior of honey bees. ABC has a simple design and is easy to use without major modifications for a variety of optimization families problems, e.g., continuous, combinatorial and mixed-integer [1,3]. The canonical ABC comprises four main steps: initialization, employed bees, onlooker bees, and scout bees. In summary, an initial solution set X is initialized based on specific rules; then solutions are sampled and updated by local and global search procedures iteratively until a stopping criterion is met. ABC has three tunable parameters: the solution set size SN ; the maximum number of

Augmented Lagrangian ABC-ADVM for Constrained Optimization

241

iterations M CN ; and the solution stagnation threshold Lit. Assuming that in the Initialization step, an initial solution set X = {x 1 , x 2 , . . . , x SN } is generated from an uniform distribution within the feasible bound, a counter li = 0 to indicate unsuccessful updates is initialized for each x ∈ X. A brief description of the remainder of the steps is given below. 2.1

Employed Bees Cycle

For each x ∈ X, a randomly chosen component xij of each solution xi ∈ X is shifted by a random step size towards the jth component of some xk ∈ X, k = i:  i xj + φ (xij − xkj ) if q = j,  xiq = (1) xiq , otherwise, where φ ∈ U(−1, 1). To verify if the update step was successful, the value of f is evaluated and a greedy selection is done for each i = 1, 2 . . . , SN :  i   x + (xij − xij )ej , li = 0 if f (xi + (xij − xij )ej ) ≤ f (xi ) x = xi , li = li + 1 otherwise, i

(2)

where ej is the jth fundamental vector. If (1) fails, (2) flags (1) as a failed update and add one to the counter li . 2.2

Onlooker Bees Phase

A solution xi ∈ X is chosen with probability pi according to a weighted roulette selection scheme to be updated using (1). This step can be thought of enhancing local search for solutions with better objective function values. Probability pi is determined for each solution xi ∈ X and F (·) is the adjusted objective function value as follows,

F (xi )

pi = SN

i=1

2.3

F (xi )

 ,

(3)

i

F (x ) =

1

i

1+f (x ) i  1 + f (x )

if f (xi ) ≥ 0 otherwise.

(4)

Scout Bees Phase

˜ = {xk ∈ X | lk ≥ Lit} denote the set of solutions flagged as stagnated. A Let X ˜ = ∅. This prevents new point in Rd is resampled by an uniform distribution if X the algorithm from premature convergence to bad local optima and increases the number of explorations. The parameter Lit is commonly defined as SN · d. ˜ is chosen to be resampled. ˜ ≥ 2, then xw ∈ argmax{f(x) | x ∈ X} If |X|

242

3

M. A. F. Mollinetti et al.

Adaptive Variable Decision Matrix Procedure

The Adaptive Decision Variable Matrix (A-DVM) was developed by Mollinetti et al. [16] as an extension of the decision variable selection procedure of Mollinetti et al. [17] made for improving the quality of derivative-free algorithms for multimodal nonlinear optimization problems. The A-DVM is made to be integrated to any metaheuristic without interfering to any subroutines. In the case of the ABC, it can be integrated to the employed and/or onlooker bees phase without interfering with any additional steps of any variant of the original algorithm. Mainly, the A-DVM builds a d × SN trinary matrix Pam that represents component xij has been chosen to be updated by an update rule (in the case of the ABC, (1)). The matrix Pam is a composition of two matrices: Pr , a binary matrix {0, 1}, whose row is determined randomly according to a uniform distribution; and Pd , a binary matrix {0, 2} generated by the fully deterministic scheme of Mollinetti et al. [17]. Pam is the result of a composition of Pr into Pd . Some solutions xi ∈ X have their j-th component randomly selected when updated by (1) while the rest have their j-th component chosen by the fully deterministic scheme. We write Pam = βPr ⊕ αPd when β% of the columns of Pam are from Pr and the remaining α(= 1 − β)% are from Pd . An example of Pam is as follows, ⎞ 0 0 0 ... 0 0 ... 0 ⎜0 2 0 . . . 0 0 . . . 0⎟ ⎟ ⎜ ⎟ ⎜ = βPr ⊕ αPd = ⎜1 0 2 . . . 0 1 . . . 0⎟ . ⎟ ⎜ .. .. .. . . . ⎝ . . . . 0 .. . . . 0⎠ ⎛

Pam

0 0 0 ... 2 0 ... 2 To maintain a healthy diversity of solutions while balancing between local search and local search, α and β are automatically adjusted along the iterations as follows, α = (1 − Δ)K1 + ΔK2 ,

β = 1 − α,

Δ = Δ1 + Δ2 =

1.75 − S1 + S2 , (5) 1.75

where Δ ∈ [0, 1] is the measure of the dispersion of the population at the current iteration and K1 and K2 are scaling parameters set to 0.3 and 0.7 as suggested by McGinley et al. [15]. Values of α close to 1 signify high population diversity and activate exploitation by the deterministic selection, while values close to 0 boost exploration using random selection. Calculation of S1 and S2 is as follows, S1 = max

    IUo − Ij + P crj2 

, P ⎡   (1−cr )    (cr )  ⎤     j j j j  P   P χc− − P   P χc+ − j φj j φj ⎥ ⎢ S2 = max ⎣ , ⎦. P P j

(6)

Augmented Lagrangian ABC-ADVM for Constrained Optimization

243

where c− = {xij ∈ X|xij ≤ crj }, c+ = {xij ∈ X|xij ≥ crj }, χ is the characteristic function that returns either 0 or 1 whether a solution belongs d to c− or c+, ¯j ], so that j=1 φj = 1 for an respectively, and φj is the range between [xj , x N -dimensional unit volume. The centroid crj of the jth components and the moment of inertia Ij of centroid crj and the inertia of an uniform distribution IUo are denoted as: P crj =

i=1

P

xij

,

P  (xij − crj )2 , Ij = i=1

I Uo =

P   i=1

i P +1

2

.

(7)

Because solutions in population-based algorithms tend to concentrate around accumulation points in the later stages of execution [14], α is biased upwards by a growth function ρ = αeγt to intensify local search around x after t iterations, where γ is set to 0.01 and an acceptable value for t can be calculated as: t = min (n·d/λt ·tmax , λt tmax ), where λt = 0.1. History H ∈ {0, 1}d stores which columns of Pd were put into Pam , and give a chance to the remaining columns of Pd to be contained in Pam at the next iteration. A bound on the number of iterations that solutions are chosen by the fully deterministic selection is enforced to be no more than 3/5K1 and no less than 1/2K2 (refer to (5)). When the entries in H are all ones, H is reinitialized and the whole process runs again. The safeguard step from the version of Mollinetti et al. [17] is kept where the last column xSN of X is moved to the first column and the remaining columns are shifted to the right. This ensures that the jth components of all solutions will be chosen in at least SN · j iterations.

4

Augmented Lagrangian Penalty Method

The more constraints a problem has, the more stringent is the feasible region, linear or nonlinear [5]. Assuming f is nonlinear and the search space is C 0 (nondifferentiable and non Lipschitz continuous), a nonlinear optimization problem with constraints of this nature is denoted as, minimize n x∈R

subject to

f (x) gi (x) ≤ 0,

i = 1, . . . , p

hi (x) = 0, ¯j , xj ≤ xj ≤ x

i = 1, . . . , m j = 1, . . . , n,

(8)

where f : Rn → R, g : Rn → Rp , h : Rn → Rm . Metaheuristics such as the ABC, as they are, cannot handle problems like (8) without the inclusion of penalty functions to account for infeasible solutions or solutions that are outside the search space or violate any constraint [20]. Penalty methods such as enforcing the decision variables to stay within their feasible bounds or adding the values of violated constraints are straightforward methods for metaheuristics. However, we argue against the use of simple methods such as the aforementioned by citing many authors on global optimization [19,20]: “Local optima are generally found

244

M. A. F. Mollinetti et al.

from solutions that started from the infeasible region which iteratively moved to the feasible region”. A penalty method that exploits this property is the augmented lagrangian with dislocated penalties [7,20] which changes (8) to an unconstrained problem that punishes infeasible solutions according to weights that are updated iteratively. We extend this concept by reformulating (8) to account for solutions out of the search space, so that it can properly penalize solutions that satisfies constraints g(x) and h(x) but are not within bounds. For any iteration k, the formulation is as follows, minimize n x ∈R

f (x) +

  2   ¯ k 2  ¯k  ρk   h(x) + λ  + g(x) + μ + z(x)2 .  2  ρk  ρk +

(9)

⎧ ⎪ ⎨xi + xi if xi < xi , z(x) = χ(β k ) + β k z(xi ) = xi − x ¯i if xi > x ¯i , for i = 1, . . . , n. ⎪ ⎩ i=1 0, otherwise,  n β k if k i=1 z(xi ) > 0, χ(β ) = 0, otherwise, (10) ¯ and μ ¯ are the equalities and inequalities Where ρk is the Lagrange multiplier, λ multipliers, respectively. Penalty function z(x) relies on a weight factor β k = ¯ [βmin , βmax ] that acts as a multiplier like λk and μk . The multipliers ρk , λ ¯ start from an initial value and are updated each iteration so that the and μ penalties may be increased or decreased according to the feasibility of solutions. The update step is as follows, firstly, compute vector V k of the inequalities dislocation,   ¯ ki μ k (11) Vi = max gi (xk ), − k , i = 1, . . . , p. ρ           ≤ τ max h(xk−1 )∞ , V k−1 ∞ , set If max h(xk )∞ , V k ∞ n

ρk+1 = ρk , otherwise set ρk+1 = γρk . It is advisable to set rules for the increase of multiplier ρk , either establishing a limit ρmax , or resetting ρk to a more acceptable value once it reaches a threshold value. We opted for the last in our ¯ k and μ ¯ k are computed, formulation. After updating ρk , suitable values for λ   ¯ k+1 = min λmax , max{λmin , λ ¯ k + ρk hi (xk )} , λ i i   μ ¯k+1 = min μmax , max{0, μ ¯ki + ρk gi (xk )} , i

i = 1, . . . , m (12) i = 1, . . . , p

where λmin , λmax and μmin are user defined parameters sensitive to each problem instance. Because the Lagrangian multipliers are more suited to methods that use only one incumbent solution at a time, multipliers ρ, μ, λ and β are reset to

Augmented Lagrangian ABC-ADVM for Constrained Optimization

245

k their starting values if the solution set had  theirvalues restarted.  Multiplier β is updated similarly to ρk : the condition z(xk )∞ ≤ z(xk−1 )∞ is verified, if it is satisfied then β k+1 = βk , otherwise β k+1 = γ b βk , γ b > 1.

5

Adapted A-DVM Using Augmented Lagrangian Penalty for Constrained Optimization Problems

The A-DVM was shown to be successful in improving the overall performance of ABC algorithms for multimodal optimization problems [16]. We propose an extension of the A-DVM, along with the incorporation of the Augmented Lagrangian penalty method (9) proposed in Sect. 4 for ABC algorithms to handle constrained nonlinear optimization problems in the form of (8). Additional deterministic rules are added to the A-DVM to take full advantage of the properties of constrained nonlinear problems. The rules are applied to the construction of deterministic matrix Pd and take precedence over the standard procedure of selecting diagonals. They are as follows, Rule 1: If a constraint gj or hj depends on only one component xj of x and xj violates gj , then xj is chosen to be updated. Otherwise, if gj or hj depend on multiple variables and is the only constraint that was not satisfied, a variable xj  associated to gj is chosen. Rule 2: In order to intensify the local search in the vicinity of the feasible bounds, a component xj of x whose 2 distance is less than a predetermined ¯j is chosen. is seen as the radius of an euclidean ball value to either xj or x B = {y ∈ Rn : y − x 2 ≤ }, centered at x. Calculation of Pam follows the same steps as in Sect. 3. The step-by-step of the A-DVM for constrained problems is detailed in Algorithm 1.

6

Experiment and Results

A numerical experiment is conducted using 7 engineering design problems to assess the robustness and performance of the proposed approach to the algorithms. Since there are no benchmarks available for this domain of problems, we chose the most prominent literature. This experiment intends to answer the following research question: “With the integration of the augmented Lagrangian and the A-DVM, given sufficient adjustments, can they improve a search algorithm such as the ABC to a competitive degree, specifically for nonlinear constrained optimization problems with nonlinear constraints?”. The original ABC from Karaboga [13] together with several variants designed for constrained optimization are compared against themselves with the A-DVM from Sect. 5 and the augmented Lagrangian from Sect. 4 as penalty method. It is important to state that we do not conduct a comparison against results found in the literature due to the discrepancy of the formulation of the engineering problems in many of the works. Chosen variants of comparison were based on the

246

M. A. F. Mollinetti et al.

Algorithm 1: Adapted A-DVM for constrained problems 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Δ ← ComputeDelta(X ); α ← ρ ((1 − Δ)K1 + ΔK2 ); β ← 1 − α; Pr ← BuildRandomMatrix(β); Pd ← BuildDeterministicMatrix(α, H ); Pd ← FindSingleViolatedConstraint(g, h)  Rule 1 ; Pd ← CheckBoundaryDistance(l, u, )  Rule 2 ; if H = (1, . . . , 1) then H ← (0, . . . , 0); H ← UpdateHistory(H , Pd , Pa m )  Index of columns of Pd in Pa m are 1 in H; Pa m ← βPr ⊕ αPd ; c ← ChooseVariables(X , Pa m ); c ← UpdateStep(c)  Use any update function; X ← UpdateSolutions(X , c ); X ← SafeguardStep(X );

most recent survey of Akay et al. [4], the variant of Garg [9] (GABC); the Dual search ABC from Dong et al. [8] (DSMABC); and the Modified ABC from Akay et al. [2]. We also extend the comparison to the hybrid variant of the ABC with evolution strategies by Mollinetti et al. [18] (ABC+ES) and the modular ABC from [6] for multimodal problems (ABC-Xm1). With the exception of the GABC, the canonical versions of the algorithms use the following penalty function:  f (x) if x ∈ S m p (13) F (x) = /S f (x) + ψ( i=1 gi (x) + i=1 hi (x)) x ∈ ¯ ∈ X such that x ¯∈ where S denotes the search space. If there is a solution x / S, then it has violated either a inequality or equality constraint. The budget of function evaluations for each algorithm is set to 104 function evaluations (FE’s). The number of solutions for all algorithms but the ABC-Xm1 is fixed at 100 and Lit = SN · n. Specific parameters for each version of the ABC are: ABCXm1, Lit = 1.06 · n and maximum population of 66 and minimum of 15; MABC, F = 4; GABC, β = 100. Regarding penalty methods, ψ is set to 104 and for the Augmented Lagrangian method, initial values of ρ, γ b , τ and β are set to 2, 1.5, 0.85, and 100 respectively. Descriptive statistics are the mean, standard deviation, median, and bestworst results obtained from 30 runs with distinct random seeds. Validation of statistical relevance is performed for non-parametric data. We first run the Friedman test to verify overall differences, then we run the Wilcoxon signed rank test to assess pairwise differences adjusted by the Holm-Bonferroni error correction. Confidence level α for tests is set to 0.95. All experiments were conducted in

Augmented Lagrangian ABC-ADVM for Constrained Optimization

247

a machine with the following hardware configuration: Intel core i7-6700 “Skylake” 3.4 GHz CPU; 16 GB RAM DDR4 3200 clocked at 3000 MHz. The running operating system (OS) is UbuntuOS 18.04. All algorithms were written in the python 3 programming language. Floating point operations were handled by the numpy package, version 1.19.1. Table 1 describes the adjusted p-values of the best algorithm in the instance against all others. Plus ‘+’ signs mean that p < 0.05 rejecting the null hypothesis that there is no difference between groups/pairs while minus ‘−’ signs signify that p > 0.05, failing to reject the null hypothesis, indicating statistical significance. Regarding the omnibus test, all instances showed that p < 0.05, so we omit displaying the results. Table 2 show the result of the experiment for all instances. For each problem instance, the row in boldface indicates the algorithm whose mean was the lowest among the others. With the exception of the WBD2 instance, it is possible to observe that algorithms using our approach were shown to have the lowest value for the mean statistic, while in all instances, save for WBD2 and DPV2, our approach showed to have the lowest best statistic. These findings indicate an improvement in the robustness of the search. However, referring to Table 1, the only instance where our approach had statistically significant results against all others was in the WBD1 instance. Although the results were positive, the statistical analysis suggests that a larger experiments with larger sample sizes would be an interesting direction in order to further study the effects of our approach. Table 1. Adjusted p-values of the wilcoxon test of the best algorithm against the others Instance

ABC ABC-n ADVM-ABC ADVM-ABC-n MABC MABC-n GABC GABC-n ABC-ES ABC-ES-n DSM-ABC DSM-ABC-n ABCx-m1 ABCx-m1-n

TBT

−

−

−

DPV1

+

+

−

−

−

−

−

+

+

+

+

+

−

+

+

−

+

+

+

+

+

DPV2

−

−

−

+

+

+

−

+

+

+

+

+

SRD11

−

−

MWTCS −

−

−

−

−

−

−

−

WBD1

+

+

+

+

+

WBD2

+

−

−

+

+

+

+ +

−

−

+

+

+

+

+

+

+

−

+

+

+

+

+

+

+

+

+

+

+

+

+

+

−

+

+

+

+

−

−

−

248

M. A. F. Mollinetti et al. Problem Algorithm ABC ABC-n ADVM-ABC ADVM-ABC-n MABC MABC-n GABC WBD2 GABC-n ABC-ES ABC-ES-n DSM-ABC DSM-ABC-n ABCx-m1 ABCx-m1-n ABC ABC-n ADVM-ABC ADVM-ABC-n MABC MABC-n GABC TBT GABC-n ABC-ES ABC-ES-n DSM-ABC DSM-ABC-n ABCx-m1 ABCx-m1-n ABC ABC-n ADVM-ABC ADVM-ABC-n MABC MABC-n GABC MWTCS GABC-n ABC-ES ABC-ES-n DSM-ABC DSM-ABC-n ABCx-m1 ABCx-m1-n ABC ABC-n ADVM-ABC ADVM-ABC-n MABC MABC-n GABC DPV1 GABC-n ABC-ES ABC-ES-n DSM-ABC DSM-ABC-n ABCx-m1 ABCx-m1-n

7

Mean 2.7457 2.6534 2.7332 2.7856 2.7336 2.6914 2.7912 2.7915 2.9719 2.9913 2.5880 2.6377 2.6421 2.5920 263.9573 263.9492 263.9557 263.961 263.9585 263.9632 263.9538 263.9692 264.0306 264.0202 264.2604 264.3975 263.9501 263.9468 0.0138 0.0136 0.0136 0.0137 0.0137 0.0135 333333.3467 0.0139 0.0178 0.0151 666666.6813 333333.3485 1000000.0134 1333333.3461 6153.7737 6178.4582 6121.3575 6112.6501 6393.0998 6365.3417 6126.9501 6170.6485 7164.0138 7336.8538 6266.2742 6206.4391 6467.0699 6420.4694

Median 2.7343 2.6425 2.7012 2.7782 2.7237 2.6676 2.7583 2.7733 2.9857 2.9962 2.5830 2.6299 2.6005 2.5492 263.9417 263.9322 263.9486 263.9464 263.9486 263.9508 263.9381 263.9617 264.0121 264.0062 264.2126 264.357 263.9397 263.9373 0.0136 0.0134 0.0134 0.0135 0.0136 0.0133 0.0136 0.0135 0.018 0.0144 0.0152 0.0152 0.0144 0.0143 6142.3162 6170.622 6137.752 6106.1346 6433.659 6319.1355 6157.0338 6160.3636 7200.5792 7366.5811 6192.8362 6185.0459 6392.3652 6422.2794

Std. Dev 0.1899 0.1418 0.2097 0.1820 0.1501 0.1821 0.2002 0.1967 0.2106 0.1986 0.1110 0.1168 0.1722 0.1563 0.0533 0.0487 0.0387 0.0585 0.0441 0.0443 0.0418 0.0446 0.0812 0.0913 0.2224 0.3241 0.0343 0.0368 0.0009 0.0008 0.0007 0.0007 0.0007 0.0006 1795054.9337 0.0011 0.0046 0.0021 2494438.2546 1795054.9334 2999999.9964 3399346.3384 126.9886 88.5261 122.2212 98.3164 203.5749 193.6483 104.3036 116.2511 421.1919 390.5162 218.1412 212.1707 345.5378 231.731

Best 2.4561 2.4268 2.4687 2.4412 2.5054 2.4515 2.4814 2.4480 2.5732 2.6344 2.3982 2.4138 2.4248 2.4012 263.9056 263.897 263.9087 263.8978 263.9045 263.9048 263.9019 263.8983 263.903 263.9046 263.9762 263.9323 263.9015 263.8966 0.0128 0.0128 0.0128 0.0128 0.0128 0.0128 0.0128 0.0128 0.0128 0.0129 0.0128 0.0128 0.0128 0.0128 5916.2534 5976.8166 5901.3161 5879.3449 5905.0226 6122.529 5901.7107 5940.6232 6307.1935 6489.4701 5958.2312 5870.9949 6053.2353 6002.5038

Worst 3.0942 2.9696 3.2651 3.2326 3.1272 3.0914 3.0929 3.2414 3.4097 3.5267 2.8228 2.8861 3.1268 3.0866 264.1075 264.1406 264.0919 264.1918 264.0747 264.0837 264.0604 264.0643 264.212 264.2195 264.8997 265.2088 264.0383 264.0642 0.0171 0.0168 0.016 0.0159 0.0156 0.0153 10000000.0025 0.0171 0.032 0.0216 10000000.0025 10000000.0025 10000000.0025 10000000.0025 6470.9351 6360.8256 6438.4008 6337.9544 6787.2161 6831.0182 6363.01 6462.3077 8204.531 8330.421 6757.7076 6850.8925 7534.1776 7076.4776

ABC ABC-n ADVM-ABC ADVM-ABC-n MABC MABC-n GABC WBD1 GABC-n ABC-ES ABC-ES-n DSM-ABC DSM-ABC-n ABCx-m1 ABCx-m1-n ABC ABC-n ADVM-ABC ADVM-ABC-n MABC MABC-n GABC DPV2 GABC-n ABC-ES ABC-ES-n DSM-ABC DSM-ABC-n ABCx-m1 ABCx-m1-n ABC ABC-n ADVM-ABC ADVM-ABC-n MABC MABC-n GABC SRD11 GABC-n ABC-ES ABC-ES-n DSM-ABC DSM-ABC-n ABCx-m1 ABCx-m1-n

2.7211 2.7321 2.7185 2.7572 2.7388 2.8067 2.6907 2.6403 2.9623 2.9317 2.5791 2.5638 2.634 2.5979 6227.0619 6218.0491 6239.5817 6224.5056 6458.3826 6498.2494 6214.9832 6197.2317 7200.5736 7328.2525 6377.096 6410.6303 6554.3604 6521.7902 2894.7012 2894.6945 2894.7517 2894.7327 2894.8356 2894.8608 2894.6999 2894.7761 2935.9379 2938.4584 2898.2751 2898.1767 2896.0631 2895.9659

2.6938 2.7183 2.6942 2.7573 2.7262 2.7915 2.6896 2.5913 2.9732 2.9284 2.5244 2.5564 2.5486 2.546 6208.3918 6227.94 6229.6174 6217.1737 6468.3317 6475.5471 6198.4474 6175.0342 7193.5074 7318.0772 6323.7024 6381.5714 6501.9594 6460.0748 2894.6314 2894.6029 2894.6119 2894.6714 2894.7755 2894.8032 2894.6198 2894.6897 2938.2581 2939.345 2898.2501 2898.2477 2896.0022 2895.7185

0.147 0.1269 0.1882 0.1633 0.1706 0.1853 0.1496 0.1409 0.1677 0.2213 0.1440 0.0824 0.1903 0.1657 105.0794 94.9543 95.6531 85.1596 213.4122 203.1428 99.3421 119.9039 399.1733 373.5517 197.3295 234.0953 390.7208 293.7438 0.2412 0.252 0.4379 0.2776 0.2431 0.2855 0.2201 0.2441 15.1853 15.7172 1.901 1.6788 0.9969 0.9396

2.4767 2.5019 2.4304 2.4974 2.4747 2.4710 2.4185 2.4693 2.5492 2.5677 2.4144 2.4069 2.4371 2.4357 6069.4265 6015.699 6087.3643 6084.6043 6040.701 6156.7328 6042.1196 5999.8275 6371.909 6708.4105 6079.7942 6002.3775 5971.1354 6133.9042 2894.4304 2894.4019 2894.424 2894.4107 2894.5251 2894.4839 2894.4236 2894.4177 2906.0677 2902.0717 2895.1792 2895.9831 2894.6734 2894.5359

3.0624 3.0946 3.1075 3.1627 3.1401 3.1700 2.971 2.9729 3.311 3.5160 2.8959 2.7908 3.1700 3.2587 6462.7758 6452.9514 6414.5279 6438.7577 6855.076 6947.5088 6411.0217 6434.6977 8212.8845 8423.0431 6732.196 6960.7568 7522.6448 7343.2997 2895.3969 2895.5988 2896.4367 2895.7321 2895.481 2895.5315 2895.3175 2895.3475 2973.2089 2969.1139 2901.8787 2903.2806 2898.5333 2898.4484

Conclusion

This work introduced two novel approaches for derivative-free methods to handle nonlinear constrained problems with nonlinear constraints. First, a self-adaptive deterministic decision variable selection (A-DVM) scheme based on Mollinetti et al. [16] for nonlinear optimization problems integrated to the Artificial Bee Colony where two new rules were added to the construction of the deterministic decision variable matrix Pd . Second, an adaptation to the augmented lagrangian method that serves as a penalty method to derivative-free algorithms that use a solution set in their optimization process. Results of the experiment show that our approach showed statistically significant improvement in relation to their original counterparts. Positive results can be observed in six of the seven instances, suggesting that the penalty method along with the selection scheme can be a promising direction for populationbased heuristics to tackle constrained optimization problems with nonlinear constraints. We can highlight possible directions from this work. On the A-DVM side, integrations to more classical metaheuristics, e.g., Particle Swarm Optimization (PSO); Genetic Algorithm (GA); Evolution Strategies (ES); and so on, using a larger pool problem instances. On the side of the augmented lagrangian penalty method, it would be fruitful to test the method to optimization problems with

Augmented Lagrangian ABC-ADVM for Constrained Optimization

249

more stringent constraints. Another interesting direction would be towards optimizing the discriminability of subspace methods for supervised learning tasks such as in [11] and [10], where the problem can be formulated as (8).

References 1. Abraham, A., Jatoth, R.K., Rajasekhar, A.: Hybrid differential artificial bee colony algorithm. J. Comput. Theor. Nanosci. 9(2), 249–257 (2012) 2. Akay, B., Karaboga, D.: A modified artificial bee colony algorithm for realparameter optimization. Inf. Sci. 192, 120–142 (2012) 3. Akay, B., Karaboga, D.: A survey on the applications of artificial bee colony in signal, image, and video processing. Signal Image Video Process. 9(4), 967–990 (2015). https://doi.org/10.1007/s11760-015-0758-4 4. Akay, B., Karaboga, D., Gorkemli, B., Kaya, E.: A survey on the artificial bee colony algorithm variants for binary, integer and mixed integer programming problems. Appl. Soft Comput. 106, 107351 (2021) 5. Audet, C.: Derivative-Free and Blackbox Optimization. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-68913-5 6. Aydın, D., Yavuz, G., St¨ utzle, T.: ABC-X: a generalized, automatically configurable artificial bee colony framework. Swarm Intell. 11(1), 1–38 (2017). https:// doi.org/10.1007/s11721-017-0131-z 7. Birgin, E.G., Mart´ınez, J.M.: Practical augmented Lagrangian methods for constrained optimization. SIAM (2014) 8. Dong, C., Xiong, Z., Liu, X., Ye, Y., Yang, Y., Guo, W.: Dual-search artificial bee colony algorithm for engineering optimization. IEEE Access 7, 24571–24584 (2019) 9. Garg, H.: Solving structural engineering design optimization problems using an artificial bee colony algorithm. J. Ind. Manage. Optim. 10(3), 777–794 (2014) 10. Gatto, B.B., dos Santos, E.M., Koerich, A.L., Fukui, K., Junior, W.S.: Tensor analysis with n-mode generalized difference subspace. Expert Syst. Appl. 171, 114559 (2021) 11. Gatto, B.B., dos Santos, E.M., Molinetti, M.A., Fukui, K.: Multilinear clustering via tensor Fukunaga-Koontz transform with fisher eigen spectrum regularization. Appl. Soft Comput. 113, 107899 (2021) 12. Karaboga, D.: An idea based on honey bee swarm for numerical optimization. Erciyes University, Technical report (2005) 13. Karaboga, D.: An idea based on honey bee swarm for numerical optimization. Technical report, Technical report-tr06, Erciyes University, Engineering Faculty, Computer Engineering Department (2005) 14. Locatelli, M., Schoen, F.: Global optimization: theory, algorithms, and applications. SIAM 15 (2013) 15. Mc Ginley, B., Maher, J., O’Riordan, C., Morgan, F.: Maintaining healthy population diversity using adaptive crossover, mutation, and selection. IEEE Trans. Evol. Comput. 15(5), 692–714 (2011) 16. Mollinetti, M.A.F., Gatto, B.B., Neto, M.T.R.S., Kuno, T.: A-DVM: a selfadaptive variable matrix decision variable selection scheme for multimodal problems. Entropy 22(9), 1004 (2020) 17. Mollinetti, M.A.F., Neto, M.T.R.S., Kuno, T.: Deterministic parameter selection of artificial bee colony based on diagonalization. In: International Conference on Hybrid Intelligent Systems (2018)

250

M. A. F. Mollinetti et al.

18. Mollinetti, M.A.F., Souza, D.L., Pereira, R.L., Yasojima, E.K.K., Teixeira, O.N.: ABC+ES: combining artificial bee colony algorithm and evolution strategies on engineering design problems and benchmark functions. In: Abraham, A., Han, S., Al-Sharhan, S., Liu, H. (eds.) Hybrid Intelligent Systems. HIS 2016. AISC, vol. 420, pp. 53–66 (2016). Springer, Cham. https://doi.org/10.1007/978-3-319-2722145 19. Nash, S.G.: Linear and nonlinear programming Engineering & Mathematics. McGraw-Hill Science, New York (1996) 20. Nocedal, J., Wright, S.J.: Numerical Optimization, vol. 2. Springer, New York (2006). https://doi.org/10.1007/978-0-387-40065-5

Remote Monitor System for Alzheimer Disease Luis B. Elvas1

, Daniel Cale1,2

, Joao C. Ferreira1,2(B)

, and Ana Madureira2,3

1 Instituto Universitário de Lisboa (ISCTE-IUL), ISTAR, 1649-026 Lisbon, Portugal

[email protected]

2 Inov Inesc Inovação—Instituto de Novas Tecnologias, 1000-029 Lisbon, Portugal

[email protected] 3 ISRC - ISEP/P.PORTO, Porto, Portugal

Abstract. Health Remote Monitoring Systems (HRMS) offer the ability to address health-care human resource concerns. In developing nations, where pervasive mobile networks and device access are linking people like never before, HRMS are of special relevance. A fundamental aim of this research work is the realization of technological-based solution to triage and follow-up people living with dementias so as to reduce pressure on busy staff while doing this from home so as to avoid all unnecessary visits to hospital facilities, increasingly perceived as dangerous due to COVID-19 but also raising nosocomial infections, raising alerts for abnormal values. Sensing approaches are complemented by advanced predictive models based on Machine Learning (ML) and Artificial Intelligence (AI), thus being able to explore novel ways of demonstrating patient-centered predictive measures. Low-cost IoT devices composing a network of sensors and actuators aggregated to create a digital experience that will be used and exposure to people to simultaneously conduct several tests and obtain health data that can allow screening of early onset dementia and to aid in the follow-up of selected cases. The best ML for predicting AD was logistic regression with an accuracy of 86.9%. This application as demonstrated to be essential for caregivers once they can monitor multiple patients in real-time and actuate when abnormal values occur. Keywords: Alzheimer disease · Dementia · Prevention · Machine Learning · Artificial Intelligence · Health Remote Monitoring Systems · Data analytics · IoT

1 Introduction Dementia is an increasing problem in modern aging societies worldwide, and particularly in medium income countries like Portugal with a frailer social and health care systems, with a high burden of disease and scoring as the fastest aging society in Europe [1]. Dementia poses multiple challenges such as the optimization of current processes for triaging, evaluating, and monitoring. Specialist skills and resources are limited and cannot scale to meet demand. Alzheimer disease (AD) is the most common type of dementia and therefore most of previous studies in the last years were performed in an AD context [2]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 251–260, 2022. https://doi.org/10.1007/978-3-030-96299-9_24

252

L. B. Elvas et al.

The costs of healthcare and long-term care for individuals with dementia are substantial [3]. The number of people living with dementia in the EU27 is estimated to be 7,853,705 and it is one major cost [4]. Even more, on 18th of February 2020, Alzheimer Europe presented a new compelling report at the European Parliament with findings on the raising prevalence rates for dementia in Europe, as one significant cause of health costs and struggling with staff storage [4]. The number of people living with the condition is set to double by 2050 according to the new Alzheimer Europe report [4], which will only place greater pressure on care and support services unless better ways of treating and preventing dementia are identified. Dementia is a major social and health concern, with increasing prevalence [5]. Having said this, this topic becomes relevant as the HRMS allow the collection of parameters on the patients in order to understand the evolution and generating alerts for atypical behaviors requiring intervention, thus saving on staff costs, which, as said before, are high. Care processes digitalization, holistic sensing supported by the Internet of Things (IoT) system and Artificial Intelligent (AI) tools are being actively applied to the health sector giving rise to the smart health paradigm [6]. This emerging market was evaluated in USD for 143.6 billion in 2019, and it has an estimated annual growth rate of 16.2% between 2020 and 2027 [7]. A Scopus search showed that were 1066 publications between 2010 and 2020 regarding Smart Health, and by Fig. 1. We can see its exponential growth.

Fig. 1. Number of studies per year

Fig. 2. Topics by subject area

On Fig. 2 we can see the top 5 subject areas of study where it is noticeable that the area with more impact is Computer Science. In this transformative process, the Health Remote Monitoring Systems (HRMS) are recognized as an emerging technology, using sensors and wearable devices to collect patient’s data. However, to present clinical value these systems have to be associated with clear clinical processes and therapeutics, so the measurements could be linked with actual patient care. The implementation of HRMS is based on a health related IoT system that incorporates, stores and communicates the information gathered by a set of wearable devices and sensors. The computer senses and records the daily physiological data of the patient by means of a data processing device, data transition, data archive, data analytics and AI [8]. The HRMS developed is based on two main pillars and aims to: (1) develop a system to identify disease development and disease prevention through the use of remote sensors (2) develop prediction models built by Artificial Intelligence implemented on

Remote Monitor System for Alzheimer Disease

253

top of the processed data, which would allow the classification of patients, discovering behavioral patterns. Through this process, alerts by email are generated when an abnormality is registered in order to help with disease prevention. Therefore, HRMS enables a data-intensive approach, in which a large amount of health data is generated, stored and available for data mining, allowing for the generation of useful knowledge. A quite recent review [9] emphasizes the relevance of taking comorbidity burden into account when investigating dementia progression. Therefore, this research work takes inspiration from the ICOPE guidelines [10] to offer an integrated and person-centered approach to AD. This research work demonstrates that simple tasks with a small IoT device can be interpreted by a trained Artificial Intelligence tool, in order to automatically and remotely determine health and disease related aspects, the progress of dementia, and make outcome predictions, in order to support health practitioners and reducing their workload.

2 Literature Review A systematic literature review was made by following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analysis) Methodology [11], and with the research question “What is the state of the art on Healthcare Remote Monitoring Systems usage in the prevention of patients suffering from Alzheimer’s disease?”. Databases searched were Scopus and Web of Science Core Collection (WoSCC) and the research was conducted through October 26th, 2021; all the results had to be articles, published between 2011–2021 and written in English. The documents collected were only about Computer Science, Medicine, Engineering and Health Professions. The search strategy was based on one query made with different focuses of research. This method allowed for the observation of the number of articles existing in both databases, considering the concept and context, and population under study. This method allowed for the observation of the number of articles existing in both databases, considering the concept and context, and population under study. It is important to note that the values corresponding to the queries still have duplicate articles. For this review only articles were considered. Grey literature, reviews, conference papers, workshops, books, and editorials were excluded, as well as works not related to the domain. All the databases were searched systematically regarding the published work related with the concept “Health Remote Monitoring Systems” or “Smart Health”, the target population “Alzheimer’s Disease” and within a “Prevention” context of the study. After performing a manual process, towards the identification of significant subjects on their research questions, identifying the outcomes and removing the duplicates, 17 documents were obtained. Considering the goals of this article is to identify the use of HRMS on Alzheimer, a list of the main topics discussed on each of the reviewed articles are described on Fig. 3, where it is noticeable the focus on the prevention of Alzheimer by resorting to the use of wearables. After assessing all the included studies, it was possible to acknowledge the growing of the remote monitoring systems across the globe in the recent years. Considering the

254

L. B. Elvas et al.

Fig. 3. Main Topics from the Literature review

Fig. 4. Main objectives of the reviewed

goals of this article is to identify the use of HRMS, a list of the main objectives of each system described in the reviewed articles was performed (Fig. 4). From the most popular topic, authors from [12] acknowledge the growing number of studies on this matter. Brid, M. et al. [13] present us that e-health tools to improve dietary behavior lacks credibility. From [14] authors don’t present any results, nevertheless, on study [15] arrive to the conclusion that there is no positive association between physical activity and working memory. On study [16] no results about the impact of the Webmobile application are presented, allowing only that users know the risks if they continue with the same lifestyle habits. Regarding HRMS, it is secure to merge the topics that are from wearables and from monitoring devices (Balance Evaluation, Instability of motor function monitoring, Sleep Monitoring and Falls Prediction and Prevention). On study [17], the authors by sleep monitoring can track cognitive changes cognitive changes related to pre-clinical Alzheimer Disease, recurring only to wearable biosensor devices. Author of study [18] review the last fifteen years of wireless sensors on neurological disorders, arriving to the conclusion that this devices have been greatly used on the assessment of this disorders, retrieving valuable data. On [19], a device is placed under the patient mattress, to monitor his sleep parameters, in order to characterize sleep disturbances in patients with AD, and arriving to the conclusion that frequent bed leaving is a highly correlated predictor of the duration of dementia. Researchers on [20] describe a noninvasive sensor for patient movement monitoring, identifying patterns that would lead to a fall and predicting them. On [21] the sleep is evaluated using wearables, and this data has been revealed as a significant predictor on identifying AD and the patient’s health status. Lastly, on study [22], by using home-based technologies combined with teleassistance service in elderly persons with Alzheimer’s disease the number of incidents has suffered a considerable reduction.

3 Methods and Materials Recent recommendations for health and care workers point out that it is dramatically necessary to develop and carry out a person-centered integrated care for older people (ICOPE) in which an integrated and person-centered approach is required as declines in intrinsic capacity are interrelated. Considering current state of art, this research work handles several technical challenges to creating an AI based solution for this challenge in computer science: 1) Fusion

Remote Monitor System for Alzheimer Disease

255

of different data sources in a big data for dementia; 2) Successfully training ML algorithm to classify the risk; and 3) Present meaningful and useful data to decision makers to support improved interventions. Our method is complemented by advanced predictive models based on ML and AI, thus being able to explore novel ways of demonstrating person-centered predictive measures. AI provides a better decision-making in the adjustment of the parameters of each evaluation process of the disease so that we achieve a user-centered approach but also a generalization to other patients. Our proposed method bases on different technological advancements to complement and to reorient current early diagnosis, prevention, and intervention of AD, including the advantage to be prepared for evaluating the progression of the AD condition through time. These technological advancements have a significant impact in terms of: Personalized early-risk diagnosis, prevention, and evaluation of AD through (1) estimating the probability of suffering AD, (2) achieving an earlier and better intervention, and (3) improving the quality of life of the citizens. Creation of pathways to manage cognitive (and related) decline by (1) getting early warning signs of likely deterioration for citizens and professionals, and (2) increasing health literacy in the interpretation of symptoms and effects. 3.1 Data Collected For the analysis purpose, it was collected data from different sensors, where a prototype was developed joining different sensos in just one device (see Fig. 5).

Fig. 5. Prototype of the IoT sensor with 4 types of health measurements

This device is a sensor prototype developed for the collection of health data from the test subjects, collecting 4 biometric parameters, including heart rate, arterial oxygen, body temperature and Galvanic Skin Response (GSR) The sensor prototype consists of two components: the microcontroller with wireless communication and sensors for biometric data collection. 3.2 Big Data Analytics and Machine Learning The multiple types of data within this research work are gathered, stored, and processed. The data was collected a the time frame between 2 and 3 weeks per test subject and with an average of over 3 measurements per day, preferably between the morning, afternoon,

256

L. B. Elvas et al.

and evening before bedtime. The prototype used not only sends our raw data, but instead sends the data already structured and pre-processed, this is the same to say, that if a given value appears to be complete nonsense the microcontroller won’t send that value. For this reason, we will make our data immediately available in our database, avoiding unnecessary operations of ETL and data treatments. In addition to the IoT sensor, a cloud server is part of the intelligent system for automatic data processing functions and standardizing the collected data and storing it in a MariaDB database. This server has Node RED [23] mechanism that will analyze all the sensor data, processing it and applying the machine learning algorithms, generating the dashboards and respective alerts when necessary, as depicted on Fig. 7. With the data stored (without personal information) of the test subjects, it can be used to develop solutions using AI and obtain relevant information to improve the quality of life and care of patients with Alzheimer’s disease. For the predictive analysis, we have implemented a Logistic Regression Algorithm on top of the data collected for 6 months from a group of 15 patients where 6 of these patients suffer from AD. The data was collected from the sensor depicted on Fig. 5, retrieving the data from sensors of the heart rate, oximeter, body temperature and GSR measurement. A dataset from the last 6 months was compiled with a classification (dependent) variable (where 1 means the patient suffers from AD and 0 means the patient does not suffer from AD). We have applied different ML algorithms, where Logistic Regression as given better results, achieving 86,89% of accuracy when compared to K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Naïve Bayes, Decision Trees amd Random Forest, as depicted on Fig. 6.

Fig. 6. Algorithm performance

To test our model, we have used the data from for different patients selected, where in Table 1 we have a summary of the data collected by using the prototype sensor. 3.3 Decision Support and Visualization The vast data acquired from all the sensors is analyzed and organized visually so health literacy can be increased by providing manageable and understandable personal data to the patient, and giving advances in professionals’ proficiency in interpretation. The definition of Key Performance Indicators (KPIs) for dementia follow-up we have defined the data stated previously, being the variables used on this research work: heart rate, arterial oxygen, arterial oxygen, temperature, GSR.

Remote Monitor System for Alzheimer Disease

257

To increase the health literacy, dashboards were designed and implemented, recurring to the set of KPIs extracted from the values produced by the sensors, being this presented to the different users in an understandable way that aids in their tasks and decisions. In this sense, two separate sets of dashboards were implemented. The first set of dashboards implemented focus on improving the proficiency of clinicians in early-risk diagnosis dashboards. The second set of dashboards focus on health literacy dashboards, aiming to understand existing risk factors and the conditions of patients in order to improve healthrelated decisions. To this aim, this task includes the output of predictive systems that aid in the diagnosis based on data collected in the past. The design and implementation of dashboards for advancing proficiency in data-oriented health services will use the KPIs already identified, being this presented to the different users in an understandable way, complemented with additional information that aids in their tasks and decisions. Table 1. Summary of biometric data collected per test subject. Test subject 1

Test subject 2

Test subject 3

Test subject 4

Age rate

~65

~70

~55

~45

Heart rate-max/min (bpm)

132/75

128/70

140/83

149/90

Daily average (bpm)

100

84

93

88

95

89

98

98

Arterial oxygen Daily average (%) Temperature Daily average (ºC)

35.7

36.5

36.9

37.2

GSR daily average (ohm)

512

889

684

856

Feeling

Happiness

Grief

Normal

Anxiety

As can be seen in Figs. 7, 8 and 9, the user dashboard shows the history of the data collected over the test periods, as well as the individual sensor measurements and an alert system. The alarm system consists of three prompts, Error Measure, Risk factors and Early-risk Diagnosis. When the system detects a failure of the sensor to collect the data or an invalid value it gives the Error Measure alert and turns red. In the case of Risk factor the alarm information and color indicates if the collected values present risk factors for the health of the test subject, where green means “low risk”, yellow means “some risk” and red alerts for health risks. In the Early-risk diagnosis alarm, the result of the analysis of the data collected from the test subject is indicated and where the system informs that it has detected a probability of diagnosis of Alzheimer’s disease. The Early-risk Diagnosis comes from the ML algorithm, where: 1) Green represents “low risk” (probability of developing AD from 0–49); 2) yellow represents “some risk” (probability of developing AD from 50–70); 3) red represents “high risk” of developing AD. The caregiver, on the dashboard can select the Risks Factors and a Dashboard is presented to him, showing which warning the visaed patient has, see Figs. 7 and 8.

258

L. B. Elvas et al.

Fig. 7. User dashboard implemented with data collected by the sensor and alarms

Fig. 8. Dashboard with the KPI requiring attention

When a certain value requires attention, an alert is sent to the caregiver as depicted in Fig. 9.

Fig. 9. Notification alert

4 Conclusions The literature review showed that the topic studied is not yet exhausted, returning 35 papers, and of these 35, only 17 papers met the inclusion criteria. Throughout the literature, we can perceive the wide use of wearables in Alzheimer’s prevention, and the most common objective of the eligible studies is lifestyle change. This topic has not proved to be the most relevant in the scientific community, being considered not very credible and is also notable for the lack of results in the studied papers. The use of HRMS is the one that prevails in quantitative terms in the cluster of the different topics. Of these, those within the HRMS are quite concise and with interesting results and impact on patients

Remote Monitor System for Alzheimer Disease

259

suffering from AD. From the data, and after the application of ML algorithms, it was realized that the best algorithm to deal with this type of data is the Logistic Regression. With this application, caregivers can keep better track of the patient, also receiving real-time notifications when a certain value to be monitored goes out of the standard. This way, patients are safer, since each caregiver can monitor different patients simultaneously. Regarding visualizations, we implemented different dashboards adapted according to: (i) the tasks and responsibility assigned to each actor (i.e. the relevant information they should be aware of), (ii) their particular perspective of the ecosystem (i.e. what are they interested in), and (iii) which other actors they interact with (i.e. people with dementia, informal careers, etc.). We complemented our solution with wearables and process their signal to extract psychological and physiological information (using edge computing), including IMU (accelerometer/gyroscope), electrodermal activity (EDA) and photoplethysmography (PPG) sensors, among others to be studied. This includes the challenging denoising of the EDA and PPG signals by filtering out the effect of motion artifacts and the ambient variations (light, temperature, humidity). Environmental IoT signal processing using edge computing that communicate with the IoT devices, are able to acquire the streams of raw data that these devices produce to quantify the ambient parameters (lighting, color, temperature, humidity, sound). With this new data, we will enlarge our dataset, therefore the use of dimensionality reduction algorithms can become necessary. Dimensionality reduction seeks and exploits the inherent structure in the data, describing the data using less information, therefore avoiding data (and model) complexity.

References 1. Dementia: Sep. 2021. https://www.who.int/news-room/fact-sheets/detail/dementia. Accessed 26 Nov 2021 2. Boustani, M., Schubert, C., Sennour, Y.: The challenge of supporting care for dementia in primary care. Clin. Interv. Aging 2(4), 631–636 (2007) 3. 2020 Alzheimer’s disease facts and figures. Alzheimers Dement. 16(3), 391–460 (2020). https://doi.org/10.1002/alz.12068 4. Alzheimer Europe – Alzheimer Europe – Our work – Annual Reports – Annual Report 2019. https://www.alzheimer-europe.org/Alzheimer-Europe/Our-work/Annual-Reports/Ann ual-Report-2019. Accessed 21 Jun 2021 5. Xia, X., Jiang, Q., McDermott, J., Han, J.-D.J.: Aging and Alzheimer’s disease: comparison and associations from molecular to system level. Aging Cell 17(5), e12802 (2018). https:// doi.org/10.1111/acel.12802 6. Röcker, C., Ziefle, M., Holzinger, A.: From computer innovation to human integration: current trends and challenges for pervasive healthtechnologies. In: Holzinger, A., Ziefle, M., Röcker, C. (eds.) Pervasive Health. HIS, pp. 1–17. Springer, London (2014). https://doi.org/10.1007/ 978-1-4471-6413-5_1 7. Global Smart Healthcare Market Size Report, 2020–2027.” https://www.grandviewresearch. com/industry-analysis/smart-healthcare-market. Accessed 21 Jun 2021 8. Malasinghe, L.P., Ramzan, N., Dahal, K.: Remote patient monitoring: a comprehensive study. J. Ambient. Intell. Humaniz. Comput. 10(1), 57–76 (2017). https://doi.org/10.1007/s12652017-0598-x

260

L. B. Elvas et al.

9. Haaksma, M.L., et al.: Comorbidity and progression of late onset Alzheimer’s disease: a systematic review. PLoS ONE 12(5), e0177044 (2017). https://doi.org/10.1371/journal.pone. 0177044 10. Integrated care for older people: guidelines on community-level interventions to manage declines in intrinsic capacity. Geneva: World Health Organization, 2017. http://www.ncbi. nlm.nih.gov/books/NBK488250/. Accessed 21 June 2021 11. Moher, D., Liberati, A., Tetzlaff, J., Altman, D.G.: Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. BMJ 339, b2535 (2009). https://doi.org/ 10.1136/bmj.b2535 12. Bott, N.T., et al.: Face-to-face and digital multidomain lifestyle interventions to enhance cognitive reserve and reduce risk of alzheimer’s disease and related dementias: a review of completed and prospective studies. Nutrients 11(9), 2258 (2019). https://doi.org/10.3390/nu1 1092258 13. Bird, M., et al.: The psychological determinants of making lifestyle and dietary behaviours after using an online cognitive health tool and its associated recommendations for protective cognitive health behaviours. Eur. J. Psychiatry 35(3), 145–156 (2021). https://doi.org/10. 1016/j.ejpsy.2021.02.001 14. Aalbers, T., Baars, M.A.E., Qin, L., de Lange, A., Kessels, R.P.C., Olde Rikkert, M.G.M.: Using an ehealth intervention to stimulate health behavior for the prevention of cognitive decline in dutch adults: a study protocol for the brain aging monitor. JMIR Res. Protoc. 4(4), e130 (2015). https://doi.org/10.2196/resprot.4468 15. Wouters, H., et al.: Physical activity and cognitive function of long-distance walkers: studying four days marches participants. Rejuvenation Res. 20(5), 367–374 (2017). https://doi.org/10. 1089/rej.2016.1876 16. Méndez-Sanz, R., de la Torre-Díez, I., López-Coronado, M.: What is your risk of contracting alzheimer’s disease? A telematics tool helps you to predict it. J. Med. Syst. 40(1), 1–8 (2015). https://doi.org/10.1007/s10916-015-0369-1 17. Saif, N., et al.: Feasibility of using a wearable biosensor device in patients at risk for alzheimer’s disease dementia. J. Prev. Alzheimer’s Dis. 7(2), 104–111 (2019). https://doi. org/10.14283/jpad.2019.39 18. Zampogna, A., et al.: Fifteen years of wireless sensors for balance assessment in neurological disorders. Sensors 20(11), 3247 (2020). https://doi.org/10.3390/s20113247 19. Higami, Y., Yamakawa, M., Shigenobu, K., Kamide, K., Makimoto, K.: High frequency of getting out of bed in patients with Alzheimer’s disease monitored by non-wearable actigraphy. Geriatr. Gerontol. Int. 19(2), 130–134 (2019). https://doi.org/10.1111/ggi.13565 20. Abbate, S., Avvenuti, M., Light, J.: MIMS: a minimally invasive monitoring sensor platform. IEEE Sens. J. 12(3), 677–684 (2012). https://doi.org/10.1109/JSEN.2011.2149515 21. Guarnieri, B., et al.: Multicenter study on sleep and circadian alterations as objective markers of mild cognitive impairment and alzheimer’s disease reveals sex differences. J. Alzheimers Dis. 78(4), 1707–1719 (2020). https://doi.org/10.3233/JAD-200632 22. Tchalla, A.E., et al.: Preventing and managing indoor falls with home-based technologies in mild and moderate Alzheimer’s disease patients: pilot study in a community dwelling. Dement. Geriatr. Cogn. Disord. 36(3–4), 251–261 (2013). https://doi.org/10.1159/000351863 23. Node-RED. https://nodered.org/. Accessed 26 Nov 2021

Detection of Social Distance and Intimation System for Covid-19 S. Anandamurugan(B) , M. Saravana Kumar(B) , K. Nithin(B) , and E. G. Prashanth(B) Kongu Engineering College, Perundurai, Erode, Tamil Nadu, India {nithink.18it,prashantheg.18t}@kongu.edu

Abstract. COVID-19’s is a novel corona virus, fast spread has caused substantial damage and affected more than tens of millions of individuals around the world. People frequently wear masks to safeguard themselves and others against the transmission of coronavirus. The world health organization conveys the people to follow the social distancing to prevent the spread of Covid. Researchers have proposed several machine learning models to classify the disease, but none have identified the algorithm which gives more accuracy. Also, similar studies that have proposed various other techniques for prediction. In addition to that maintaining social distancing is also a major factor. In these regions, personally monitoring whether individuals are maintaining social distancing or not is quite impossible. This study aims to develop a highly accurate and real-time technique for the automatic detection of individuals who are not maintaining social distancing. Three state-of-the-art object identification models, namely YOLOv4, Tiny-YOLOv4 are used to detect the objects. Many results suggest that YOLO v4 has the greatest mAP value of 88.90%, followed by YOLO v4 and Tiny-YOLO v4 with mAP values of 82.24% and 74.80%, respectively. Keywords: You Look Only Once Version3 (YOLOV3) · Convolutional Neural Networks (CNN) · Region of Interest (ROI) · Structured Query Language (SQL) · You Look Only Once Version4 (YOLOV4) · Artificial Intelligence (AI)

1 Introduction The image recognition and distance between two frames are the key factor in this project. It also included the alert system to maintain the people to follow the social distancing and the violation count, date, place, time are also included in the database for future reference. From this information we come to know that the area has a highest chance of spread of Covid-19. In addition to that the alert is maintained when the people are not maintaining social distancing. These data are stored in database for future reference. With these data we can find the probability of spread of covid-19 in that particular place. The objective is creating a system established by camera and YOLOV4. YOLOV4 is used to improve the efficiency of identifying person with better accuracy and speed. The purpose of the system is to minimize the manual work by identifying the people through video analysis and calculating the social distancing violation of people. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 261–269, 2022. https://doi.org/10.1007/978-3-030-96299-9_25

262

S. Anandamurugan et al.

1.1 You Look Only Once Version4 YOLOV4 is advance level of YOLOV3. It includes mean average is increased to ten percentage higher when compared to YOLOV3. It includes the fifteen percentage higher of frames per second detection. The detection speed is twice that of YOLOV3. The YOLOV4 performance can be increased by adopting ten methods. The first one gathering of more data. The second one is Preprocessing and other eight techniques to improve YOLOV4 performance. It has a higher speed of interference. In this project we had include some of the advanced function for detecting the objects and to have a high accuracy rate. YOLOV4 is the name comes from one time object detector to have a higher execution rate and speed. Before getting into the object detection the YOLOV4 will do object segregation and object localization and instance segmentation to find the particular object [1]. YOLOV4 contains several layers in which the algorithm splits into layer by layer and finds the particular object. The PyCharm is used to edit the code and it also act as IDE. The Python 3.3.9 is used in this project The mathematical formula Centroid for finding the center of the frame and Euclidean formula is used here for distance between two frames in the unit of pixel. 1.2 Distance Between Two Frames The distance between two frames is calculated by distance formula. In this project we use pixel pitch algorithm i.e. (Euclidean formula) between two frames. We use python scipy package to find the distance between two frames. In this we take centroid of the frames with the help of python package NumPy to find the centroid of the frames. With the help of these formula, we can find the distance between the two frames. If the distance is minimum than minimum distance a red with violation occurs. This formula, can be obtained in the package of scipy package in python. With the help of Euclidean formula to find the distance between two frames. If the distance is lesser than minimum distance a violation occurs with the red frame and those who follow the social distancing will encounter for green frame.  E(X , Y ) = (x1 − x2)2 + (y1 − y2)2

1.3 Alert System and Database If the violation is high an alert system will be produced. This violation can be stored in a database with place and time for future use. We use XAMPP and PHP MYADMIN to store the data in the table. We use MYSQL to communicate between the database. The full form of SQL is Structured Query Language. In this Project we used SQL query to insert the violation count into the database. The SQL is first defined by ANSI. The insert command is used here to update the value in the database. If the violation count is higher an alert system is produced to maintain the social distancing. This frequency will be changed with respective to violation count. The alert system is the major point to follow the social distancing. The sound is beep sound which

Detection of Social Distance and Intimation System for Covid-19

263

is the speaker tells to follow the social distancing. It also tells the violation count in that particular place itself to make a people to had an extra idea to follow the social distancing (Fig. 1).

Fig. 1. Shows violation if it is less than 6 feet.

2 Problem Statement With the help of social distancing, we can able to control the spread of Covid-19. People always forget to follow the social distancing which may leads to spread of Covid-19. In this type of situation, we can make our people to follow the social distancing to avoid the spread the Covid-19. The problem statement is to make our people to follow the social distance in the crowded area. This may help to prevent the spread of Covid-19 in that crowded area. There is a problem in an existing system which may video stream to find the violation. This is not an accurate measure of finding the social distancing because it may lead to time waste and also needs a person work to collect the video frame and to find the social distancing violation.

3 Proposed System Of all of the approaches proposed in the literature, the YOLOV4 appears to the most promising for the counting of people. Since YOLOV4 is fast and accurate object detection algorithm. The YOLOV4 approach allows both finding the people and also identification/extraction of the pixels associated with each individual person. The proposed social distancing violation counting System. Video analysis method is used for social distancing Violations. While the existing method requires hardware like drones, sensors, proposed method does not require any hardware since the algorithm can able to process the video that is captured by camera. YOLOV4 though correctly identifies and counts the people with better accuracy it is comparatively slower than YOLOV3 [2] (Fig. 2).

264

S. Anandamurugan et al.

Fig. 2. Shows full execution system of our project.

4 Related Work 1) YOLOV4 is fast and accurate object detection algorithm. YOLOV4 divides the given images into grid cells and these grid cells are processed simultaneously and this is why the YOLOV4 is faster. YOLOV4 is capable of detecting multiple objects in each image that makes it more efficient and accurate than other algorithms. This Fig. 3 shows the architecture of YOLOV4. 2) The project only requires camera as its hardware requirement since the algorithm can able to process the video that is captured by camera. The camera can be of any type like mobile camera, digital camera which is easily available among users. By using the proposed method, time and cost is saved and greater accuracy is achieved. Here, live detection analysis method is used for detecting the people. If the video is taken with the help of the Euclidean distance formula, we can find the distance between the two-centroid frame. There are certain limitations in the existing dataset. So, we have used weights file to train the YOLOV4 algorithm to detect people. This figure shows the flow of program. 3) If the violation count is higher an alert system is produced to maintain the social distancing. This frequency will be changed with respective to violation count. Eg: If the violation is 2 a lesser frequency sound will produce if the violation is 20 a higher frequency sound will produce. The data will be stored in the database for future use. The information contains Violation count date Place, time to identify the location. The sound pronunciation is shown in Fig. 5. 4) Of all of the approaches proposed in the literature, the YOLOV4 appears to the most promising for the counting of people. Since YOLOV4 is fast and accurate object detection algorithm. The YOLOV4 approach allows both finding the people and also identification/extraction of the pixels associated with each individual person. The proposed social distancing finding System, Video analysis method is used for social distancing Violations. While the existing method requires hardware like drones, sensors, proposed method does not require any hardware since the algorithm can able to process the video that is captured by camera. YOLOV3 though correctly identifies and counts the people with better accuracy it is comparatively slower than YOLOV4.

Detection of Social Distance and Intimation System for Covid-19

265

5) Storing the data in database may useful in future for other works. With the help of mysql connector package and Xampp a local server is turned in the system and the violation are transferred to the database. It includes insert query to insert the values into the database. The database may take average after one hour and store it in another table (Fig. 4).

Fig. 3. Flow of an algorithm.

Fig. 4. The sound pronunciation when the violation is greater than two.

Fig. 5. Database insertion

266

S. Anandamurugan et al.

5 Methodology A) YOLOv4 Architecture and Functioning YOLOV4 is an object detection algorithm and it was found a long years ago. There are multiple objects can be found by YOLOV4 algorithm. The YOLOV4 algorithm detects not only the person it also detects the object like car, bus, and other eighty objects. The architecture of YOLOV4 works based on the ROI in which it converts the image into a layer and then it does a preprocessing step to find the particular objects. The function of this algorithm differs from object to object. Incase if the algorithm detects a person, it is a neural network to find the people. The YOLOV4 gives high accuracy if its work with the TPU based systems. YOLOV4 is the best algorithm when compared to another detecting algorithm. The advantages of this algorithm are trained only once to have a faster execution rate B) Social Distancing Violation The violation is founded by the real time access of camera and violation are given in the monitor screen with the help of Euclidean formula to find the distance between two frames are founded if it is less than 6 feet a red frame occurs for violation. It includes python package like NumPy, SciPy, Sixer package. The green color frame occurs those who not violate It also alerts sound violation and database connectivity. The database attributes include date, place, time, Violation Count at last we can find the violation in that particular place for that particular time.

C) Results This table shows the violation occurs if the distance is less than 6 feet and doesn’t show violation if it is greater than six. The violation and alert sound change of frame color change is obtained when the distance is greater than six feet. The red color frame is obtained when the distance is greater than six feet. This color frame will obtain for all the person in the crowd those who violate it [3] (Table 1). The above images show that the testing of social distancing using real time live detection using camera. The principle behind this distance between two frames are calculated every minute. When the person comes to the frames also moves to that person to find the distance between two frames. Once the minimum distance is less a violation with red color frame with alert sound [4] (Fig. 6).

Detection of Social Distance and Intimation System for Covid-19

267

Table 1. Violation counting Label

Distance

Prediction

Output 1

Less than 6 feet

Violated

Output 2

Greater than 6 feet

Not-violated

Fig. 6. Social distancing testing. Camera height: 3 m. Precision rate: 15 frames per second.

Table 2. Social distancing testing Label

Distance

Prediction

a

6.6 feet (250-pixel pitch)

Not-violated

b

6.3 feet (260-pixel pitch)

Not-violated

c

6.1 feet (265-pixel pitch)

Not-violated

d

5.8 feet (270-pixel pitch)

Violated

The above Table 2 shows that the testing along with pixel pitch are counted to find the social distancing violations. If the two frames are above six feet it shows green frame when it is lesser than six feet it shows red color frame for both the person and the violation count are also increased to two. Violation will be more in particular period. Critical density is defined as number of violations divided by particular place. Critical density can be found by having the violation count date, place and time. If the Critical density is high there may be high possibilities of spread of Covid-19. By Critical density value it is easy to find the spread of Covid-19 in that particular place. The proposed method can identify the person and find the social distance with the help of Euclidian formula. In order to improve the functionality and flexibility we added alert system and information transfer to the database for future use. Our Project will give an accuracy of 85% and the existing project gives about 80% accuracy. We use YOLOV4 algorithm to improve the accuracy. The violation count may differ from time to time. The violation count will be higher from morning period and after that the violation count will be less up to evening. After that the violation count may high. This is due to people move from one place to another place is high in that

268

S. Anandamurugan et al.

particular time. This is the graph shows that violation is higher in the peak period and lesser in the remaining period (Fig. 7).

Fig. 7. Data insertion with date, place, violations.

This diagram shows that violation along with date, place is stored into the database. This data helpful to find the critical density with the help of these we can find the possibilities of spread of Corona in that particular for the given time and place. This data also may take a decision to protect the area to prevent the spread of Covid-19 [5].

6 Conclusion and Future Work This project can be completed with low cost and it has many advantages when compared to existing solution. We had included live detection alert system and information transfer. In future it should be conclude more accurately with all places and to find the average of violation in a particular hour. The future work also include it also takes the person who is violating more. In this project we used only YOLOV4 algorithm and Euclidean formula to find the distance between two frames. The alerts frequency sound may differ violation count. The database also takes the average value violation for an hour and store it in another table. This project also helps to find the spread of Covid-19 in the particular place. This data may help the government to take the necessary steps in that pandemic area. With the

Detection of Social Distance and Intimation System for Covid-19

269

help of these violation count people may alert to follow the social distancing from the next day onwards. This project helps other to have the data about the social distancing violations.

References 1. Courtemanche, C., Garuccio, J., Le, A., Pinkston, J., Yelowitz, A.: “Strong social distancing measures in the United States reduced. The COVID-19 growth rate” study evaluates the impact of social distancing measures on the growth rate of confirmed COVID-19 cases across the United States. Health. Aff. 39, 1237–1246 (2020) 2. Nguyen, C.T., et al.: Enabling and Emerging Technologies for Social Distancing: A Comprehensive Survey (2020). arXive 3. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 8759–8768 (2018) 4. Zheng, Z., Wang, P., Liu, W., Li, J., Ye, R., Ren, D.: Distance-IoU loss: faster and better learning for bounding box regression. Proc. AAAI Conf. Artif. Intell. 34(07), 12993–13000 (2020). https://doi.org/10.1609/aaai.v34i07.6999 5. The Math of camera Field of View Calculations (FOV): Scantips.com (2020). https://www.sca ntips.com/lights/fieldofviewmath.html

Automatic Shoe Detection Using Image Processing K. R. Prasanna Kumar(B) , D. Pravin(B) , N. Rokith Dhayal(B) , and S. Sathya(B) Department of Information Technology, Kongu Engineering College, Perundurai, Tamil Nadu, India

Abstract. Image processing is a technique for applying operations on an image in order to improve it or extract relevant information from it. Object detection is a computer vision approach for identifying and locating things in images and videos. Deep learning has showed enormous promise in a wide range of realworld applications. Recent object detection based on image processing and deep learning models has yielded promising results in terms of object detection in images. Educational institutes and industries are insisting students and employees to wear shoes to enhance their safety and as well as for the professional appearance. It is observed that some of the students or employees are not wearing shoes which may lead to some serious injuries at labs and work places . To avoid this institutions and industries are involving with the physical verification process. The physical verification process is overhead to them. The intension of the proposed work is to automate the shoe detecting process using image processing and deep learning techniques. A shoe detection dataset consists of with shoe images which has been used to train model. The trained model is used to identify whether the person is wearing shoes or not in real time. If the person is not wearing shoe then the appropriate notification will be given. This proposed work will be helpful to the industries or institutions to ensure the safety measures. Keywords: Shoe detection · Image processing · Deep learning

1 Introduction Object detection is the process of locating semantic instances of specific category like human, buildings, vehicles and so on in digital photographs and videos. Face detection and pedestrian detection are two well-studied object detection areas. Object detection may be used in a variety of computer vision applications, such as retrieval of images and in surveillance cameras. Having a professional appearance is especially important in numerous emerging countries working cultures. Shoes are quite vital for a professional appearance. As a result, students in high school and college prepare to present themselves professionally. Because some students do not follow directions, it is vital to keep an eye on them in order to preserve a professional image. Similarly in industries are asked to wear shoes to ensure their safety. A person is currently assigned to keep an eye on the students © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 270–277, 2022. https://doi.org/10.1007/978-3-030-96299-9_26

Automatic Shoe Detection Using Image Processing

271

or employees [8]. It’s difficult to keep track of all of the students or employees. Using Python, OpenCV, Tensor Flow, and Keras, the proposed work will develop a shoe detector with computer vision using the dataset. For feature extraction, deep transfer learning with three traditional machine learning techniques are integrated [7]. The proposed model can be combined with scrutiny cameras to prevent injuries in the workplace.

2 Related Work Face detection models using several algorithms and techniques to detect the face mask. The SSDMNV2 method use the Single Shot Multi Box Detector and the MobilenetV2 architecture as a framework for the classifier, which is relatively light and may even be utilized in embedded devices (such as the NVIDIA Jetson Nano or the Raspberry Pi) to conduct real-time mask detection. The technique we used yielded a 0.9264 accuracy score and a 0.93 F1 score. The dataset were manually cleaned to improve the accuracy of the findings, the problem of numerous incorrect predictions was successfully eliminated from the model. The SSDMNV2 model should, ideally, aid the concerned authorities in dealing with this massive pandemic that has spread throughout much of the globe [6]. A machine learning-based face mask identification technique. Since 2019, the COVID-19 epidemic has spread rapidly over the world. Human existence is becoming increasingly involutes and complex as a result of this pandemic. Due to various typical mistakes made by humans this disease is spreading rapidly. So, the mask detector is presented based on the machine learning facial categorization algorithm. It is used to assess whether a person is wearing a mask or not. It is used to discern a person with a mask or a covered face. According to the findings, the ML-based topology produces better results with more precision and is more successful in suppressing the COVID-19 pandemic [2]. Existing detection methods, on the other hand, have a number of flaws, such as the inability to track specific motorcyclists across many frames or to distinguish drivers from passengers when wearing helmets. It proposes a CNN-based multi-task learning (MTL) system for identifying and monitoring particular motorcycles, as well as registering rider helmet wear. It demonstrates that using MTL for concurrent visual similarity learning and helmet use categorization improves the efficiency of our technique when compared to previous studies, allowing for a processing speed of more than 8 frames per second on consumer hardware [5]. It is made up of two parts. The first component is based on the ResNet-50 deep transfer learning model and is geared for feature extraction. The second component is based on YOLO v2 and is meant to detect medical face masks. As a detector, the adam optimizer earned the highest average precision and proved it is the effective model for detecting the medical face mask [4]. To achieve short inference time and high accuracy, the proposed technique uses an ensemble of one-stage and two-stage detectors. It starts with ResNet50 as a baseline and uses the transfer learning principle to combine high-level semantic information across many feature maps [1]. The suggested approach accurately detects the face from the image and then determines whether or not it has a mask on it. It can also detect a face and a mask in motion as a

272

K. R. Prasanna Kumar et al.

surveillance task performance. On two separate datasets, the technique achieves accuracy of up to 95.77% and 94.58%, respectively. It uses the Sequential Convolutional Neural Network model to investigate optimum parameter values in order to correctly detect the existence of masks without generating over-fitting. The model should be enhanced further to recognize whether the mask is virus-prone or not, i.e. whether it is surgical, N95, or not [9]. A Convolutional Neural Network (CNN)-based architecture for identifying cases of face mask misuse. This system, which is compatible with CCTV cameras, employs a two-stage CNN architecture that can recognize both masked and unmasked faces. After comparing its results with other models such as D-lib and MTCNN, the first stage includes a pre-trained Retina Face model for stable face identification. There were no biased photos in the collection of masked and unmasked faces. In addition, a technology known as Centroid Tracking was incorporated to our system, which improved overall stability and video data outcomes [3].

3 Proposed Methodology The first step in predicting whether a person is wearing shoes or not is to train the algorithm with a suitable dataset. The dataset for the model is ready, and it includes shoe photos. For the train and test CSV files, create the TFrecord. For the dataset, create a label map. When a class name is not specified in the annotation file, a label map is utilized to find the name of the class. The label map is a separate source of record for class annotations. An accurate detection model for shoes is required for training the classifier, so that the EfficientDet Object detection can classify whether the person is wearing shoe. The training procedure will begin when the record files have been created. The model will create an inference graph at the end of the training. The flow diagram in Fig. 1 depicts the technique utilized in this work.

Fig. 1. Flow diagram

3.1 Optimization Because it necessitates a considerable amount of computer power, training a deep neural network is a time-consuming and expensive operation. To train the network more quickly and efficiently, the deep learning-based transfer learning method is applied. The taught

Automatic Shoe Detection Using Image Processing

273

knowledge in terms of parametric weights from the neural network may be transferred to a new model by transfer learning [1]. The new model improves performance even when just trained on a small dataset. AlexNet, MobileNet, ResNet50, and other pre-trained models were all trained using 14 million images from the ImageNet dataset. The SSD is treated as a pre-trained model in the recommended technique for easy shoe classification [1]. 3.2 Pre-processing Because the quality of the model trained on a decent dataset is determined by it, the data were collected manually by collecting real-time photographs. They were then analyzed, and all repeats were manually eliminated. To eliminate the faulty photos, the data cleaning was done manually. Finding these photographs was a crucial aspect of the process. As is widely known, cleaning a corrupt image was a difficult process, but thanks to commendable efforts, the work was split and the data set was cleaned with shoe pictures. Any prediction model may be improved by cleaning, detecting, and fixing flaws in a dataset. This section explains how to prepare data for training [6]. Pre-processing is a function in the dataset that receives a folder as input, loads all of the pictures from the folder, and resizes the photographs using the SSD EfficientDet model. After that, the photographs are converted into tensors and the list is sorted alphabetically. To make the procedure go faster, the list is converted to a NumPy array. When that, after the model has been trained, data augmentation is utilized to increase its accuracy [6]. 3.3 Data Augmentation Due to the lack of a suitable amount of data for training the suggested model, an immense amount of data is required for effective training of the SSD EfficientDet model. This problem is solved using the data augmentation approach. Methods like as rotating, zooming, shifting, shearing, and flipping the image are employed in this approach to create several variations of a similar image. For the data augmentation procedure in the suggested approach, picture augmentation is employed. For image augmentation, a function image data generation is constructed, which returns test and train batches of data. 3.4 Model Training The categorization challenge was solved with the help of a deep neural network. Corrected linear measure (ReLU) and GHB Pooling layers follow the main Convolution layer. After creating the blueprint for data analysis, the model must be trained using a specific dataset before being compared to another dataset. When developing a prediction, a correct model and optimum train test split help to offer accurate results. The test size is set as 0.1, which means that 90% of the dataset will be coached and the remaining 10% will be used for testing. Mistreatment Model Checkpoint is used to track validity loss. The images from the training and test sets are then fitted to the subsequent model. In this case, 20% of the coaching data is used as validation data. The model is trained for a total of twenty epochs (iterations) to maintain a balance between accuracy and likelihood Fig. 2 shows a visual representation of the intended model. The foundation layers

274

K. R. Prasanna Kumar et al.

are then frozen to prevent the loss of previously learnt characteristics. Then additional trainable layers are added, and these layers are trained on the acquired dataset to find the attributes that may be used to distinguish a shoe-wearing foot from a shoe-less foot. The pipeline of the training process using pre-trained model is shown in the Fig. 2. The trained model is saved for future purpose as well as the weight values are also saved. It is observed that the pre-trained models are rapidly reducing the unnecessary computational costs.

Fig. 2. Pipeline of using pretrained model

4 Discussion Images and real-time video feeds were used to evaluate the models. The model’s correctness is obtained, and the model’s optimization is a continuous process that results in a highly accurate answer by setting the hyperparameters. The loss is defined by how well the model performs in these two sets of data, and it is calculated using training and validation data. The mistake in the training set of data is known as training loss. Validation loss is the error that occurs after the validation batch of data has been sent through the trained network. Training accuracy refers to the trained model’s capacity to recognize independent pictures that were not used in training, whereas test accuracy refers to the trained model’s ability to recognize images that were not utilized in training. Figure 3 and Fig. 4 depict the loss in training and validation, respectively, whereas Fig. 5 and Fig. 6 depict the accuracy in training and validation. After each optimization iteration, the loss value reveals how well or poorly a model performs. Adding more photos to the training dataset and changing the batch size for training can improve the model’s accuracy even further. The use of high-performance personal GPUs might aid the training process. This model might be used as an example of edge analytics in action. Furthermore, using a shoe dataset, the suggested technique delivers state-of-the-art results. The system can recognize if the user is wearing shoes thanks to the invention of shoe detection.

Automatic Shoe Detection Using Image Processing

Training Loss 1 0.8

Loss

0.6 train_loss 0.4 0.2 0 0.0

2.5

5.0

7.5

10.0 12.5 15.0 17.5

Fig. 3. Training loss

Validaon Loss 1 0.8

Loss

0.6 Val_Loss 0.4 0.2 0 0.0

2.5

5.0

7.5

10.0 12.5

15.0

Fig. 4. Validation loss

17.5

275

276

K. R. Prasanna Kumar et al.

Training Accuracy 1.2 1

Accuracy

0.8 0.6 0.4 0.2 0 0.0

2.5

5.0

7.5

10.0

12.5 15.0 17.5

Fig. 5. Training accuracy

Validaon Accuracy 1

Accuracy

0.8

0.6 val_acc 0.4

0.2

0 0.0

2.5 5.0

7.5 10.0

12.5

15.0

Fig. 6. Validation accuracy

17.5

Automatic Shoe Detection Using Image Processing

277

5 Conclusion At initially, the suggested system included a brief explanation of the work’s rationale. The real time shoe detection was performed by using live camera source. It showed a rapid performance. This system depicts the model’s learning and performance task. The method has attained a reasonable level of accuracy using basic machine learning tools and simplified methodologies. It may be used for a wide range of purposes. The proposed approach will make a significant contribution to the education institutions or industries. It might be extended in the future to identify industry-specific footwear to meet industry standards.

References 1. Sethi, S., Kathuria, M., Kaushik, T.: Face mask detection using deep learning: An approach to reduce risk of Coronavirus spread. Elsevier Inc. (2021) https://doi.org/10.1016/j.jbi.2021. 103848 2. Gupta, S., Sreenivasu, S.V.N., Chouhan, K., Shrivastava, A., Sahu, B., Potdar, R.M.: Novel face mask detection technique using machine learning to control COVID’19 pandemic. Elsevier Ltd. (2021) https://doi.org/10.1016/j.matpr.2021.07.368 3. Chavda, A., Dsouza, J., Badgujar, S., Damani, A.: Multi-stage CNN architecture for face mask. In: 2021 6th International Conference for Convergence in Technology (I2CT), IEEE (2021) 4. Loey, M., Manogaran, G., Taha, M.H.N., Khalifa, N.E.M.: Fighting against COVID-19: A novel deep learning model based on YOLO-v2 with ResNet-50 for medical face mask detection. Elsevier Ltd. (2020) https://doi.org/10.1016/j.measurement.2020.108288 5. Lin, H., Deng, J.D. (Member, IEEE), Albers, D., Siebert, F.W.: Helmet use detection of tracked motorcycles using CNN-based multi-task learning. In: IEEE, vol. 8 (2020) 6. Nagrath, P., Jain, R., Madan, A., Arora, R., Kataria, P., Hemanth, J.: SSDMNV2: A real time DNN-based face mask detection system using single shot multi-box detector and MobileNetV2. Elsevier Ltd. (2020) https://doi.org/10.1016/j.scs.2020.102692 7. Logeswaran, K., Suresh, P., Ponselvakumar, A.P., Renuga, G., Priyadharshini, M., Nivetha, R.: Contemplate study of contemporary evolutionary approaches used for mining high utility item set. Int. J. Adv. Sci. Technol. 29(5), 2599–2607 (2020) 8. Prasanna Kumar, K.R., Kousalya, K.: Amelioration of task scheduling in cloud computing using crow search algorithm. Neural Comput. Appl. 32(10), 5901–5907 (2019) 9. Das, A., Wasif, M., Rohini Basak, A.: Covid-19 face mask detection using TensorFlow, Keras and OpenCV. In: IEEE 17th India Council International Conference (INDICON) (2020)

Recognition of Disparaging Phrases in Social Media K. R. Prasanna Kumar(B) , P. Aswanth(B) , A. Athithya(B) , and T. Gopika(B) Department of Information Technology, Kongu Engineering College, Perundurai, Tamil Nadu, India

Abstract. Social media provides a space for individuals, where they can share their view or opinion very easily. The use of social media is growing at a rapid pace. Users can communicate more swiftly through social media platforms like Twitter, Facebook, and YouTube. The content of social media texts is generally made up of code-mixed comments/posts and replies, and it may contain harmful or nonoffensive words. Sentiment analysis on social media provides businesses with a quick and effective approach to monitor public opinion on their brand, business, goods, and other topics. In recent years, a variety of features and approaches for training sentiment classifiers for fetched datasets have been investigated, with mixed results. Twitter is a popular social media platform. It provides businesses with a quick and efficient way to assess customers’ viewpoints on issues that are crucial to their performance in the marketplace. Expanding a sentiment analysis software is a way to utilise computers to measure consumer perceptions. Analyze the sentiment from employee tweets regarding work from home is the main goal. The employees work from home tweets dataset as input was collected from twitter. Then, to analyse the sentiment the NLP techniques, text classification and deep learning algorithm were used. The experimental results shows the performance metrics such as accuracy and analyse the sentiment based on sentiment analyser into positive, negative and neutral. Keyword: Social media · Sentiment analysis · Long Short Term Memory · NLP techniques · Tf-idf Vectorizer

1 Introduction Social media plays a vital role to interact with people all over the world. The impact of societal changes is bending in the direction of people’s expressed thoughts on social media. Some of the well-known social media platform are Facebook, Instagram and Twitter so on. Twitter, which has 313 million users worldwide and over 1.4 billion active tweets posts, has become a great resource for organisations and people with a solid social, democratic, or financial stake in order to maintain and growing their moral authority and notoriety [21]. At present days knowing individual/public opinion has more impact on the business and public domain activities. Industries & political part are showing interest in knowing the public opinion & emotion through social media. So, exploring necessary information from social media platform become a trust area. These kind of analyse © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 278–285, 2022. https://doi.org/10.1007/978-3-030-96299-9_27

Recognition of Disparaging Phrases in Social Media

279

where cannot out for different purposes likes improve the product quality, reputation of company to known pros & cons and so on. There are good number of examples where social media information become a highlight news and bring change in the society [7]. Sentiment analysis, sometimes called sentiment AI, is the systematic observation, extraction, decision-making, and assessment of emotional states and subjective information employing Natural Language Processing, text mining, computational linguistics, and bio metry [15, 19]. These firms can monitor multiple social media sites in real time using sentiment analysis. The approach of automatically recognising emotional or opinionated information in a text segment, as well as assessing the text’s polarity [4], is known as sentiment analysis. The purpose of Twitter sentiment classification is to assign a positive, negative, or neutral sentiment polarity to a tweet. Tweets frequently contain incomplete, noisy, and badly structured sentences, irregular phrasing, ill-formed words, and non-dictionary terms [19]. Before selecting features, a number of pre-processing processes are done to decrease the amount of noise in the tweets. In tweets, many sorts of negation may be observed. Negation, in general, is critical in determining a tweet’s mood [14]. The negation process transforms “won’t,” “can’t,” and “not” into “will not,” “cannot,” and “not,” respectively. The majority of experts say that URLs don’t reveal anything about a tweet’s tone. Here [19], Twitter’s short URLs are transformed to URLs and tokenized [28]. The token-matching URLs are then removed from tweets, allowing the message’s text to be improved. Three letters differentiate words like “cool” from “cooooool.” Numbers are being deleted from the equation [24]. Stop words like “the,” “is,” and “at” frequently relate to a language’s most common phrases. Most researchers remove stop words before selecting features because they are thought to be harmful to the purpose of sentiment classification [6, 28]. The standard approach of eliminating stop words is based on precompiled lists. In general, sentiment analysis aims to detect a speaker’s, essayist’s, or other subject’s disposition in terms of theme by noticing strong emotional or passionate responses to an archive, communication, or incident [15]. The disposition might be an emotional judgement or evaluation (i.e., the creator’s or speaker’s passionate state) or an anticipation of enthusiastic responses (i.e., the creator’s or buyer’s intended impact) [19]. Client surveys or suggestions on a broad range of issues are now freely available on the Internet, and audits may include surveys on themes like clients or film fault-finding [24, 27]. Because individuals desire to express themselves online, surveys are becoming increasingly popular. In document-level sentiment classification, a document can be completely categorised as “positive,” “negative,” or “neutral.” At the sentence level, each sentence is graded as “positive,” “negative,” or “unbiased” [27]. At the aspect and feature levels, sentiment classification [14]: At this level, depending on specific features of the sentences archives, sentences/documents can be classed as “positive,” “negative,” or “non-partisan” [28]. This is referred to as “perspective-level evaluation grouping.”

280

K. R. Prasanna Kumar et al.

2 Proposed Methodology Deep learning algorithms such as LSTM and others are employed. Long short-term memory is a sort of deep learning artificial recurrent neural network. LSTM features feedback connections, unlike standard feed forward neural networks. It can handle complete data sequences as well as single data points. LSTM networks are well-suited to classifying, assessing, and generating predictions based on time series data [6] because major occurrences in a time series may have unanticipated delays. The data is retrieved from the UCI repository, loaded as training and testing, and then visualized [24]. For data preprocessing, nltk is employed. Remove urls, html tags, digits, hashtags, mentions, and stop words from the dataset, then use Tf-idf Vectorizer & count vectors to tokenize, label encode, and extract features before splitting the train and test vectors [8]. Finally, filter the tweets into three categories: good, negative, and neutral, and evaluate the findings based on precision, accuracy, f1-score and the recall. The approach used in this paper is depicted in the flow diagram in Fig. 1.

Fig. 1. Flow diagram

Recognition of Disparaging Phrases in Social Media

281

2.1 Data Selection Work from home employees tweet dataset is collected from twitter. The data selection process is the process of assessing the sentiment and categorising it as positive, neutral, or favourable for work from home employees [8]. 2.2 Data Preprocessing Data pre-processing is the process of deleting undesired data as well as missing data from a dataset [6]. The dataset is transformed into a structure appropriate for machine learning using pre-processing data transformation methods. It also entails cleaning the dataset by deleting extraneous or corrupted data that could impair the dataset’s correctness, making it more efficient. The great majority of machine learning methods need numerical input and output variables. 2.3 NLP Techniques NLP is a branch of machine learning that involves a computer’s capacity to comprehend, interpret, modify, and perhaps synthesise human language [16]. Cleaning the data typically consists of a number of steps: 1. 2. 3. 4.

Remove any punctuation marks. Tokenization. Stemming. Sentiment analysis.

Remove Punctuation: Punctuation can help us grasp a statement by providing grammatical context. Tokenization: Tokenization is the process of breaking down text into smaller parts, such as phrases or words. It provides previously unstructured text structure. Plata o Plomo, for example, is ‘Plata’,‘o’,‘Plomo’. Stemming: This technique aids in the reduction of a word to its stem form. Sentiment Analysis: Using the sentiment analyser, we may categorise the sentiment as positive, neutral, or negative (polarity score) [15]. Sentiment analysis breaks down a communication into subject chunks and assigns a sentiment score to each topic.

2.4 Data Splitting Data is required during this phase in order for learning to take place. In addition to the data necessary for training, test data is required to assess the algorithm’s performance and determine its effectiveness. In this method, 70% of the twitter dataset is used as training data, while the remaining 30% is used as testing data. Data splitting is the process of dividing accessible data into two halves for cross-validation reasons [16]. One part of the

282

K. R. Prasanna Kumar et al.

data is used to create a predictive model, while the other is utilised to assess the model’s performance. Part of analysing data mining models is separating data into training and testing sets. When a data set is split into training and testing sets, the bulk of the data is utilised for training and just a small portion for testing. 2.5 Classification Proposed system uses deep learning algorithms, such as LSTM, in our method. Long Short-Term Memory (LSTM) is a deep learning architecture that employs a recurrent neural network design that is artificial. Unlike simple feed forward neural networks, LSTM has feedback connections and can handle huge data sequences as well as single data points. Because significant occurrences in a time series might have unpredictable delays, LSTM networks are well-suited to categorising, analysing, and producing predictions based on time series data.

3 Discusssion The final result will get generated based on the overall classification and prediction. The performance of this proposed approach is evaluated using some measures, Once the model is trained using training dataset it is evaluated by test dataset [19]. The metrics used to evaluate the model are accuracy, confusion matrix, F1 score, precision, recall [8]. The Table 1 shows the confusion matrix calculation. Table 1. Confusion matrix Actual positive

Actual negative

Predicted positive

TrPo

FP

Predicted negative

FaNe

TrNe

TrPo – True Positive, FaPo – False Positive. FaPo – False Positive, FaNe – False Negative.

The Confusion matrix is the representation of actual and predicted results. True Positive, True Negative are the results that model predictions and actual results are same. False Positive and False Negative are the wrong prediction results. F1 score, precision, recall and accuracy are the prediction based metrices. These metrices are shown in Eqs. (1)–(4) Accuracy = TrPo/(TrPo + TrNe) ∗ 100

(1)

Recognition of Disparaging Phrases in Social Media

283

Precision = TrPo/(TrPo + FaPo)

(2)

Recall = TrPo / (TrPo + FaNe)

(3)

F1 Score = 2 ∗ (Precision ∗ recall/precision + recall)

(4)

Precision is used to determine the number of positive class predictions that are actually positive class predictions. Recall is the number of positive class predictions produced from all positive cases in the dataset. The F1 score is a single value that combines accuracy and recall into one.

Fig. 2. Sentiment analysis

The Fig. 2 depicts the sentiment analysis as classifying the phrase as positive, negative and neutral sentiment. By this analysis, the work from home dataset has positive tweets. So, the proposed system can be conclude as people are more interested in work from home. In Fig. 3 due to increase in precision accuracy is increased. So, that in proposed system accuracy level is increased to the average of 80–85%.

284

K. R. Prasanna Kumar et al.

Fig. 3. Performance analysis

4 Conclusion The proposed system conclude that, the employees twitter dataset was taken as input. It implements the NLP techniques and deep learning algorithm such as Long short term memory (LSTM). Then visualize the sentiment into positive, neutral and negative in the form graph. In future, it is possible to provide extensions or modifications to the proposed clustering and classification algorithms to achieve further increased performance. Apart from the experimented combination of data mining techniques, further combinations and other clustering algorithms can be used to improve the detection accuracy. Finally, the sentiment analysis detection system can be extended as a prevention system to enhance the performance of the system.

References 1. Carreras, X., Màrquez, L.: Boosting trees for anti-spam email filtering. arXiv:cs/0109015. https://arxiv.org/abs/cs/0109015 (2001) 2. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent Dirichlet allocation. J. Mach. Learn. Res. 3, 993– 1022 (2003) 3. Hammouda, K.M., Kamel, M.S.: Efficient phrase-based document indexing for Web document clustering. IEEE Trans. Knowl. Data Eng. 16(10), 1279–1296 (2004) 4. Das, S.R., Chen, M.Y.: Yahoo! for Amazon: Sentiment extraction from small talk on the web. Manage. Sci. 53(9), 1375–1388 (2007) 5. Factiva, D.: Quick Study: Direct Correction Established Between Social Media Engagement and Strong Financial Performance. PR News (2009) 6. Gupta, V., Lehal, G.S.: A survey of text mining techniques and applications. J. Emerg. Technol. Web Intell. 1(1), 60–76 (2009) 7. Kaplan, A.M., Haenlein, M.: Users of the world, unite! The challenges and opportunities of Social Media. Bus. Horiz. 53(1), 59–68 (2010)

Recognition of Disparaging Phrases in Social Media

285

8. Tumasjan, A., Sprenger, T.O., Sandner, P.G., Welpe, I.M.: Predicting elections with twitter: What 140 characters reveal about political sentiment. Icwsm 10(1), 178–185 (2010) 9. DuVander, A.: Which APIs are handling billions of requests per day? Programmable Web (2012) 10. Aggarwal, C.C., Reddy, C.K.: Data Clustering: Algorithms and Applications. CRC Press, Boca Raton, FL, USA (2013) 11. Sharma, R., Nigam, S., Jain, R.: Opinion mining of movie reviews at document level. arXiv preprint arXiv:1408.3829 (2014) 12. Sharma, R., Nigam, S., Jain, R.: Polarity detection at sentence level. Int. J. Comput. Appl. 86(11), (2014) 13. Chae, B.K.: Insights from hashtag #supplychain and Twitter analytics: Considering Twitter and Twitter data for supply chain practice and research. Int. J. Prod. Econ. 165, 247–259 (2015) 14. Giachanou, A., Crestani, F.: Like It or Not: A Survey of Twitter Sentiment Analysis Methods. ACM Comput. Surv. 49(2), 1–41 (2016) 15. Abirami, A., Gayathri, V.: A survey on sentiment analysis methods and approach. In: Advanced Computing (ICoAC), 2016 Eighth International Conference on, pp. 72–76, IEEE (2017) 16. Ahmad, N., Siddique, J.: Personality assessment using Twitter tweets. Procedia Comput. Sci. 112, 1964–1973 (2017) 17. Carvalho, J.P., Rosa, H., Brogueira, G., Batista, F.: MISNIS: An intelligent platform for Twitter topic mining. Expert Syst. Appl. 89, 374–388 (2017) 18. Carducci, G., Rizzo, G., Monti, D., Palumbo, E., Morisio, M.: TwitPersonality: Computing personality traits from tweets using word embeddings and supervised learning. Information 9(5), 127 (2018) 19. Ahmad, T., Ramsay, A., Ahmed, H.: Detecting emotions in English and Arabic tweets. Information 10(3), 98 (2019) 20. Devlin, J., Chang, M.-W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidirectional transformers for language understanding. In: Proc. Conf. North Amer. Chapter Assoc. Comput. Linguistics, Human Lang. Technol., vol. 1, pp. 4171–4186. Association for Computational Linguistics, Minneapolis, MN, USA, June 2019 21. Fung, I., et al.: Pedagogical demonstration of Twitter data analysis: A case study of world AIDS day, 2014. Data 4(2), 84 (2019) 22. Boldog, P., Tekeli, T., Vizi, Z., Dénes, A., Bartha, F.A., Röst, G.: Risk assessment of novel coronavirus COVID-19 outbreaks outside China. J. Clin. Med. 9(2), 571 (2020) 23. Bhat, R., Singh, V.K., Naik, N., Kamath, C.R., Mulimani, P., Kulkarni, N.: COVID 2019 outbreak: The disappointment in Indian teachers. Asian J. Psychiatry 50, (2020) Art. no. 102047 24. Han, X., Wang, J., Zhang, M., Wang, X.: Using social media to mine and analyze public opinion related to COVID-19 in China. Int. J. Environ. Res. Public Health 17(8), 2788 (2020) 25. Depoux, A., Martin, S., Karafillakis, E., Preet, R., Wilder-Smith, A., Larson, H.: The pandemic of social media panic travels faster than the COVID-19 outbreak. J. Travel Med. 27(3), (2020). Art. no. taaa031 26. El Zowalaty, M.E., Järhult, J.D.: From SARS to COVID-19: A previously unknown SARSrelated coronavirus (SARS-CoV-2) of pandemic potential infecting humans—Call for a one health approach. One Health 9, (2020). Art. no. 100124 27. Logeswaran, K., Suresh, P., Ponselvakumar, A.P., Renuga, G., Priyadharshini, M., Nivetha, R.: Contemplate study of contemporary evolutionary approaches used for mining high utility item set. Int. J. Adv. Sci. Technol. 29(5), 2599–2607 (2020) 28. Prasanna Kumar, K.R., Kousalya, K.: Amelioration of task scheduling in cloud computing using crow search algorithm. Neural Comput. Appl. 32(10), 5901–5907 (2019)

Machine Learning Model for Identification of Covid-19 Future Forecasting N. Anitha(B) , C. Soundarajan(B) , V. Swathi(B) , and M. Tamilselvan(B) Department of Information Technology, Kongu Engineering College, Perundurai, Erode, India {anitha.it,soundarajanc.18it,swathiv.18it, tamilselvanm.18it}@kongu.edu

Abstract. The global proliferation of COVID-19 has put humanity in jeopardy. The assets of the world’s most powerful economies are at risk due to the disease’s high infectivity and contagiousness. The ability of ML algorithms to forecast the number of future patients COVID-19 has an effect on, which is currently regarded as a potential threat to humanity. In this study, five common guaging models, including LR, LASSO, SVM, ES, and LSTM, were used to assess the COVID-19 underpinning variables. Each model makes three types of predictions: the number of recently contaminated cases, the number of passings, and the number of recoveries. However, it is impossible to predict the exact prognosis for the patients. To combat the problem, a proposed technique based on the long transient memory (LSTM) predicts the number of COVID-19 cases in the following 10 days and the influence of preventive measures such as social seclusion and lockdown on COVID-19 spread.

1 Introduction 1.1 Overview of Covid-19 Coronavirus, a global pandemic, has revealed human culture’s vulnerability to extreme, irresistible infections, as well as the difficulty of addressing this problem in a globally interwoven, complicated framework. In just a few weeks, the Coronavirus spread to over 100 countries. As a result, the entire human species must work together to battle the epidemic, as well as rationally plan to return to work and production based on the actual situation in each location and a full topographical risk assessment. Many efforts have been made to find an appropriate and rapid method of identifying tainted patients in the early stages. After doing chest CT sweeps of 21 people infected with COVID19 in China, Guan et al. detected two-sided pneumonic parenchymal ground-glass and consolidative aspiratory opacities, sometimes with an adjusted morphology and a fringe lung conveyance. As a result, the COVID-19 conclusion can be treated as a picture division problem, allowing the disease’s basic ingredients to be eliminated. Coronavirus Illness 2019 (COVID-19), a disease caused by the new Covid, is quickly spreading over the world. It had infected over 1,436,000 persons in over 200 countries and domains as of April 9, 2020. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 286–295, 2022. https://doi.org/10.1007/978-3-030-96299-9_28

Machine Learning Model for Identification of Covid-19 Future Forecasting

287

Covid-19 infection is an infectious disease that affects the respiratory and circulatory systems and is caused by a severe respiratory illness called Covid 2 (SARS-CoV-2). Initially distinguished in Wuhan, China and at present it is a continuous pandemic condition. Symptoms include fever, hacking, exhaustion, breathing difficulty, and a loss of smell and taste. Symptoms arise one to fourteen days after exposure to the illness. While the great majority of people experience minor side effects, some people get acute bronchial pain (ARDS). This can be accelerated by cytokine storms, multi-organ disappointment, septic shock, and blood clumping. Long-term organ damage has been observed, and there is concern about countless patients who have recovered from the acute phase of the illness but continue to experience a variety of side effects—known as long COVID—for a long time afterward, including severe exhaustion, cognitive decline and other intellectual issues, second-rate fever, muscle weakness, and shortness of breath. 1.2 Exponential Smoothing Outstanding smoothing is a method for smoothing time series data using spectacular window work as a guideline. While past sensations are weighted similarly in the basic moving normal, exceptional capacities are used to assign radically decreasing loads with time. It is a practical, academic, and efficiently used methodology for obtaining some confidence based on the client’s prior assumptions, such as irregularity. For the analysis of time-series data, dramatic smoothing is routinely used. Dramatic smoothing is one of the window capabilities commonly used to smooth data in signal processing, acting as low-pass channels to remove high-recurrence turbulence. Poisson’s use of recursive dramatic window capabilities in convolutions from the nineteenth century, as well as Kolmogorov and Zurbenko’s use of recursive moving midpoints from their disturbance research, set the precedent for this method. There is no officially right system for picking {\display style \alpha}\alpha. Now and then the analyst’s judgment is utilized to pick a suitable factor. This is as opposed to a basic moving normal, where a few examples can be ommited without much loss of data because of the consistent weighing of tests inside the normal. If a known number of tests will be recalled favourably, a weighted normal can be changed for this as well, by assigning equivalent load to the new example and skipping each of the previous ones. A dramatically weighted moving normal is another name for this basic sort of dramatic smoothing (EWMA). It’s also known as an autoregressive incorporated moving normal (ARIMA) (0,1,1) model because it doesn’t have a constant term. 1.3 Future Forecasting Estimating is the process of predicting what will happen in the future based on data collected over a long period of time and, most commonly, pattern analysis. An ordinary model could be used to examine a specific factor at a future period. The phrase “expectation” is a comparative term that encompasses a wider range of concepts. Both may refer to formal factual procedures that make use of time series, longitudinal, or cross-sectional data, as well as reduced formal critical strategies. The terms “figure” and “determining” are sometimes used in hydrology to describe attributes at specific future events, whereas the phrase “prediction” is used in more general assessments, such as when when floods

288

N. Anitha et al.

are expected to occur over a long duration of time. Hazard and vulnerability are essential components of expecting and anticipating; it is generally seen as excellent practise to indicate the level of hazard and vulnerability. To ensure that the figure is as precise as possible, all of the data should be current. It’s not uncommon for the information used to forecast the parameter to be judged. Subjective judging processes are abstract and rely on the evaluation and judgement of purchasers and experts; they are useful when prior knowledge is unavailable. They’re frequently utilised in conjunction with transitional or long-reach choices. Educated assessment and judgement, the Delphi procedure, statistical surveying, and recorded life-cycle similarity are examples of subjective determining techniques. To estimate future information as a component of past data, quantitative determining models are used. When prior mathematical data is accessible and it’s acceptable to suppose that some of the examples in the data can be trusted to continue into the future, they’re appropriate to employ. Short- and mid-range range options are frequently used with these approaches. Quantitative deciding procedures include last period interest, basic and balanced N-period shifting midpoints, uncomplicated notable smoothing, poisson measure model based guaging, and multiplicative occasional lists. Different techniques can result in differing degrees of determining precision, according to previous study. In terms of guaging execution, GMDH neural organisation beat traditional anticipating computations such as Single Exponential Smooth, Double Exponential Smooth, ARIMA, and back-proliferation neural organisation. 1.4 Supervised Machine Learning Regulated learning is an AI challenge that entails developing a capacity that converts an input to a yield using model identify sets. It creates a capacity from indicated preparation data, which is made up of a set of prepping models. Every directed learning model is made up of a pair of an information object (usually a vector) and a target yield value. A training calculation that is administered deconstructs the training data and creates an induced capacity that may be utilised to plan new models. In an ideal circumstance, the formula will be used to precisely determine the class names for inconspicuous occurrences. This necessitates the use of calculation to sum up information from the preparation to hidden situations in a “reasonable” manner. Idea learning is a term used frequently in human and animal brain studies to describe the same task. To take care of the given issue of managed learning, one needs to play out the accompanying steps: Determine the sort of preparing models. Prior to doing whatever else, the client ought to choose what sort of information is to be utilized as a preparation set. On account of penmanship examination, for instance, this may be a solitary transcribed person, a whole manually written word, or a whole line of penmanship. Accumulate a preparation set. The preparation set should be illustrative of this present reality utilization of the capacity. In this way, a bunch of information objects is accumulated and comparing yields are likewise assembled, either from human specialists or from estimations. Decide the information highlight portrayal of the learned capacity. The precision of the learned capacity relies unequivocally upon how the information object is addressed. Commonly, the info object is changed into an element vector, which includes various provisions that are elucidating of the article. The quantity of provisions ought not be

Machine Learning Model for Identification of Covid-19 Future Forecasting

289

excessively huge, on account of the scourge of dimensionality; yet ought to contain sufficient data to precisely foresee the yield. Decide the construction of the learned capacity and relating learning calculation. For instance, the architect might decide to utilize support vector machines or choice trees. Complete the plan. Run the learning calculation on the assembled preparing set. Some directed learning calculations require the client to decide specific control boundaries. These boundaries might be changed by enhancing execution on a subset (called an approval set) of the preparation set, or through cross-approval. Assess the precision of the learned capacity. After boundary change and learning, the presentation of the subsequent capacity. 1.5 Related Work Coronavirus is as of now viewed as a possible risk to humankind. Support Vector Machine (SVM) was used to predict COVID-19 compromising elements in this evaluation using four typical expectation models, including straight relapse (passed on to right), essentially entire rundown, and pick administrator. Forecasts are made for every models, such as the amount of new diseases, the amount of passings, and the amount of repeats for a period of 10 days. In terms of the review’s implications, it shows that it is a promising tool for implementing these techniques in the current COVID 19 contaminated situation. Coronavirus doesn’t appear to influence kids harshly; numerous pediatrics wards have been centered more around the crisis of COVID-19-related issues. Consequently, consideration on numerous other intense and ongoing sicknesses, particularly those more extraordinary, might be inadequate. This shortage of interest might cause, especially in adolescence, extreme issues, or even passing. The motivation behind this article, according to Alaa A. R. Alsaeedy and Edwin K. P. Chong et al., is to familiarise another technique with recognising locations with high human population and mobility, which pose a threat to spread COVID-19. Highly populated regions with rapidly travelling people (known as red zones) are powerless to stop the disease from spreading, especially if they include people who are asymptomatic but are contaminated, as well as persons who are healthy. Techniques: Our strategy detects high-risk areas by exploiting current cell network features such as cell (re)selection and handover, which are utilised to maintain consistent inclusion for portable end-client hardware (UE). We take advantage of previously existing cell network functions designed to ensure end-client portability and consistency. Because nearly everyone carries a cell phone (also known as client gear or UE), these function as human trackers that are always on. To put it another way, the more UEs there are and the more versatile they are, the more people there are and the more portable they are. According to a new investigation, SARS-CoV-2 can be visible for up to three hours (lasting viable in vapour sprayers), and can be breathed out by infected people when chatting, hacking, or in any case, breathing, if suggestive. We are particularly concerned about situations in which infectious persons are present in areas where there are many additional persistently portable individuals. [1]. In this work, Richard f. Singe, Nicolas Velásquez, and others claim that a large amount of potentially dangerous COVID-19 misinformation is circulating on the internet. In this paper, we apply artificial intelligence to assess COVID-19 content among online opponents of foundation wellbeing management, specifically inoculations (“anti vax”).

290

N. Anitha et al.

We discovered that the anti-vaccine local area is promoting a less involved debate around COVID-19 than its counterpart, the pro-vaccine local area (“supportive of vax”). In any event, the counter vax local region displays a broader range of COVID-19 “flavours,” and so can speak to a broader cross-section of people looking for COVID-19 direction on the web, such as those concerned about a mandatory optimal COVID-19 immunisation or those seeking elective remedies. Thus the counter vax local area looks better situated to draw in new help going ahead than the supportive of vax local area. This is disturbing since an inescapable absence of reception of a COVID-19 immunization will mean the world misses the mark concerning giving group invulnerability, leaving nations open to future COVID-19 resurgences. We propose an intuitive model for deciphering these results, which could aid in determining the acceptable adequacy of intercession approaches. Our methodology is adaptable, allowing us to address the critical problem that web-based media foundations have in dismantling massive amounts of online health misinformation and disinformation [2]. In this study, Shaoping Hu, Yuan Gao, and others suggest Subsequently late December 2019, a new Covid illness (COVID-19) has flared up in Wuhan, China, and has since spread around the world. Despite the fact that COVID-19 is a highly treated disease, the beginning of genuine ailment might bring about death as an outcome of considerable alveolar harm and moderate respiratory disappointment. Despite the fact that research facility testing, e.g., utilizing reverse record polymerase chain response (RT-PCR), is the brilliant norm for clinical analysis, the tests might deliver bogus negatives. Besides, under the pandemic circumstance, lack of RT-PCR testing assets may likewise defer the accompanying clinical choice and treatment. Under such conditions, chest CT imaging has turned into a significant device for both determination and guess of COVID-19 patients. We suggest a pitifully directed profound learning methodology for identifying and organising COVID-19 disease from CT images in this review. The suggested technique can reduce the need for manual CT image tagging while still allowing for precise contamination detection and differentiation of COVID-19 from non-COVID-19 cases. [3]. In this work, Yan Zhang, Yingbing L, and colleagues claim that Corona Virus Disease 2019 (COVID-19) cases in Wuhan were eliminated, and the plague situation was well under control. Such a contagious public health virus has a great deal of strain on the public economy. A few nations and districts around the world are still plague-stricken, and there is a pressing need to assess the disease situation and travel risk in the area. In general, a fine scale down to see the overall situation, and then intelligent drafting selections to speed up the return to invention and labour. Using multiple sources of information, points for assessing the COVID-19 plague were created in this study. Using the GeoDetector model and the decision tree model, a computational assessment of 736 fine-grained lattices was carried out. The investigation discovered that the risk level in more experienced areas was much higher than in newer areas; population thickness was the main determinant of disease; the number of metropolitan people dwindled to 37 percent of what it was in normal times after the “city conclusion,” according to Tencent data; the model this paper used depicts the central point in characterising okay regions and high-hazard regions, and proposes ideas and appraisals. The most significant not really set in stone the level of territorial contamination is the populace thickness after

Machine Learning Model for Identification of Covid-19 Future Forecasting

291

conclusion of the city, trailed by the traffic thickness and day by day populace thickness. Starting here of view, a little region with a little day by day populace and a little populace after the conclusion that is situate a long way from the traffic center point is moderately protected. The proportions of the lockdown are viable, fundamental and convenient. To confirm the aftereffects of the geo-identifier and create the danger standards according to the viewpoint of the choice tree, three primary rules are produced [4]. In this paper, Mohamed Abdel-Basset, Reda Mohamed, and colleagues propose that many countries be tested by the clinical assets needed for COVID-19 location, which necessitates the development of a low-cost, quick apparatus that can distinguish and analyse the infection adequately for a large number of tests. Despite the fact that a chest X-Ray examination is a useful up-and-comer instrument, the images generated by the outputs must be broken down accurately and quickly if large numbers of tests are to be prepared. Coronavirus induces reciprocal aspiratory parenchymal ground-glass and consolidative pneumonic opacities, which can have an altered morphology and spread to the fringes of the lungs. We hope to swiftly isolate similar tiny spots in chest X-Ray images that may include COVID-19 recognising pieces in this study. As a result, this study presents a half-and-half COVID-19 location model for X-Ray picture division based on a further developed marine hunters computation (IMPA). To improve the presentation of the IMPA and arrive at better arrangements in fewer cycles, the positioning based variety decrease (RDR) approach is used. RDR works on finding the particles that couldn’t find better arrangements after a certain number of cycles, and then relocating those particles to the best arrangements up to that point. The presentation of IMPA was approved based on nine chest X-Ray images with edge levels ranging from 10 to 100 and contrasted, as well as five condition of workmanship calculations: balance analyzer (EO), whale streamlining calculation (WOA), sine cosine calculation (SCA), Harris-falcons calculation (HHA), and salp swarm calculations (SSA). The trial results show that for a given set of measurements, the suggested mixture model outperforms all other calculations. Furthermore, in the Structured Similarity Index Metric (SSIM) and Universal Quality Index (UQI) measures, the presentation of our suggested model was simultaneous on all quantities of edges level [5].

2 Proposed Methodology Due to naturally distinguishing vital aspects from the preparation tests, taking care of the enactment from the previous time venture as contribution for the current time step, and organisations self-associations, AI methods were viable for prediction. According to the findings of the model inquiry, we believe that crisis intervention measures taken in the early stages of the pandemic, such as obstructing, limiting individual progression, and increasing assistance, had a critical regulating effect on the pandemic’s initial spread. The AI calculation LR, LASSO, SVM, ES, LSTM every one of these calculation are utilized and the best calculation are arranged in the r-squared blunder and the changed r-squared mistake. Maintaining interest in various clinical assets to ensure that suspected patients may be investigated and treated without wasting time is an extremely feasible anticipation and treatment method. Long transient memory (LSTM) of the plague were first fitted

292

N. Anitha et al.

and broken down to verify the validity of the existing numerical models. The findings were then used to fit and analyse COVID-19’s situation. For various borders and in various locations, the expected results of three different numerical models are diverse. The results obtained by the suggested technique for various parts (number of positive cases recovered, number of cases, and so on) will be precise within a particular range and will be a useful tool for managers and health officials. In this task, contains the accompanying modules: • • • •

INPUT DATASET DATA PRE PROCESSING CLASSIFICATION PREDICTION OF ACCURACY

2.1 Input Dataset Here we use a dataset from Kaggle. The data contains the total number of confirmed cases, the total number of deaths, the total number of newly confirmed cases, and the total number of healed patients per province. 2.2 Data Pre Processing Data preprocessing is a method for converting unclean data into a clean data set. The dataset is frequently insufficient, inconsistent, and/or missing in specific behaviours or trends, as well as containing several errors. Preprocessing data is a tried and true means of resolving such problems. When the first incidence of COVID-19 was reported in India, 80% of the data was used for training, while the remaining 20% was used for forecasting and validation. The resulting plot depicts the total number of confirmed cases, with observed data used for training, official data (green line) indicating official data accessible, and predicted data indicating a total number of confirmed cases forecast. The anticipated number of total confirmed positive cases roughly matches the available official statistics, as seen in this graph. 2.3 Classification For each data set point, the classification technique predicts the target class. By examining patients’ disease patterns, a risk factor might be associated with them using the categorization approach. The following machine learning methods will be used: LR, LASSO, ES, SVM, and LSTM. The r-squared score, mae (mean absolute error), rmse (root mean square error), precision, accuracy, and recall are all displayed with the results. 2.4 Prediction of Accuracy Predictive neural networks or characteristic data, such as infection event or non-event binomial effects, can be used with this technique. Various metrics’ forecast accuracy can be used for a variety of applications. The rate at which normal (non-expected prediction

Machine Learning Model for Identification of Covid-19 Future Forecasting

293

accurately forecasts sensitivity (non-infectious disease), accuracy (predicted percentage of predicted trend), positive predictive value, negative predictive value (correctly predicted infection rate is), and the ratio are among them. Expected predictions are a gauge of the possibility that the overall process will improve faster than the individual’s accuracy. In the form of a graph, the machine learning methods utilised in the suggested as well as the other are compared. It is possible to display a representation of the accuracy.

Patient’s data

Collection of data

Collection of relevant data

Data Preprocessi ng

Machine Learning

Number of positive cases Number of Negative cases

Predicting (positive /Negative)

Long shortterm Integrated Average (LSTM)

Death cases Output Result

3 Experimental Setup Using AI techniques, construct a framework for calculating the number of cases impacted by COVID-19 in the future. The dataset used in the analysis includes information on step-by-step reports of the number of late-tainted cases, recoveries, and deaths in the wake of COVID-19 around the world. The global community is becoming increasingly concerned as the fatality rate and confirmed cases continue to rise. The number of people who could be affected by the COVID-19 pandemic in different parts of the world is quite small. This study aims to establish the amount of persons who may be affected, as well as the number of new contaminated cases and deaths, as well as the expected recovery times over the following 10 days. There are a total of four AI models. To estimate the number of late debased instances, passings, and recoveries, the algorithms LR, LASSO,

294

N. Anitha et al.

SVM, ES, and LSTM were utilised. On the fifth sheet, the plot of certifiable scenario derived from the genuine data reports of the assessment’s analysing season follows the plots of declared cases, passes, and recoveries on the bottom four sheets. The figures show that the machine learning models used in this evaluation were appropriate for the assessing task, indicating a more comfortable study and future inquiry of the surrounding environment.

ALGORITHM

Accuracy

ACCURACY

LR

82

LASSO

78

SVM

52

ES

96

LSTM

98

120 100 80 60

40

Accuracy

20 0

4 Conclusion The number of possible COVID-19 positive cases in India for the next 10 days has been estimated using an information-driven anticipating/assessment technique. Using and bend fitting, the number of recovered cases, long transient memory (LSTM) per day positive cases, and expired cases was also determined. The influence of prevention methods such as friendly segregation and lockout has also been observed, demonstrating that these preventive measures can significantly reduce the transmission of the infection. Regardless of the fact that this technique frequently needs sufficient information that will assist it, in the early stages of pandemic transmission, this strategy can be used to more exactly predict scourge transmission pointers for the time being, in order to provide mediation control at all levels of the offices and strategy execution provides transient crisis avoidance programmes. Three different numerical models predict different outcomes for different boundaries and districts. In general, the Logistic model may have the most fitting influence among the three models.

References 1. Alsaeedy, A.A.R., Chong, E.: Identifying regions at risk for spreading COVID-19 Using existing cellular wireless network functionalities. IEEE Open J. Eng. Med. Biol. 1, 187–189 (2020) 2. Sear, R.F., et al.: Evaluating COVID-19 substance in the internet based wellbeing assessment war utilizing AI. IEEE Access, 1 (2020)

Machine Learning Model for Identification of Covid-19 Future Forecasting

295

3. Hu, S., et al.: Pitifully supervised deep learning for COVID-19 infection detection and classification from CT images. IEEE Access, 1 (2020) 4. Zhang, Y., Li, Y., Yang, B., Zheng, X., Chen, M.: Hazard assessment of COVID-19 based on multisource data from a geographical view. IEEE Access, 1 (2020) 5. Abdel-Basset, M., Mohamed, R., Elhoseny, M., Chakrabortty, R.K., Ryan, M.: A mixture COVID-19 discovery model utilizing a further developed marine hunter’s calculation and a positioning based variety decrease procedure. IEEE Access, 1 (2020) 6. Petropoulos, F., Makridakis, S.: Guaging the novel covid coronavirus. Plos one 15(3), e0231236 (2020) 7. Grasselli, G., Pesenti, A., Cecconi, M.: Basic consideration use for the Coronavirus flare-up in lombardy, italy: early experience and conjecture during a crisis reaction. Jama (2020) 8. Novel, C.P.E.R.E., et al.: The epidemiological attributes of an episode of 2019 novel Covid infections (Coronavirus) in china. Zhonghua liu xing bing xue za zhi=Zhonghua liuxingbingxue zazhi, 41(2), 145 (2020) 9. Grushka-Cockayne, Y., Jose, V.R.R.: Joining expectation stretches in the m4 rivalry. Int. J. Forecast. 36(1), 178–185 (2020) 10. Mediaite, N.C.: Harvard educator sounds caution on ‘probable’ Covid pandemic: 40% to 70% of world could be tainted for the current year. Gotten to on 2020.02.18. https://www.mediaite.com/news/harvardprofessor-sounds-alerton-possibleCovi dpandemic-40-to-70-ofworld-could-be-contaminatedforthispresentyear/

Alzheimer’s Disease Detection Using Machine Learning and Deep Learning Algorithms K. Sentamilselvan(B) , J. Swetha(B) , M. Sujitha(B) , and R. Vigasini(B) Department of Information Technology, Kongu Engineering College, Erode, Tamil Nadu, India [email protected]

Abstract. Alzheimer’s disease (AD) is a kind of neurodegenerative illness that affects memory-related brain cells and capacities. It is one of the most common neurodegenerative disorders that is not unusual. Clinical trials of Alzheimer’s disease pills failed 99.6% of the time. This research looked into machine learning approaches for using empirical statistics to predict Alzheimer’s disease development in the coming years. Diagnosing due to the degree of moderate cognitive impairment, Alzheimer’s disease can be difficult to diagnose, especially early in the illness’s course (MCI). The study purpose of CNN is to discover the most intricate pathways that are directly linked to alterations in Alzheimer’s disease. A range of imaging modalities are employed to diagnose Alzheimer’s disease, and the diagnosis is aided by the use of image modes. Early detection of Alzheimer’s illness, this research employs a CNN for training and a Random Forest Classifier, KNeighborsClassifier, XGBClassifier, and Logistic Regression for testing and classification algorithms. This study looks at how different types of machine learning algorithms can be used to solve AD diagnostic challenges. Deep learning has been proved to be a capable technique for handling a variety of picture identification issues Although the bulk of these published systems owe their effectiveness to rigorous training on a large number of data samples, according to current research. Keywords: Alzheimer’s disease · Deep learning · Convolutional neural network · Artificial neural networks · Random Forest Classifier · KNeighborsClassifier · XGBClassifier · Logistic regression

1 Introduction One of our body’s most important organs is the brain. The brain controls and facilitates all of the actions and responses that allow us to think and believe. It also strengthens our feelings and memories. Alzheimer’s disease is a progressive and permanent kind of mental decline. Every four seconds, someone in the world is diagnosed with Alzheimer’s disease. It improves at a sluggish rate and tears down memory cells, obliterating a person’s ability to think. It’s a degenerative nerve condition that causes neurons to lose function or perhaps die. After an Alzheimer’s diagnosis, the typical life expectancy is only four to eight years. This disorder affects one out of every ten persons over the age of 65, but it can also hit at a younger age, since it has been diagnosed in numerous people in © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 296–306, 2022. https://doi.org/10.1007/978-3-030-96299-9_29

Alzheimer’s Disease Detection

297

their 20s. This is the leading cause of dementia among the elderly. Alzheimer’s disease causes a deterioration in cognitive skills that are utilised to complete activities on a daily basis. Alzheimer’s disease is responsible for 60–80% of dementia cases. Plaques and tangles in the brain are the hallmarks of this illness as well as the destruction and death of brain cells. Dr. Alois Alzheimer was the first to notice it, after witnessing a lady die as a result of internal brain tissue alterations. The doctor scanned her brain after she died, and while doing so, he discovered the creation of numerous clumps. This was determined to be the primary cause of the sickness. They messed up the brain’s synchronisation with the rest of the body. As a result, persons with this disease find it difficult to carry out daily tasks like as driving, cooking, and so on. The symptoms aren’t always obvious in the early stages, but they can include trouble remembering names, misplacing critical belongings, having trouble planning, and so on. The middle stage of Alzheimer’s disease is the severe, with symptoms such as extreme mood swings, disorientation, impulsivity, short attention span, poor object recognition, and so on. The last stage is the most severe. Symptoms include trouble remembering names, misplacing crucial belongings, having trouble planning activities, and so on in the initial phases. The middle stage of Alzheimer’s disease is the most severe, with symptoms such as extreme mood swings, disorientation, impulsivity, short attention spans, poor object recognition, and so on. The most serious stage is the last. As a result, the disease has a significant risk of affecting the elderly. Although there is no cure for dementia, early treatment can help to halt its course. A healthy diet, physical activity, being social, protecting the head from injuries, reading, playing musical instruments, and engaging in intellectual activities have all been linked to lowering the risk of Alzheimer’s disease. These activities can strengthen overall brain health and cognitive effectiveness. As a result, the Convolutional Neural Networks (CNN) method is employed to detect the existence of Alzheimer’s disease. CNNs have the most important the benefit of lowering the total of parameters in ANNs. This achievement has prompted researchers and developers to examine larger models in order to address difficult problems that were previously impossible to solve using regular ANNs. The most important assumption about CNN problems is that they should not have spatially dependent features. It doesn’t matter where they are in the photographs as long as they are discovered. CNN’s ability to access information is another important feature.

298

K. Sentamilselvan et al.

2 Related Work People are impacted by a range of disorders that are connected to their surroundings and lifestyle, according to Dahiwade et al. As a result, forecasting ability has improved. The first stages of the illness are essential. Doctors, on the other hand, are having a difficult time predicting the future based on symptoms. The most challenging difficulty is determining a precise prognosis for a disease. Data mining is critical in sickness prediction to overcome this challenge. Medical research creates a vast amount of data every year. Increased data production in the medical and healthcare industries needs a more exact assessment of medical data obtained from early patient therapy. Data mining exposes hidden pattern knowledge in massive volumes of medical data using sickness data. The prognosis of an illness is influenced by the patient’s symptoms in general. P. Lodha and colleagues proposed the idea. Machine learning is becoming more prevalent in a range of medical fields. Improved data for detecting early indications of a variety of illnesses has emerged from medical technological breakthroughs. Alzheimer’s disease is a long-term condition in which brain cells degenerate, resulting in memory loss. Individuals with cognitive mental illnesses such as forgetfulness and ambiguity, as well as other symptoms such as psychological and behavioural concerns, may benefit from neuroimaging. The goal of this study is to use machine learning algorithms to understand data from neuroimaging technology so that Alzheimer’s disease can be detected early. According to Khan and Usman, Alzheimer’s disease is an incurable condition. As the aggregate size of the mind shrinks, a brain illness develops that impairs cognition and memory, finally leading to death. Early detection is required for the development of more widely used Alzheimer’s disease therapeutics. Machine learning (ML) is a field of artificial intelligence that uses a variety of probabilistic and optimization methodologies to allow computers to profit from big, complex datasets. As a result, researchers are focusing their efforts on using machine learning to detect Alzheimer’s disease at an early stage. The study examines, analyses, and critically evaluates existing research on using machine learning techniques to detect Alzheimer’s disease early. Several approaches have shown promise in terms of accuracy prediction, and they’ve all been tested on a variety of pathologically untested data sets from various imaging modes, making a fair comparison difficult. Francisco J. Martinez-Murcia et almachine’s learning from electronic health data methods may only predict a single endpoint. The ability to model hundreds of patient traits at once is a huge step toward customised Alzheimer’s disease treatment. Use the unsupervised machine learning model Conditional Restricted Boltzmann Machine to simulate accurate patient trajectories (CRBM). Using 18-month trajectory data from a patient, a model for personalised disease progression prediction was developed. There were 1909 people with mild to moderate intellectual disability or Alzheimer’s disease, and 44 clinical parameters were recorded. Synthetic patient data reflect the progression of each component of IQ abilities, analytical processes, and their sub-components’ links to critical clinical indicators. The CRBM’s synthetic patient data duplicates the means, margins of error, and correlations of each variable with such consistency over time that logistic regression is unable to distinguish between synthetic and genuine data.

Alzheimer’s Disease Detection

299

3 Materials and Methods The data for the study were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. The ADNI was founded in 2003 as a public–private collaboration directed by Principal Investigator Michael W. Weiner, MD. The primary purpose of the ADNI project was to see if serial MRI, PET, other biological markers, and clinical and neuropsychological testing could be used to track MCI and early AD development. According to Table 1, AD = 137, 18-month MCIc = 76, MCInc = 134, and HC = 162. The study included 162 cognitively normal elderly controls (HC), 137 Alzheimer’s disease (AD) patients, 76 MCI patients who converted to AD within 18 months (MCIc), 134 MCI patients who did not convert to AD within 18 months (MCI), and 137 MCI patients who did not convert to AD within 18 months (MCI) (MCInc). Patients with MCI who had been followed for less than 18 months were excluded from the research [3]. A total of 509 participants from 41 different radiology centres were investigated. The inclusion criteria for HC 20 and 26 years old were Mini Mental State Examination (MMSE) scores of 24 to 30; Clinical Dementia Rating (CDR) [18] of zero; and lack of depression; the exclusion criteria were CDR of 0.5 or 1.0; and NINCDS/ADRDA criteria for probable Alzheimer’s disease [17]. Table 1. The training and testing dataset. Variable

AD

MCIc

MCInc

HC

N

137

76

134

162

Gender (male: female)

67:70

43:33

84:50

86:76

Age (year: mean, std)

76.0, 7.3

74.8, 7.3

74.5, 7.2

76.3, 5.4

Weight (kg: mean, std)

70.9, 14.0

72.7, 14.3

76.2, 12.9

73.8, 13.6

MMSE (mean, std)

23.2, 2.0

26.47, 1.84

27.19, 1.71

29.18, 0.96

4 Convolutional Neural Network A CNN model’s convolution layer is frequently broken into two parts: feature extraction and feature mapping [11]. Each neuron in the feature-extraction segment is coupled to the higher layer’s local receptive field to extract local features. It is recognised Convolution is conducted on the input when a spatial link between a local when a spatial link between a local feature and other features is discovered data using a learnable filter or kernel to construct a feature map in the feature-mapping stage. Each neuron computes the outputs of neurons associated to receptive fields in the input by computing the dot product of its weight (i.e., filter) and a local receptive field (equal to filter size) via feature mapping (the input volume). To create a variety of feature maps, a series of learnable filters may be employed. The number of parameters that must be adjusted in the CNN is greatly reduced using this technique. After a convolutional layer, a pooling layer, such as the

300

K. Sentamilselvan et al.

max-pooling layer [25], conducts a spatial down-sampling operation (e.g., X, Y for a transverse slice). We can efficiently reduce feature resolution using our innovative dual-feature extraction approach [11] (Fig. 1).

Fig. 1. Working of CNN

4.1 Stride Indeed, CNN adds new variables, giving you more possibilities for lowering the parameters while removing some of the negative effects. Stride is one of these alternatives. Based on the areas, it is simply assumed that the node of the following layer has several overlaps. We may change the overlap by controlling the stride. Figure 2 shows a 7 * 7 picture that was supplied. If we adjust the filter one node at a time, we can only receive a 5 * 5 result. Figure 2 shows how the outputs of the three left matrices overlap (as do the outputs of the three middle matrices and the three right matrices). N +F (1) S However, if we continue and complete each step, the result will be 3 * 3. Simply simply, there will be less overlap, which will result in less production Eq. (1) formalises the output size O given the picture NN dimension and the FF filter size, as shown in Fig. 3. O=1+

4.2 Feature Map The input picture will be separated into layers in Fig. 4, and the outcomes of applying the filters will be collected. Units in a convolutional neural network’s hidden layer are divided into feature maps, with each feature map having the same weight matrix. The fact that hidden units on a feature map are related to independent units on the bottom layer distinguishes them. As a result, different parts of the hidden layer’s input picture will be associated to different units inside a feature map. A hidden layer is divided into feature maps, each of which looks for the same feature in the input image at many locations.

Alzheimer’s Disease Detection -1

1

1

1

-1

1

1

-1

1

-1

1

-1

1

-1

-1

1

1

1

-1

1

1

-1

-1

-1

1

-1

1

-1

-1

-1

-1

1

-1

1

-1

-1

-1

1

-1

-1

-1

1

-1

1

-1

-1

-1

-1

-1

301

Fig. 2. Stride 1, the filter window moves only one time for each connection.

N F

F

Fig. 3. The effect of stride in the output

-1 -1 -1 -1 -1 -1 -1

1 1 1 -1 -1 -1 1

1 -1 1 -1 -1 1 -1

1 1 1 1 1 -1 -1

-1 -1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1 -1

-1 -1 -1 -1 -1 -1 -1

-0.11

1

-0.11

-1

1

1

-0.55

0.11

-0.33

-1

1

-1

-0.33

0.33

-0.33

-1

1

1

-0.22

-0.11

-0.22

-0.33

-0.33

-0.33

Fig. 4. Feature map

302

K. Sentamilselvan et al.

4.3 ReLU Layer Right present, the ReLU is the most widely utilised activation function on the planet. Since then, it’s been employed in nearly all convolutional neural networks and deep learning systems. That all negative values become zero instantly, reducing the model’s capacity to fit or train from the input correctly. That is, any negative input to the ReLU activation function immediately converts the value to zero, as seen in Fig. 5. Without the activation function, there would just be a massive linear mapping between inputs and outputs, and classification skills would be lost. As a result, ReLU will aid in preventing the exponential development of the compute needed to run the neural network. 4.4 Max Pooling After a convolutional layer termed ReLU the convolutional layer generated by a convolutional layer, a pooling layer is added. Pooling is similar to applying a filter to feature maps in that it entails selecting a pooling procedure. The pooling operation or filter has a lower footprint than the feature map. It’s usually always 22 pixels wide with a 2 pixel stride. The pooling layer will always lower each feature map’s size by a factor of two. As illustrated in Fig. 6, Max pool will compute the maximum value for each patch of the feature map. -0.11 -0.55 -0.33 -0.22 -0.33

1 0.11 0.33 -0.11 -0.33

-0.11 -0.33 -0.33 -0.22 -0.33

0 0 0

1 0.11 0.33

0 0 0

0 0

0 0

0 0

Fig. 5. ReLU layer

0

1

0

0

0.11

0

0

0.33

0

0

0

0

0

0

0

1 0.33 0.33 0

1 0.33 0.33 0

Fig. 6. Max pooling

4.5 Fully-Connected Layer The fully linked layer of the CNN represents the input feature vector. The data in this feature vector is crucial for the input. This feature vector may be utilised for classification, regression, or as an input into another network to convert into a new type of output, as seen in Fig. 6, once the network has been trained. It can also be used as a vector that has been encoded. This feature vector is used to compute the loss and aid in the training

Alzheimer’s Disease Detection

303

of the network during training. The convolution layers store information about local properties in the input picture, such as edges, shapes, and so on, before the fully coupled layer(s). Each convolution layer has a number of filters, each representing one of the local features. The Fully Connected Layer is all about feed forward neural networks. The network’s last tiers are the Fully Connected Tiers. The input to the completely connected layer is the output from the final Pooling or Convolutional Layer, which is flattened and then fed into the fully connected layer. In most machine learning models, the last few levels are full connected layers that combine the data acquired by preceding layers to produce the final output. After the characteristics have been retrieved, we must categorise them. The Fully Connected Layer is all about feed forward neural networks. The network’s last tiers are the Fully Connected Levels. The input to the completely connected layer is the output from the final Pooling or Convolutional Layer, which is flattened and then fed into the fully connected layer. In most machine learning models, the last few levels are full connected layers that combine the data acquired by preceding layers to produce the final output. After feature extraction, we need to categorise the data into many classes, which we can do with a fully connected neural network. However, to make the model trainable from start to finish, we generally end up adding totally connected layers. The fully connected layers learn a function between the convolutional layers’ high-level properties.

5 Conclusion A correct diagnosis of Alzheimer’s disease allows the patient to receive the best therapy possible. Many researchers are working on this difficult issue, and they have built a number of CAD systems to aid in the diagnosis of Alzheimer’s disease. To do the categorization, we designed a deep learning technique in our process. There were a total of 6400 photos utilised, with 896 mildly demented, 64 moderately demented, 3200 nondemented, and 2240 very mildly demented being used The algorithm’s performance was evaluated using 1013 test pictures, including 139 mildly demented, 10 moderately demented, 530 non-demented, and 334 very mildly demented. Figure 8 shows that utilizing (Fig. 7).

Fig. 7. Fully connected layer

304

K. Sentamilselvan et al.

Fig. 8. Confusion matrix

Convolutional Neural Networks, a 95% accuracy level may be achieved. The confusion metrics were calculated as shown in Fig. 8, with Precison, Recall, and F1 score being used for further study of the classification problem. Based on characteristics in the feature extractor of a Convolutional Neural Network, the suggested approach can diagnose the disease and offer the accuracy of alzheimer’s disease. It performs feature map by separating the picture into layers to increase the usefulness and flexibility for discovering the illness, and then maxpool is used to reduce the dimensionality. Health Lab is a website where each unique image may be uploaded and the accuracy for that image is displayed (Fig. 9).

Fig. 9. Graph of training and validation accuracy with epochs

Alzheimer’s Disease Detection

305

References 1. Alzheimer’s Association: 2018 Alzheimer’s disease facts and figures. Alzheimer’s Dementia 14(3), 367–429 (2018) 2. Cuingnet, R., et al.: Alzheimer’s Disease Neuroimaging Initiative. Automatic classification of patients with Alzheimer’s disease from structural MRI: a comparison of ten methods using the ADNI database. Neuroimage 56(2), 766–781(2011) 3. Christian, S., Antonio, C., Petronilla, B., Gilardi, M.C., Aldo, Q., Isabella, C.: Magnetic resonance imaging biomarkers for the early diagnosis of Alzheimer’s disease: a machine learning approach. Front. Neurosci. 9, 307 (2015). https://doi.org/10.3389/fnins.2015.00307 4. Eke, C.S., Jammeh, E., Li, X., Carroll, C.: Early detection of Alzheimer’s disease with blood plasma proteins using support vector machines. IEEE J. Biomed. Health Inform. 25(1), 218– 226 (2021). https://doi.org/10.1109/JBHI.2020.2984355 5. De Strooper, B., Karran, E.: The cellular phase of Alzheimer’s disease. Cell 164(4), 603–615 (2016) 6. Dahiwade, D., Patle, G., Meshram, E.: Designing disease prediction model using machine learning approach. In: 2019 3rd International Conference on Computing Methodologies and Communication (ICCMC), pp. 1211–1215. IEEE (2019) 7. Fisher, C.K., Smith, A.M., Walsh, J.R.: Machine learning for comprehensive forecasting of Alzheimer’s disease progression. Sci. Rep. 9(1), 1–14 (2019) 8. Martinez-Murcia, F.J., Ortiz, A., Gorriz, J.-M., Ramirez, J., Castillo-Barnes, D.: Studying the manifold structure of Alzheimer’s disease: a deep learning approach using convolutional autoencoders. IEEE J. Biomed. Health Inform. 24(1), 17–26 (2020). https://doi.org/10.1109/ JBHI.2019.2914970 9. Galvin, J.E.: Prevention of Alzheimer’s disease: lessons learned and applied. J. Am. Geriatr. Soc. 65(10), 2128–2133 (2017) 10. Jain, R., Jain, N., Aggarwal, A., Hemanth, D.J.: Convolutional neural network based Alzheimer’s disease classification from magnetic resonance brain images. Cogn. Syst. Res. 57, 147–159 (2019) 11. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with deep convolutional neural networks. In: Proceedings of the International Conference on Neural Information Processing Systems, pp. 1097–1105. MIT Press, Cambridge (2019) 12. Khan, A., Usman, M.: Early diagnosis of Alzheimer’s disease using machine learning techniques: a review paper. In: 2015 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K), vol. 1, pp. 380–387 (2015) 13. Liu, M., Zhang, D., Shen, D., Alzheimer’s Disease Neuroimaging Initiative: Ensemble sparse classification of Alzheimer’s disease. NeuroImage 60(2), 1106–1116 (2012) 14. Nie, L., Zhang, L., Meng, L., Song, X., Chang, X., Li, X.: Modeling disease progression via multisource multitask learners: a case study with Alzheimer’s disease. IEEE Trans. Neural Netw. Learn. Syst. 28(7), 1508–1519 (2017). https://doi.org/10.1109/TNNLS.2016.2520964 15. Long, X., Chen, L., Jiang, C., Zhang, L., Alzheimer’s disease Neuroimaging Initiative: Prediction and classification of Alzheimer disease based on quantification of MRI deformation. PloS one 12(3), e0173372 (2017) 16. Lodha, P., Talele, A., Degaonkar, K.: Diagnosis of Alzheimer’s disease using machine learning. In: 2018 Fourth International Conference on Computing Communication Control and Automation (ICCUBEA), pp. 1–4 (2018) 17. McKhann, G., Drachman, D., Folstein, M., Katzman, R., Price, D., Stadlan, E.M.: Clinical diagnosis of Alzheimer’s disease report of the NINCDS ADRDA Work Group∗ under the auspices of department of health and human services task force on Alzheimer’s disease. Neurology 34, 939 (1984). https://doi.org/10.1212/wnl.34.7.939

306

K. Sentamilselvan et al.

18. Morris, J.C.: The Clinical Dementia Rating (CDR): current version and scoring rules. Neurology 43, 2412–2414 (1993). https://doi.org/10.1212/WNL.43.11.2412-a 19. Dadar, M., et al.: Validation of a regression technique for segmentation of white matter hyperintensities in Alzheimer’s disease. IEEE Trans. Med. Imaging 36(8), 1758–1768 (2017). https://doi.org/10.1109/TMI.2017.2693978 20. Schelke, M.W., et al.: Mechanisms of risk reduction in the clinical practice of Alzheimer’s disease prevention. Front. Aging Neurosci. 10, 96 (2018) 21. Tong, T., Gray, K., Gao, Q., Chen, L., Rueckert, D., Alzheimer’s Disease Neuroimaging Initiative: Multi-modal classification of Alzheimer’s disease using nonlinear graph fusion. Pattern Recogn. 63, 171–181 (2017) 22. Zhou, T., Liu, M., Thung, K.-H., Shen, D.: Latent representation learning for Alzheimer’s disease diagnosis with incomplete multi modality neuroimaging and genetic data. IEEE Trans. Med. Imaging 38(10), 2411–2422 (2019). https://doi.org/10.1109/TMI.2019.2913158 23. Veitch, D.P., et al.: Understanding disease progression and improving Alzheimer’s disease clinical trials: recent highlights from the Alzheimer’s Disease Neuroimaging Initiative. Alzheimer’s Dementia 15(1), 106–152 (2019) 24. Wechsler, D.: Manual: Wechsler Memory Scale-Revised. Psychological Corporation, San Antonio (1987) 25. Weng, J., Ahuja, N., Huang, T.S.: Cresceptron: a self-organizing neural network which grows adaptively. In: Proceedings of the International Joint Conference on Neural Networks, Baltimore, MD, vol. 1, pp. 576–581 (1992). https://doi.org/10.1109/IJCNN.1992.287150 26. Zhang, D., Wang, Y., Zhou, L., Yuan, H., Shen, D., Alzheimer’s Disease Neuroimaging Initiative: Multimodal classification of Alzheimer’s disease and mild cognitive impairment. Neuroimage 55(3), 856–867 (2011)

Crime Factor Anaysis and Prediction Using Machine Learning N. Anitha(B) , S. Gowtham, M. Kaarniha Shri, and T. Kalaiyarasi Kongu Engineering College, Erode, India [email protected]

Abstract. A crime, often known as a criminal offence, is an act that harms not just people but also a community, society, or state. Every second, a large number of crimes occur in various locations, patterns, and times, and the number continues to rise. Murder, retaliation, theft, rape, kidnapping, and burglary are only a few examples of many sorts of crimes. As a result, it’s critical to recognise various reasons and criminal patterns. This study examines how machine learning algorithms may be created and analysed to classify crimes as high, medium, or low based on real-time parameters such as single parent crime, health, and crime rate, punished crime. Keywords: Single parent crime · Health factor · Crime rate · Punished crime

1 Introduction A crime, often known as a criminal offence, is an act that harms not just people but also values of a given society. Such behaviour act as prohibited and penalised under the rules. Crime is one of the biggest problem and its preventing is critical duty. The Department of Police has enormous challenges in crime prediction and criminal identification due to the vast volume of crime data available. Every day, a large number of crimes occur in various locations and in various patterns, with the number of crimes increasing with each passing day. From a strategic standpoint, the capacity to forecast any crime based on geography, age groupings, and other factors can aid in delivering important information to law enforcement. However, because crime is expanding at an alarming rate, precisely anticipating crime is a difficult undertaking. As a result, crime factor prediction and analysis approaches are critical for detecting and reducing future crimes. In order to forecast and assist prevent future crimes, the Crime Factor Prediction technique includes employing algorithms to analyse huge volumes of data. The most extensively used approach, place-based prediction, often analyses pre-existing crime data to identify locales and age groups of those who committed harmful behaviours with a high chance of crime. Machine Learning Algorithm used for extract and detect the patterns in huge data sets. It is a computer science and statistics multidisciplinary topic with the purpose of extracting details from a data collection into intelligible structure for subsequent. Find the analytical stage is called as data mining. Statistical algorithms and machine learning techniques are used in predictive analytics to determine the values of © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 307–313, 2022. https://doi.org/10.1007/978-3-030-96299-9_30

308

N. Anitha et al.

upcoming events based on previous records. It is provide the best judgement of what will happen in the future, rather than just knowing what has happened. Predictive analysis entails the creation of models that aid in prediction. 1.1 Literature Review Jha et al. (2020), Crime forecasting has been one of the most complex challenges in law enforcement today. In this article, we use a strategy for a time series model and machine testing systems for crime estimation. First, we used the existing machine learning techniques to predict the number of crimes. Second, we proposed two approaches, a modified autoregressive integrated moving average model and a modified artificial neural network model. The first important objective is to compare the applicability of a unchanging time series model against that of a changing time series model for crime forecasting. More than two million crime informations are tested. The work concludes that variate time series model yields better forecasting results than the predicted values from existing techniques. Aarthi et al. (2019), made an article which is the process of understanding data through cleaning raw data, finding patterns, creating models, and testing those models. It includes statistics, machine learning, and database systems. Data mining is a technique that can assist us with crime detection problem. We can capture the years of human experience into computer models through data mining by performing the clustering the population or data points are divided into number of groups. By obtaining the final product which is a project that can be predicted by using primary data sets and the output will be simple to understand to the user. Algorithm that guarantees convergence in probability are derived in the paper. Then we can start the centroid positions with new examples, it is not easy to predict the value of K by performing global cluster. When we initiate a partition differently we will get different final clusters. Mansour and Lundy (2019) published an essay on crime prevention and resolution tactics that rely on the presence of a crime team in the right location at the right time. The goal of our present study is to discover the geographical, temporal, meteorological, and event characteristics that are most frequently linked to distinct crime categories. The results reveal that the XG boost method performs the best and that geographical, temporal, and meteorological factors all play a role in crime type categorization. The following are some of the benefits of random forest: It’s one of the most precise learning algorithms on the market. The biggest drawback of random forest is that it might become too sluggish and useless for real-time forecasts if there are too many trees. For linearly separable data, linear regression works remarkably well. Shinde et al. (2020) suggested every culture has a large amount of crime. On this subject, a lot of analysis has been done on it. The approaches and algorithms utilised in those publications for inspection and anticipation of violation are the major emphasis. When used in a given business, data mining offers several advantages. Aside from such benefits, data mining has its own drawbacks, such as privacy, security, and information abuse.

Crime Factor Anaysis and Prediction Using Machine Learning

309

2 Proposed System A suitable algorithms and statistical tools can be used to do crime factor analysis. Navie Bayes algorithm and the decision tree algorithm are two more strategies for crime prediction and analysis covered here. As raw data, the dataset was gathered from a criminology journalist. And after removing the noise data from the dataset, the cluster and cluster overhead were formed. Then, using PSO (Particle Swarm Optimization), the features were selected and the optimum G data point and optimised data boundary were calculated. This method is fitted the train data set using Naive Bayes Algorithm. Finally, based on the crime factor data, the offences were classified as high, low, or medium. 2.1 Particle Swarm Optimization It is one of the bio’s algorithms, and it is simple to get an optimal solution with this approach. In comparison to other approaches, the PSO provides precise results in a shorter span and at a cost of low. Each particle’s movement is guided not just by its local best location, but also by the best places in the search space, which are updated when other particles find a better area. As a result, the swarm is likely to migrate toward the best options. 2.2 PSO Design Steps Step 1: The number of route nodes, as well as the number of interpolation, is dictated by the current environment. Step 2: Set the particle’s parameters uniformly and set the particle’s location. The particle population and velocity are then set to zero. Step 3: Calculate the coordinates of m interpolation points in each particle’s x and y directions. Step 4: Calculate the particle’s fitness value. Step 5: Update the particle’s velocity and location. P best, the local optimal value, and G best, the global ideal value. Step 6: Determine if the updated particle intersects and, if so, create a path using the path node coordinates. One iteration is added to the total number of iterations. Step 7: The algorithm stops and the optimal path is output if the termination condition is fulfilled (the threshold mistake is small enough to be ignored). If not, it returns to Step 3 and restarts the process. 2.3 Multinomial NB (MNB) It is an algorithm that uses Bayes’ Theorem for classify the objects. A Naive Bayes assume strong or naïve between attributes of data. Naive bayes design steps: Step 1: Create a frequency table from the data collection. Step 2: And then create Likelihood table by calculating some probabilities such as Overcast probability. Step 3: Calculate the posterior probability for each class using the Naive Bayesian equation. The outcome of prediction is the class with the highest posterior probability.

310

N. Anitha et al.

2.4 Categorizing the Crimes After that, divide the dataset into low, medium, and high categories and use the Nave Bayes technique to fit the model to the training dataset. And putting the statistics to the test by providing information on criminal factors such as single parenthood, crime rate, relative crime rate, poverty, and punishment. If the total of input variables such as single parent, poverty, mental state, past crime, health factor, number of crimes per day, parents crime rate, friends crime rate, relative crime rate, neighbour crime rate, and region crime rate is less than 5, the model will output the category as low. If the summing result is less than or equal to 6.5, the classification is output as medium. If the summation value is more than 6.5, the output classification will be high. The criminology journalist sets the threshold values (Fig. 1).

Fig. 1. Graphical workflow of proposed models for detection phishing of websites.

3 Results and Discussion Find the crime factor using machine learning algorithm is more advanced in study, with the goal of lowering crime rates. This study focuses on various crime categories and their incidence in various locations. The efficiency of the suggested approach based on categorising crimes based on crime variables such as single parent crime, health factor, poverty, punishment, mental condition, and so on was effectively categorised. Experiments are run on a real-time dataset using Naive Bayes machine learning methods.

Crime Factor Anaysis and Prediction Using Machine Learning

311

312

N. Anitha et al.

4 Conclusion Find the crime factor using machine learning algorithm is more advanced in study, with the goal of lowering crime rates. The focus of this study is on the many sorts of crime. The offences were correctly classified as high, low, and medium using the Nave Bayes method. Experiments are carried out utilising machine learning techniques, such as the Nave Bayes algorithm, using areal-time dataset. The study will be expanded in the future by applying it to large datasets and producing findings based on a comparison of different Machine Learning techniques.

References Premalatha, M., Vijayalakshmi, C., Vijayalakshmi, C.: SVM approach for non-parametric method in classification and regression learning process on feature selection with $epsilon$-insensitive region. Malaya J. Mat. S(1), 276–279 (2019) Aarthi, S., Samyuktha, M., Sahana, M.: Crime hotspot detection with clustering algorithm using data mining. In: 2019 3rd International Conference on Trends in Electronics and Informatics (ICOEI), pp. 401–405. IEEE (2019) Al Saidi, W., Zeki, A.M.: The use of data mining techniques in crime prevention and prediction. In: 2nd Smart Cities Symposium (SCS 2019), pp. 1–4. IET (2019) Chauhan, C., Sehgal, S.: A review: crime analysis using data mining techniques and algorithms. In: 2017 International Conference on Computing, Communication and Automation (ICCCA), pp. 21–25. IEEE (2017) Iqbal, R.: An experimental study of classification algorithms for crime prediction. Indian J. Sci. Technol. 6(3), 1–7 (2013) Jha, S., Yang, E., Almagrabi, A.O., Bashir, A.K., Joshi, G.P.: Comparative analysis of time series model and machine testing systems for crime forecasting. Neural Comput. Appl. 33(17), 10621– 10636 (2020) Kim, S., Joshi, P., Kalsi, P.S., Taheri, P.: Crime analysis through machine learning. In: 2018 IEEE 9th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), pp. 415–420. IEEE (2018) Kumar, R., Nagpal, B.: Analysis and prediction of crime patterns using big data. Int. J. Inf. Technol. 11(4), 799–805 (2019) Krishnendu, S.G., Lakshmi, P.P., Nitha, L.: Crime analysis and prediction using optimized Kmeans algorithm. In: 2020 Fourth International Conference on Computing Methodologies and Communication (ICCMC), pp. 915–918. IEEE (2020) Mansour, H.A., Lundy, M.: Crime types prediction. In: Alfaries, A., Mengash, H., Yasar, A., Shakshuki, E. (eds.) Advances in Data Science, Cyber Security and IT Applications: First International Conference on Computing, ICC 2019, Riyadh, Saudi Arabia, December 10–12, 2019, Proceedings, Part I, pp. 260–274. Springer International Publishing, Cham (2019). https:// doi.org/10.1007/978-3-030-36365-9_22 Shinde, V., Bhatt, Y., Wawage, S., Kongre, V., Sonar, R.: Application of data mining for analysis and prediction of crime. In: Senjyu, T., Mahalle, P.N., Perumal, T., Joshi, A. (eds.) Information and Communication Technology for Intelligent Systems: Proceedings of ICTIS 2020, Volume 1, pp. 91–102. Springer Singapore, Singapore (2021). https://doi.org/10.1007/978-981-15-707 8-0_8 Singh, N., Kaverappa, C.B., Joshi, J.D.: Data mining for prevention of crimes. In: Yamamoto, S., Mori, H. (eds.) HIMI 2018. LNCS, vol. 10904, pp. 705–717. Springer, Cham (2018). https:// doi.org/10.1007/978-3-319-92043-6_55

Crime Factor Anaysis and Prediction Using Machine Learning

313

Sivanagaleela, B., Rajesh, S.: Crime analysis and prediction using fuzzy c-means algorithm. In: 2019 3rd International Conference on Trends in Electronics and Informatics (ICOEI), pp. 595– 599. IEEE, (2019) Yadav, S., Timbadia, M., Yadav, A., Vishwakarma, R., Yadav, N.: Crime pattern detection, analysis & prediction. In: 2017 International conference of Electronics, Communication and Aerospace Technology (ICECA), vol. 1, pp. 225–230. IEEE (2017)

Detection of Fake Reviews on Online Products Using Machine Learning Algorithms H. Muthu Krishnan(B) , J. Preetha(B) , S. P. Shona(B) , and A. Sivakami(B) Department of Information Technology, Kongu Engineering College, Erode, Tamil Nadu, India {hm.it,preethaj.18it}@kongu.edu

Abstract. Internet reviews may affect a customer’s selection, and they can assess a product by correlating it to other brands. If the reviews are authentic, the customer may only pick a product that fulfils their demands. If the reviews, on the other hand, are phoney, the buyer is duped. It is vital to obtain the identification of fake consumer opinions in order to address this issue. To evaluate if a review is fraudulent or not, the actions of reviewers are extracted based on a semantic analysis of his review content. In this study, a data set for a mixed product was retrieved from the web, along with reviews and other information about the reviewers, to identify false reviewers using four algorithms: Support Vector Machine, Logistic Regression, K-Nearest Neighbour and Decision Tree. The accuracy rate along with precision rate of the aforementioned four methods are used to validate the significance of the features on the choice. Experiments were carried out on a large number of reviews gathered from the internet, demonstrating the efficacy of the proposed method. Keywords: Support Vector Machine (SVM) · Decision Tree (DT) · Logistic Regression (LR) · K- Nearest Neighbour (KNN)

1 Introduction In this age of the internet, customers may now write evaluations or thoughts on a number of websites. These reviews benefit both companies and potential customers who want to get a feel of what items or services are offered before making a purchase. In recent years, the quantity of consumer evaluations has constantly increased. Customer feedback influences the decisions of potential purchasers. In other words, when shoppers read product evaluations on social media, they decide whether to purchase the goods or not. As a result, customer reviews provide consumers with a crucial service. Positive reviews provide significant financial benefit, whilst negative reviews frequently have a negative financial impact. As a result, as users become more prominent in the marketplace, firms are increasingly depending on customer input to enhance their products, services, and marketing. The spammers are referred to as a group of spammers when numerous clients who acquired a specific model spammer to attain a specific aim. Many research has looked into the subject of detecting fake reviews and the obstacles that come with it. The most © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 314–319, 2022. https://doi.org/10.1007/978-3-030-96299-9_31

Detection of Fake Reviews on Online Products

315

difficult part of detecting fraudulent reviews is determining whether they are genuine or not. We have offered a complete review of the literature in this survey study in order to better identify existing challenges and future research opportunities in this field. It has both statistical machine learning and data mining algorithms, making it easy for researchers interested in identifying fraudulent reviews to choose the right machine learning approach. This article incorporates relevant papers from Google Scholar, Web of Sciences, and a few high-profile conferences to assist the reader comprehend the topic of phoney review detection. Finally, publications from 2007 to 2021 have been picked for summary and analysis.

2 Related Work Numerous researches have been published in the topic of identifying fake reviews. The effectiveness of a number of methods for classifying and detecting review spam, as well as popular machine learning methodologies for dealing with the problem of review phishing detection by M. Crawford et al. [1]. A technique was proposed that uses Opinion Mining to identify phoney or spam product reviews by A. Danti et al. [2]. To take use of qualities of Autoencoder and Random Forest, an end-to-end trainable model was built, and a stochastic model has now been established to lead the global parameter process. The effectiveness of a variety of ways for detecting and classifying phishing reviews, as well as the most well-known machine learning techniques designed to address the challenge of detecting review spam. It is feasible that identifying spam and phoney reviews using sentiment analysis, as well as a find review that include the expletives or swear words by S. Kumar et al. [3]. It has been proposed that machine learning techniques be used to Sort movie reviews into categories based on whether they’re good or negative by E. Elmurngi et al. [4]. A sentiment analysis-based strategy for detecting untruthful user assessments with discrete semantic content, akin to movie reviews, has been provided by M. R. Adike and V. S. Reddy et al. [5]. A. Tewari and S. Jangale et al. [6] suggested a method for categorizing tweets into spam and non-spam categories. The language and rating attributes of a review have been proposed as a means to identify fake product reviews by E. Wahyuni and A. Djunaidy et al. [7].

3 Machine Learning Artificial intelligence has a section called machine learning (AI) that centers on the concept of a computer programme learning and adapting to new data without requiring human intervention. A computer is programmed with a complex algorithm or source code that allows it to recognize data and make predictions based on that data. Machine learning can help in decision-making by parsing the massive amount of data that is consistently and freely available around the world. Machine learning has a wide range of applications, including investment, advertising, financing, news organization, fraud detection, and more.

316

H. M. Krishnan et al.

4 Logistic Regression This is a machine learning method for addressing problems difficulties in categorization. It is a probability-based predictive analytic method. A Logistic Regression model is similar to a Linear Regression model, but instead of using a linear function, it employs the ‘logistic function’ or ‘Sigmoid function’ which is a more advanced cost function. According, the logistic regression hypothesis, the cost function should be confined to a value between 0 and 1. As a result, linear functions fail to characterize it since it might have a value more than 1 or less than 0, which the logistic regression hypothesis says is impossible. This function generates a ‘S’ curve that may be used to convert any real-valued integer to a value between 0 and 1. y predicted becomes 1 when the curve reaches positive infinity, while y predicted becomes 0 when the curve reaches negative infinity If the sigmoid function’s output is greater than 0.5, we can classify the result as 1 or YES, and if it is less than 0.5, we can classify it as NO., we can classify the result as 0 or NO. Sigmoid Function, P = 1/1 + eˆ(−y).

5 Support Vector Machine A supervised learning system called a support vector machine is used to solve classification and regression problems. Many people prefer the support vector machine because it produces significant correctness while using less computing power. It’s primarily used to solve categorization challenges. Learning can be classified into three categories: supervised, unsupervised, and reinforcement learning. The hyperplane is divided to produce a support vector machine, which is a selective classifier. The algorithm generates the optimum hyperplane from labelled training data in order to classify new samples. This hyperplane is a line that divides a plane into two halves in two-dimensional space, with each class on either side. This method tries to find a hyperplane in N-dimensional space in the graph that labels individual data points.

6 K-Nearest Neighbor (KNN) This is a method for grouping new data points into target classes based on adjacent data points’ properties. Features of KNN Method: KNN is a method groups with similar of features, which means it makes no assumptions about the data set, unlike other methods. The method is effortless and more effective realistic data, which means it memorizes a discriminative function rather than learning it from the training data set. KNNs are a form of neural network frequently utilized to solve classification and regression problems. Drawbacks: After several iterations, it was shown that the approach works poorly on huge datasets because the significant expense of determining the length between every new & old point, which slows down the operation. It’s also been noted that using this approach to

Detection of Fake Reviews on Online Products

317

work with high-dimensional data is problematic because the distance calculation in each dimension is incorrect. Before applying the KNN method to any dataset, it is necessary to conduct feature scaling, i.e., standardization and normalizing. If these phases are skipped, the KNN algorithm may make incorrect predictions. Sensitive to outliers, missing values, and noisy data: The KNN algorithm is sensitive to dataset noise. Missing values must be manually imputed, and outliers must be removed.

7 Decision Tree A decision tree is a simple method of classifying samples. A class is the name given to each element of the categorization domain. A decision tree, also called as a classification model, is a tree wherein every interior node has an input feature labelled on it. The curves that emerge from a system are made with an input feature are either labelled with several of the target feature’s potential values or they are routed to a subordinate point for a various input. The leaves are marked with a distribution function, indicating that the tree has placed the data set into one of the classes or a probability distribution over the classes (which, if the decision tree is flawless, is skewed towards certain subsets of classes). By separating into subgroups based on value test, a tree may be “trained”. Recursive partitioning is the process of repeating this approach on subset. When the target variable’s value is the same in all subgroups, or when splitting no longer adds value to the estimations, the recursion is complete. Because it doesn’t need domain knowledge or parameter selection, a decision tree classifier is ideal. The accuracy of the decision tree classifier is usually rather excellent. A common inductive strategy for obtaining categorization information is decision tree induction.

8 Proposed Methodology Reviews on website products are mostly used to make purchasing decisions, which can lead to users being misled in their decisions. Find out how to avoid the negative aspects of online reviews. For discovering false reviews, the dataset was trained and evaluated with four different types of algorithms and the best performing approach was chosen. In order to detect bogus news, the dataset is fed into various algorithms. To get at the final outcome, the correctness of the obtained results is examined.

9 Training and Testing The dataset has been trained and validated. The necessary libraries have been imported. Only the relevant characteristics are taken from the dataset and trained. The Logistic Algorithm is one of four algorithms used. Regression is a Machine Learning technique for resolving classification issues. It’s a predictive analysis tool based on the probability concept. As an algorithm, it uses a decision tree classifier. Decision Trees are a sort of Supervised Machine Learning that describes what the input in the training data is and what the related output is, with the data being divided according to a parameter on a regular basis. KNN classifies data point into the target class based on the data points. A supervised learning system called a support vector machine is used to solve classification and regression problems. Then there was the model.

318

H. M. Krishnan et al.

10 Accuracy Rate Fake reviews have been classified by comparing algorithms such as Support Vector Machine, which has a 90% accuracy rate, Decision Tree, which has a 68% accuracy rate, K closest neighbor, which has a 90% accuracy rate, and Logistic regression, which has a 92% accuracy rate. That Logistic Regression Algorithm has a high rate of precision and accuracy. In comparison, the logistic regression has a greater accuracy rate. When it comes to determining whether online product reviews are fake or real, Logistic Regression is more effective. The number of true positives and negatives divided by the number of true positives, true negatives, false positives, and false negatives generates the accuracy rate (TP + TN)/(TP + FP + TN + FN).

11 Conclusion Finally, the effective detection of bogus reviews demonstrated the possibility. We can forecast fake reviews on internet products by calculating its probability. To determine if a review is fake or not, various algorithms are used to identify fake reviews from internet reviews. The logistic algorithm has a greater accuracy and is suggested for detecting bogus reviews. The recommended technique has a 92% success rate in terms of efficiency.

Detection of Fake Reviews on Online Products

319

References 1. Crawford, M., Khoshgoftaar, T.M., Prusa, J.D., Richter, A.N., Al Najada, H.: Survey of review spam detection using machine learning techniques. J. Big Data 2(1), 1–24 (2015). https://doi. org/10.1186/s40537-015-0029-9 2. Danti, A., Sanjay, K.S.: Detection of fake opinions on online products using decision tree and information gain. In: 2019 3rd International Conference on Computing Methodologies and Communication (ICCMC), pp. 372–375, 27 March 2019 3. Chauhan, S., Anupam, G., Prafull, G., Avishkar, C., Mahendra, G.: Research on product review analysis and spam review detection, pp. 390–393 (2017). https://doi.org/10.1109/SPIN.2017. 8049980 4. Elmurngi, E., Gherbi, A.: Fake reviews detection on movie reviews through sentiment analysis using supervised learning techniques. Int. J. Adv. Syst. Meas. 11(1&2), 196–207 (2018) 5. Adike, M.R., Reddy, V.S.: Detection of Fake Review and Brand Spam Using Data Mining Technique (2016) 6. Tewari, A., Jangale, S.: Spam filtering methods and machine learning algorithm-a survey. Int. J. Comput. Appl. 154, 8–12 (2016). https://doi.org/10.5120/ijca2016912153 7. Wahyuni, E., Djunaidy, A.: Fake review detection from a product review using modified method of iterative computation framework. In: MATEC Web of Conferences, vol. 58 (2016). https:// doi.org/10.1051/matecconf/20165803003

Deep Neural Network Model for Automatic Detection of Citrus Fruit and Leaf Disease S. Anandamurugan(B) , B. Deva Dharshini(B) , J. Ayesha Howla(B) , and T. Ranjith(B) Kongu Engineering College, Erode, Tamilnadu, India

Abstract. Citrus yield decreases are mostly caused by citrus fruit and leaf diseases. As a result, developing an automated detection method for citrus plant diseases is critical. Deep learning algorithms have recently shown promising and favourable outcomes, prompting us to take on the problem of identifying citrus fruit and leaf illnesses. MLP classifiers are suggested using a unified approach in this paper. The suggested Multilayer perceptron classifier model is focused on distinguishing healthy fruits and leaves from those with common citrus illnesses such as Scab, Melanose, canker, Black spot, and greening disease. The preliminary findings show that the MLP classifier model outperforms the other model by a significant margin. Keywords: Disease detection · MLP classifier · Deep learning

1 Introduction Agriculture has a crucial role in a country’s economic development. Its study aims to boost food output and quality while cutting costs and increasing business margins. Citrus disease comes in a variety of forms, the most common of which are citrus canker, greening disease, scab disease, melanose, and black spot. Citrus canker is a disease caused by the bacteria xanthomonas axonopodis that affects citrus crops by causing lesions on the stems, leaves, and fruits. Bacteria leak out of the citrus plant when there is free moisture, infecting new growth and development. Greening disease is characterised by yellow veins and neighbouring tissues, as well as splotchy mottling of the entire leaf, twig dieback, and feeder rootlet and lateral root degradation. Scab is a minor disease that affects mature plants and causes scab symptoms on leaves, twigs, and fruits, as well as a yellow dazzling edge. After the fruit reaches full maturity, black spot manifests itself as a dark brown to black spot. Melanose is a fungal infection that causes plant and fruit damage in citrus trees.

2 Related Works This section contains the most recent research publications on disease detection using DL models for Citrus illnesses. To recognise illnesses, most citrus detection systems use traditional machine learning (ML) methods. The purpose of this research is to figure © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 320–331, 2022. https://doi.org/10.1007/978-3-030-96299-9_32

Deep Neural Network Model for Automatic Detection

321

out which of the five types of citrus diseases there are: greening, Black Spot, Melanose, Scab and canker. To automatically identify illnesses in grain crops, a system has been proposed [1]. Agricultural productivity makes a substantial contribution to the economy of any country, according to this paper. sorghum, rice, wheat, corn, oats, and millets are all typical grain plants farmed around the world. Information on early diagnosis of many diseases can aid in disease control through correct pest control technology selection to increase grain output For assessing student satisfaction and feedback, a fuzzy-based sentiment analysis system has been suggested. Overall purpose of this study is to assign relevant sentiment scores to opinion phrases and polarity shifters in the input evaluations; a fuzzy-based sentiment analysis system can analyse student feedback and satisfaction [2]. Our method uses a module of fuzzy logic to analyse and perfect assessment of satisfaction after computing the sentiment score of student feedback evaluations. Deep learning for biomedical image reconstruction has been proposed as described in this work, and is a tremendous resource in medicine because it allows us to see into the body of human beings and gives scientists and physicians a plethora of information that is essential for disease knowledge, modelling, therapy and diagnosis [3]. A unified model based on several convolutional neural networks for autonomous grape leaf disease detection has been proposed. Grape diseases are the main causes of serious grape decline in this research. Deep learning algorithms have lately shown outstanding results in a variety of computer vision issues prompting us to apply them to the identification of grape illnesses. In this study, an integrated strategy is used to propose a united convolutional neural networks (CNNs) architecture.

3 Methodology 3.1 Data Collection The photos in the proposed collection come from a variety of public sources, including TEST, which has variety of images of infected and uninfected foliage and fruits. The images utilised for training, validation, and testing in this investigation are listed in Table 1. The sampling images of the proposed dataset are depicted in Fig. 1. Table 1. Particulars of the proposed datasets Image category

Quantity

Training samples

Validation samples

Testing samples

Affected images

115

115

85

35

Uninfected leaf and fruit images

200

200

98

112

322

S. Anandamurugan et al.

Fig. 1. Sampling images from the dataset

3.2 MLP Models Processing digital images is used in a variety of areas of biology to identify and analyse real-world problems. The goal of this method is to apply image processing techniques to detect citrus disease through leaf inspection. The leaf images depict the plant’s health condition and make it easier to monitor and detect disease levels at an early stage. The examination of leaf images is a crucial stage in the diagnosis of a variety of plant diseases. The current method for increasing the clarity and quality of leaf photographs consists of two stages. This method’s preliminary stage employs Recursively Separated Weighted Histogram Equalization (RSWHE), which improves the image’s contrast level. A Median filter method is used in the second stage to remove the unwanted noise. This proposed method uses these strategies to improve image clarity and applies them to the detection of lemon citrus canker disease [4]. Using the Median filter, this can reach to the higher end of conserving, reducing noise, and isolating noise spikes. If half of the numbers are less than m and half are bigger, the median might be a group of numbers in this group. The median m could be a pixel value selected from the neighbourhood sorted distribution values as a middle pixel value. For identifying citrus illnesses into multiple groups, the proposed method used a progressive MLP model, which included Scab, Canker, Black spot, greening, and Melanose [6]. In the deep learning model, this model incorporates a suitable number of layers. By comparing the efficacy of the suggested

Deep Neural Network Model for Automatic Detection

323

model to that of a similar study technique, more sophisticated practical applications in the field of plant disease recognition based on visual symptoms can be developed. Python Python is a programming language that is designed to be simple to read, implement, and use. It is open source, which means it is free to use. It is a high-level language that is used to construct web applications on a server. It has been translated to Java and .NET virtual machines and can run on Windows, Mac, and even Unix platforms. Python, like Ruby or Perl, is a scripting language that is frequently used to create Web applications and dynamic Web content. Python is also supported by a variety of 2D and 3D imaging tools, allowing users to write custom plug-ins and extensions. Anaconda Anaconda Cloud is Anaconda’s package management service. Cloud enables finding, accessing, storing, and sharing public notebooks, environments, and conda and PyPI packages a breeze. For a wide range of applications, Cloud hosts hundreds of helpful Python packages, notebooks, projects, and environments. To search for, download, and install public packages, you do not need to log in or even have a Cloud account. You may use conda-build to create new conda packages, which you can subsequently publish to Cloud to share with others or access from anywhere. You may manage your account via the Anaconda Cloud command line interface (CLI), anaconda-client, which includes authentication, tokens, upload, download, removal, and search. Connect to your Anaconda Cloud account and manage it. Adaptive Image Scaling In widely adopted target detection models, images vary in size, so the standard way is to scale the images to a universal size before feeding them into the detection network. Numerous images may have different aspect ratios during the inspection, the black image borders may be varied after filling and zooming. Additional fillings will cause additional information redundancy, which will slow down the speed. As a result, the letterbox function in the MLP code is adaptively updated to the original image with the least amount of black border, in contrast to the earlier work [8]. The black frames at either end of the image height are reduced, which reduces the number of calculations required during inference, allowing the target detection speed to be significantly improved. The inference speed has increased by 37% due to this simple improvement, which can be considered highly effective. Methodology The proposed methodology is a deep learning method based on a convolutional neural network. In the proposed methodology, MLP CLASSIFIER and GAUSSIAN FEATURE extraction are applied. The proposed model’s performance was evaluated using stateof-the-art methodologies to ensure the highest level of accuracy. The train picture is used as the input, then the grey scale image is transformed, and then the binary image

324

S. Anandamurugan et al.

is calculated, with the illness region being detected with the highest level of accuracy. A feedforward artificial neural network called a multilayer perceptron (MLP) is a type of feedforward artificial neural network (ANN). Each successive layer is made up of a collection of nonlinear functions that are the weighted sum of all the previous layer’s outputs (completely linked) [8]. The Gray Level Co-occurrence Matrix (GLCM) method is a technique for obtaining statistical texture information of second order. Third and higher order textures consider the relationships among three or more pixels, and have been employed in a variety of applications. These are theoretically possible, but due to calculation time and interpretation difficulty, they are rarely used. a) Epoch The number of epochs is a hyper parameter that specifies how many times the learning algorithm will go over the full training dataset. Each sample in the training dataset has an opportunity to update the internal model parameters once each epoch [9]. One or more batches make up an epoch. Consider a for-loop with the number of epochs multiplied by the training dataset for each iteration. Another nested forloop within this for-loop iterates over each batch of samples, each batch having the provided “batch size” number of samples [10] b) Gray Scale Conversion Grayscale images are a type of black-and-white or grey monochrome image made up entirely of shades of grey in image classification. The contrast goes from black to white, with black being the lowest and white being the highest. Grayscale images are distinct from one-bit bi-tonal black-and-white images, which are images with only two colours: black and white in the context of computer imaging. There are various shades of grey in grayscale photographs. When only a single frequency is collected, grayscale images are created by measuring the intensity of light at each pixel according to a certain weighted mixture of frequencies, and they are monochromatic proper [11]. In theory, the frequencies might come from any part of the electromagnetic spectrum. c) Binary Image Classification Because image binarization is such a difficult problem, determining the best threshold value for each scenario is impractical. All picture binarization methods have their own set of flaws and advantages. As a result, rather than a single threshold value, we aim for an algorithm that determines the best binarization approach in this study [6]. Our goal is to take advantage of the great features of various binarization algorithms and apply them wherever they are effective. Using some basic features like standard deviation, mean, and maximum intensity to determine the appropriate binarization approach for every image collection.

Deep Neural Network Model for Automatic Detection

325

4 Experiments 4.1 Training Process The percentage of training data used in the picture training process may vary based on the needs of the experiment. The training data includes both the input and the expected result. Image segmentation is the technique of splitting a digital image into several parts in digital image processing and computer vision. The purpose of segmentation is to make an image more intelligible and easier to examine by simplifying and/or changing its representation. Objects and boundaries in images are often located via image segmentation. Stage 1: Choose the appropriate model and configure the configuration file to match the target objects. Stage 2: Use pre-trained weights to provide starting parameters for the model and expedite the training process. Stage 3: Establish the training parameters and begin the training process. 4.2 Experimental Results The proposed model’s performance is compared to that of current research, and the results are reported. A detailed comparison of published approaches, on the other hand, is difficult for a variety of reasons. To begin, such models were evaluated on a variety of datasets, which made comparison challenging. Furthermore, the articles by the contributing writers present the approaches in an abstracted form with little detail, making them unusable for future scholars. Performance of Each Model Photos from the test dataset as well as real-time images collected on a smartphone or tablet are used to evaluate the proposed MLP model’s detection capability [8]. The photos used comprise single and multiple views of fruit and leaves with and without illness, allowing for a more accurate assessment of the model’s performance under diverse conditions. The effectiveness of each model is evaluated using 3 test cases: • • • • •

Case 1: Individual with disease & without disease from dataset Case 2: Multiple image with and without disease from dataset Case 3: Individual with disease & without disease from real-time images Case 4: Multiple images with and without disease from real-time images Case 5: Real-time images of individuals posed at different angles with and without infection.

326

S. Anandamurugan et al.

These same test cases are applied for image dataset as well as real-time images. The results of the accuracy of various kinds of diseases are listed in the following figures Fig. 2.

Fig. 2. Accuracy chart for various kinds of Citrus diseases

Performance Comparison of the Models The performance of each version’s models has been discussed in detail in the preceding

Deep Neural Network Model for Automatic Detection

327

section. For examining the recognition performance of the recommended model, it is required to compare the prediction performance of the MLP model for disease detection with various variants of the CNN model. As performance evaluation criteria, The proposed method can identify various Citrus diseases in Fruits and Leaves based on MLP

Fig. 3. Detection of Citrus Canker disease

Table 2. Comparison of performance evaluation criteria for the proposed study Algorithm

Accuracy

MLP

94%

CNN

87%

328

S. Anandamurugan et al.

Classifier and GAUSSIAN feature extraction of Deep Learning models. The suggested model produces better categorization accuracy than the CNN model [8]. The high rate of image is successfully trained and the testing sample produces better result then the other machine learning methods and the epoch value generation is high level so that the different kinds of disease can be predicted through the binary image classification. In this study, it was demonstrated that the proposed model ensures recognition accuracy and effectively realizes the lightweight characteristics of the model [4] (Table 2 and Figs. 3, 4, 5, 6 and 7).

Fig. 4. Detection of Melanose disease

Deep Neural Network Model for Automatic Detection

Fig. 5. Detection of greening disease

Fig. 6. Detection of Scab disease

329

330

S. Anandamurugan et al.

Fig. 7. Detection of Black Spot disease

5 Conclusion Healthy and unhealthy Citrus fruits and leaves can be distinguished using the suggested neural network-based leaf disease diagnostic model. In this paper, we use the MLP classifier to solve the problem of diagnosing diseases from photos of citrus fruit and leaves. Our model’s modules are the following: Data capture, data preparation, and MLP classifier are the three steps. In the proposed MLP classifier, two convolutional layers were used. The foremost convolutional layer removes low-level features from the image, while the next convolutional layer groups high-level qualities. As a result, citrus fruit/leaves are classified as diseased. into the Scab, Canker, Black spot, greening, and Melanose classes. In terms of accuracy, the proposed MLP classifier outperformed the other classifiers, reaching a classification accuracy of 94% in citrus fruit/leaf disease investigations.

References 1. Khattak, A., et al.: Automatic detection of citrus fruit and leaves diseases using deep neural network model. https://doi.org/10.1109/Access.2017

Deep Neural Network Model for Automatic Detection

331

2. Manavalan, R.: Automatic identification of diseases in grains crops through computational approaches: a review. Comput. Electron. Agric. 178, 105802 (2020) 3. Yedder, H.B., Cardoen, B., Hamarneh, G.: Deep learning for biomedical image reconstruction: a survey. arXiv [eess.IV] (2020) 4. Ji, M., Zhang, L., Wu, Q.: Automatic grape leaf diseases identification via United Model based on multiple convolutional neural networks. Inf. Process. Agric. 7(3), 418–426 (2020) 5. Zhu, X., He, Z., Du, J., Chen, L., Lin, P., Tian, Q.: Soil moisture temporal stability and spatiotemporal variability about a typical subalpine ecosystem in northwestern China. Hydrol. Process. 34(11), 2401–2417 (2020) 6. Liu, Z., Xiang, X., Qin, J., Tan, Y., Zhang, Q., Xiong, N.N.: Image recognition of citrus diseases based on deep learning. Comput. Mater. Continua 66(1), 457 (2020) 7. Richey, B., Majumder, S., Shirvaikar, M.V., Kehtarnavaz, N.: Real-time detection of maize crop disease via a deep learning-based smartphone app. In: Real-Time Image Processing and Deep Learning 2020 (2020) 8. Singh, H., Rani, R., Mahajan, S.: Detection and classification of citrus leaf disease using hybrid features. In: Pant, M., Sharma, T.K., Verma, O.P., Singla, R., Sikander, A. (eds.) Soft Computing: Theories and Applications. AISC, vol. 1053, pp. 737–745. Springer, Singapore (2020). https://doi.org/10.1007/978-981-15-0751-9_67 9. Barman, U., Choudhury, R.D., Sahu, D., Barman, G.G.: Comparison of convolution neural networks for smartphone image based real time classification of citrus leaf disease. Comput. Electron. Agric. 177, 105661 (2020) 10. Khanramaki, M., Asli-Ardeh, E.A., Kozegar, E.: Citrus pests classification using an ensemble of deep learning models. Comput. Electron. Agric. 186, 106192 (2020). https://doi.org/10. 1016/j.compag.2021.106192 11. Kukreja, V., Dhiman, P.: A Deep Neural Network based disease detection technique for Citrus fruits. In: The 2020 International Conference on Smart Electronics and Communication (ICOSEC) (2020)

Reducing Time Complexity of Fuzzy C Means Algorithm Amrita Bhattacherjee1 , Sugata Sanyal2 , and Ajith Abraham3(B) 1 Department of Statistics, St. Xavier’s College, Kolkata 700016, India 2 School of Technology and Computer Science, Tata Institute of Fundamental Research,

Mumbai 400005, India 3 Machine Intelligence Research Labs (MIR Labs) Scientific Network for Innovation and

Research Excellence, Auburn, WA 98071, USA [email protected]

Abstract. The Fuzzy C-Means clustering technique is one of the most popular soft clustering algorithms in the field of data segmentation. However, its high time complexity makes it computationally expensive, when implemented on very large datasets. Kolen and Hutcheson [1] proposed a modification of the FCM Algorithm, which dramatically reduces the runtime of their algorithm, making it linear with respect to the number of clusters, as opposed to the original algorithm which was quadratic with respect to the number of clusters. This paper proposes further modification of the algorithm by Kolen et al., by suggesting effective seed initialisation (by Fuzzy C-Means++, proposed by Stetco et al. [2]) before feeding the initial cluster centers to the algorithm. The resultant model converges even faster. Empirical findings are illustrated using two synthetic and two real-world datasets. Keywords: Clustering · Fuzzy partitions · Time complexity · Fuzzy C-means algorithm · Unsupervised machine learning

1 Introduction Cluster analysis or clustering is a method of grouping data points into different clusters or categories such that objects within the same cluster are more similar to each other than objects in different clusters. The objects are grouped together based on some similarity measure, which is specified depending on the data at hand and the objective of the task. This method has widespread application, ranging from pattern recognition and market segmentation to image processing and various other fields of data analysis. The Fuzzy C-Means algorithm is one such clustering algorithm, which facilitates soft partitioning of the objects in the dataset. The earliest applications of clustering primarily focused on ‘crisp’ partitions of objects, where each point either fully belongs to a category or does not belong to a category at all. This approach relied on the idea that

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 332–347, 2022. https://doi.org/10.1007/978-3-030-96299-9_33

Reducing Time Complexity of Fuzzy C Means Algorithm

333

an object in a category does not bear any resemblance to any of the categories except to the one it belongs to. Soft partitions, on the other hand, rely on the idea that each object is characterised by the extent to which they belong to all the clusters/categories. A measure of this extent of an object’s resemblance to each cluster is introduced by Zadeh (1965) [11] in the form of what is now known as a ‘membership function’. The final goal is to create partitions or clusters with soft or fuzzy margins. As stated by Bezdek et al. [3]: “A fuzzy c-partition of (the dataset) X is one which characterizes the membership of each sample point in all the clusters by a membership function which ranges between 0 and 1”. The detailed definition of fuzzy c-means (FCM) partitioning and the corresponding algorithm, as proposed by Bezdek et al. [3], is given in Sect. 3.1. The main limitation of this algorithm is its time complexity and memory requirements. The algorithm alternates between estimating cluster centers from the membership matrix and updating the membership matrix based on the cluster centers. As such, the membership matrix, which is of the order of the number of objects to be clustered, is repeatedly accessed and updated, on every iteration. This greatly affects the speed of the algorithm when the dataset is very large. This problem has been widely addressed in the literature. This paper focuses on the modification proposed by Kolen and Hutcheson (2002) [1], where the membership matrix is not generated (or updated) iteratively. This modification generates an algorithm which has a time complexity of O(ncp) as opposed to Bezdek’s original FCM Algorithm, which had a time complexity of O(nc2 p), where n is the number of objects in the dataset, c is the number of clusters and p is the number of features of each object/point in the data. Let us call this algorithm FCM-U, where U refers to the membership matrix. This paper employs the FCM-U algorithm and pairs it with the popular approach of effective seed initialisation for even faster convergence. Here, the FCM++ algorithm (proposed by Stetco et al. [2]) is implemented for effective seed initialization. On clubbing these two algorithms together, the model runs faster and empirically converges earlier than the FCM-U algorithm. The following section discusses some related works in reducing time complexity of the FCM Algorithm, followed by short descriptions of the original FCM algorithm, the FCM++ approach and the FCM-U algorithm. Then, the proposed model is defined, followed by a comparative analysis of the results obtained when this algorithm is employed for clustering datasets. Finally, some further scopes of improvement are discussed.

2 Related Works Several researchers have proposed methods to tackle the problem of high computational cost that comes with implementation of the Fuzzy C-Means algorithm. In 1986, Cannon, Dave and Bezdek [4] proposed an Approximate Fuzzy C-Means algorithm where the exact variates in the equation are replaced with integer/real-valued estimates. Tolias and Panas [5] applied spatial constraints on image segmentation problems using a fuzzy rule-based system, which showed reduced computational time.

334

A. Bhattacherjee et al.

In 1994, Kamel and Selim [6] proposed two algorithms that converged faster than the FCM algorithm, having adopted a continuous process where the algorithm starts updating the membership values as soon as a part of cluster centers are updated. In 1998, Cheng et al. [7] proposed a multi-stage random sampling approach where the cluster centers are estimated after taking repeated random samples from the data. Then, the centroids are initialised over the entire data. This process reported a speed-up of 2–3 times than the original algorithm. Hore et al. [8] proposed a single-pass fuzzy c-means algorithm using weighted point calculation. In 2002, Kolen and Hutcheson [1] proposed a modification which eliminates the task of repeatedly updating the membership matrix, this reducing the time complexity to a linear function of the number of clusters; as opposed to the original algorithm which was a quadratic function of the number of clusters. This was particularly beneficial for large datasets. In fact, this paper implements this approach in the proposed algorithm along with effective seed initialisation. Another angle of attack adopted by researchers is manipulating the data itself. Hung and Yang [9] proposed the psFCM algorithm which used a simplified subset of the original data to speed up the convergence. Several approaches were made to eliminate initial bias and reduce the time taken for convergence of the FCM algorithm. These research works mainly focused on modifying the initial centroids which are passed to the algorithm. Effective seed initialisation shows promising result in removing initial bias of the FCM algorithm. In 2015, Stetco, Zeng and Keane [2] extended the idea of K-Means++ [10] algorithm into the standard version of Fuzzy C-Means.

3 Fuzzy C-Means (FCM) Algorithm and Its Variants Let X = {X1 , X2 , . . . , Xn } be a set of n points in Rp , the p-dimensional Euclidean space. For 1 ≤ c ≤ n, c ∈ N , the set of natural number, a fuzzy c-partition of X is represented by (U , X ) where, U is a matrix of order n × c, that is –   U = uij n×c where, uij denotes the membership value of the ith point in X to the j th fuzzy set. Here, 1 ≤ i ≤ n and 1 ≤ j ≤ c. The values of the membership matrix are subject to the following conditions: 1.  0 ≤ uij ≤ 1, ∀i, j c u = 1, ∀i 2. j=1  ij 3. 0 < ni=1 uij < n, ∀j

Reducing Time Complexity of Fuzzy C Means Algorithm

335

The FCM algorithm defines a constant m, which is called the fuzziness parameter and corresponds to the degree of fuzziness of the clusters. By convention, we take m > 1. The FCM Algorithm then defines ‘cluster centers’ vj , 1 ≤ j ≤ c as: n

m i=1 xi uij m i=1 uij

vj = n

(1)

The membership function is typically defined as:  uij =

 2 c   dij m−1

k=1

dik

−1 ,

(2)

for 1 ≤ i ≤ n and 1 ≤ j ≤ c where, dij = xi − vj  is the distance of the ith point in X to the jth cluster center. The cost function is defined as: Jm (U , V ; X ) =

c n

ujim xi − vj 2

(3)

i=1 j=1

Therefore, the Fuzzy C-means algorithm as proposed by Bezdek is given by:

3.1 Effective Seed Initialization and Eliminating the U-Matrix The Fuzzy C-Means++ algorithm as proposed by Stetco et al. uses effective seed initialisation to determine the starting values for the FCM algorithm. Before stating the algorithm, we state some notations:

336

A. Bhattacherjee et al.

They defined a value Pi , corresponding to the ith data point in X, given by: Pi =

d s (xi , V ) sum(d s )

where, d s (xi , V ) denotes the distance (raised to the power s) from a point xi ∈ X to its closest representative in R. The value of s controls the spreading factor of the algorithm. A small value of s will choose centers which are very close to each other, whereas a very large value of s might lead to the choice of outliers as cluster centers. When s is taken to be zero, the algorithm reduces to random seed initialisation. Further, the first point is randomly chosen and determines the selection of all the other centers. With the values and parameters defined above, the FCM++ algorithm by Stetco et al. is as given in Algorithm 2.

We now state the algorithm as proposed by Kolen et.al. which constitutes the main body of the algorithm. In 2002, John Kolen and Tim Hutcheson proposed a modification in the algorithm which reduced the time of computation drastically. They eliminated the storage of the membership matrix at every iteration, and directly computed the updated cluster centers and is detailed in Algorithm 3. Notations:

Reducing Time Complexity of Fuzzy C Means Algorithm

337

3.2 Proposed Algorithm This paper implements an algorithm which combines the previous methods into a single implementation. In other words, we first generate a prototype matrix using effective seed initialisation (FCM++), and then use this initial prototype matrix as the starting point of the algorithm as stated in Sect. 3.1. Additionally, some modifications were made so that the algorithm works even when the cluster centers are points from the dataset itself. The algorithm is as stated below:

338

A. Bhattacherjee et al.

4 Experimental Results Kolen and Hutcheson [1] illustrated the performance impacts of their modification in great detail. The algorithm implemented in this paper shows further improvement in computation speed owing to effective seed initialisation. The results are illustrated on 4 datasets – the Iris Dataset [12], Wine Dataset [13] and 2 synthetic datasets generated from gaussian distributions. The time for convergence (to reach the same cost value) was measured (in seconds) for both the original FCM algorithm and the proposed modified algorithm while varying the number of clusters. The empirical findings are Tabulated in Tables 1, 2, 3 and 4. Table 1 indicate that the proposed algorithm provides a considerable gain in time due to faster convergence with the same cost value. The time taken for convergence is plotted in Fig. 1. The black points are the time taken (in seconds) by the original algorithm, plotted against the number of clusters specified to the algorithm. To compare the rate of change in time taken for each algorithm, a simple linear regression is fitted for each of them. The following graph gives a visual representation of the results obtained.

Reducing Time Complexity of Fuzzy C Means Algorithm

339

Table 1. Time for convergence for the Iris dataset Number of clusters

Algorithm used Original FCM

Proposed FCM

2

0.058

0.021

3

0.147

0.047

4

0.193

0.072

5

0.299

0.093

6

0.384

0.101

Fig. 1. Iris dataset performance

The regression equations obtained are: TimeOriginal = (0.0804 × N ) − 0.1054 TimeProposed = (0.0206 × N ) − 0.0156 where, N represents the number of clusters. The regression is done keeping the number of features in the dataset constant. It can be noted visually from the graph that the time taken by the original algorithm is consistently higher than that by the proposed algorithm. In addition, the rate of increase in time as the number of clusters increases can be obtained from the regression equations as follows: Slope for Original Algorithm = 0.0804 Slope for Proposed Algorithm = 0.0206 Clearly, the rate of increase in time for a unit increase in the number of clusters is approximately 4 times higher for the original algorithm than that for the proposed

340

A. Bhattacherjee et al.

algorithm. This validates a considerable amount of savings in time, especially for higher number of clusters. The time taken are recorded while keeping the cost value constant for a given number of clusters, which enables a fair comparison. The cost is calculated using (3). For perspective, the performance of the proposed algorithm in predicting the correct clusters can be visually estimated by looking at the following graphs. Figure 2 represents the true clusters as available in ground truth labels of the dataset.

Fig. 2. Iris dataset: True and predicted clusters (red stars indicate the predicted cluster centers)

The Wine Dataset contains data on the results of a chemical analysis of 3 different types of wine grown in the same region in Italy. The 13 different features for each datapoint are actually the amount of each of the 13 different constituents found in the analysis. The attributes are real-valued numbers. There is a total of 178 datapoints. The time for convergence (to reach the same cost value) was measured (in seconds) for both the original FCM algorithm and the proposed modified algorithm while varying the number of clusters and the results are depicted in Table 2. Table 2. Time for convergence for the Wine dataset Number of clusters

Algorithm used Original FCM

Proposed FCM

2

0.184

0.074

3

0.619

0.321

4

0.794

0.288

5

2.493

0.504

6

2.501

0.915

Th proposed algorithm, once again, shows significant economy in terms of time taken till convergence. A similar study is done to obtain simple linear regression equations for

Reducing Time Complexity of Fuzzy C Means Algorithm

341

each of the algorithms. The regression lines are plotted against the number of clusters in Fig. 3.

Fig. 3. Wine dataset performance

The regression equations obtained are: TimeOriginal = (0.6508 × N ) − 1.285 TimeProposed = (0.1865 × N ) − 0.326 where, N represents the number of clusters. The regression is done keeping the number of features in the dataset constant. It can be noted visually from the graph that the time taken by the original algorithm is consistently higher than that by the proposed algorithm. In addition, the rate of increase in time as the number of clusters increases can be obtained from the regression equations as follows: Slope for Original Algorithm = 0.6508 Slope for Proposed Algorithm = 0.1865 Here, the rate of increase in time for a unit increase in the number of clusters is approximately 3.5 times more for the original algorithm than that for the proposed algorithm. The time taken are recorded while keeping the cost value constant for a given number of clusters, which enables a fair comparison. The cost is calculated using (3). Figure 4 illustrates the true clusters and the predicted clusters for the Wine dataset. Isotropic gaussian blobs are generated using Python’s Scikit-learn library. The dataset generated for this problem contains 3 clusters where cluster centers are generated at random from the interval (−10, 10). The standard deviation for each cluster is set at 1 (to maintain homoscedasticity). The random state is fixed at ‘0’. Under the above conditions, 300 points are generated, each having 3 features. The points are plotted on a 2-dimensional space for visualisation in Fig. 5.

342

A. Bhattacherjee et al.

Fig. 4. Wine data set: True and predicted clusters (red stars indicate the predicted cluster centers)

Fig. 5. Isotropic gaussian blobs

The time for convergence (to reach the same cost value) was measured (in seconds) for both the original FCM algorithm and the proposed modified algorithm while varying the number of clusters and the results are illustrated in Table 3 and the time taken over the number of clusters is depicted in Fig. 6. Referring to Fig. 6, the regression equations obtained are: TimeOriginal = (0.197 × N ) − 0.345 TimeProposed = (0.064 × N ) − 0.115 where, N represents the number of clusters. The number of features in the dataset is kept constant. It can be noted visually that the time taken by the original algorithm is consistently higher than that by the proposed algorithm. In addition, the rate of increase

Reducing Time Complexity of Fuzzy C Means Algorithm

343

Table 3. Time for convergence for the Gaussian dataset (Type 1) Number of clusters

Algorithm used Original FCM

Proposed FCM

2

0.115

0.043

3

0.126

0.037

4

0.461

0.126

5

0.642

0.225

6

0.840

0.267

Fig. 6. Isotropic gaussian blobs dataset performance

in time as the number of clusters increases can be obtained from the regression equations as follows: Slope for Original Algorithm = 0.197 Slope for Proposed Algorithm = 0.064 Here, the rate of increase in time for a unit increase in the number of clusters is approximately 3 times more for the original algorithm than that for the proposed algorithm. Hence, we can conclude that the proposed algorithm facilitates a significant amount of savings in time to converge to the same clustering result. Figure 7 illustrates the true and predicted clusters of this simulated dataset. As illustrated in Fig. 8, samples from 4 gaussian distributions of varying means and standard deviations are taken to create overlapping clusters. For this particular evaluation, the means of the 4 distributions are taken as (−3,1), (2,2), (1,−3) and (5,4) with respective standard deviations 1, 0.5, 1.5 and 2 respectively. 250 points are generated from each of these distributions (making a total of 1000 datapoints). The time for convergence (keeping the cost same) is measured in seconds for both the original and the

344

A. Bhattacherjee et al.

Fig. 7. Isotropic gaussian blobs dataset: True and predicted clusters (red stars indicate the predicted cluster centers)

proposed algorithm are depicted in Table 4 and the clustering results are illustrated in Figs. 9 and 10.

Fig. 8. Gaussian Dataset (Type 2)

Table 4. Time for convergence for the Gaussian dataset (Type 2) Number of clusters

Algorithm used Original FCM

Proposed FCM

2

0.738

0.321

3

1.973

0.619

4

1.264

0.442

5

5.288

0.910

6

12.016

2.402

Reducing Time Complexity of Fuzzy C Means Algorithm

345

The time taken by each of the 2 algorithms is regressed separately on the number of clusters, and two regression equations are obtained. Note that even though the regression line seems to suggest that, for 2 clusters, proposed algorithm takes more time than the original algorithm, it can be seen from the plotted points that, in the data, the proposed algorithm does in fact take less time for all clusters. The regression equations obtained are: TimeOriginal = (2.587 × N ) − 6.093 TimeProposed = (0.445 × N ) − 0.842 where, N represents the number of clusters. The number of features in the dataset is kept constant. It can be noted that the rate of increase in time as the number of clusters increases can be obtained from the regression equations as follows – Slope for Original Algorithm = 2.587 Slope for Proposed Algorithm = 0.445 Here, the rate of increase in time for a unit increase in the number of clusters is approximately 5 times more for the original algorithm than that for the proposed algorithm, which is especially pronounced for high number of clusters. Hence, we can conclude that the proposed algorithm facilitates a significant amount of savings in time to converge to the same clustering result.

Fig. 9. Gaussian Dataset (Type 2) performance

For visualisation, the true clusters are plotted below, followed by a graph illustrating the predicted clusters –

346

A. Bhattacherjee et al.

Fig. 10. Gaussian datset (Type 2): True and predicted clusters

5 Conclusions Comparative analyses of the time taken for Algorithm 2 and Algorithm 3, when implemented individually are already elaborated in [3] and [2] respectively. This paper combined these algorithms and compared its performance with the original Fuzzy C-Means algorithm to empirically confirm that it indeed accelerates the speed of the algorithm, which becomes more evident for larger datasets and higher number of clusters. In fact, the cluster accuracy stays intact (and in some cases, improves over the original FCM algorithm). Empirical results indicate faster convergence with very high cluster accuracy (as confirmed by Adjusted Rand Index during runtime). One can be interested in tailoring the algorithm to the specific data in hand. In this context, feature normalisation, feature engineering, sampling from the dataset could be viable options for further speeding up the convergence. The FCM algorithm largely depends on the initial centers selected. Further attempts could be made to eliminate the initial bias to ensure that the algorithm converges to a better solution. FCM++ has been proven to be a good approach in this context. However, testing other methods of effective seed initialisation (preferably along with Hutcheson and Kolen’s [1] algorithm) might yield promising results. Combining other time-reduction approaches like random sampling of the datapoints or multi-stage random sampling [7] have been proven to be very successful. Pairing this strategy with the proposed algorithm is expected to perform extremely well for large datasets. Another open field of application is image segmentation. FCM algorithm finds manifold implementations in image segmentation problems, where the image sizes are quite high. In such a scenario, modifying the algorithm to accommodate image data and effectively reducing its runtime will open new avenues. The authors of this paper are looking into a similar implementation on image data, and tailor the time complexity reduction approach towards image-segmentation problems.

Reducing Time Complexity of Fuzzy C Means Algorithm

347

References 1. Kolen, J.F., Hutcheson, T.: Reducing the time complexity of the fuzzy c-means algorithm. IEEE Trans. Fuzzy Syst. 10(2), 263–267 (2002) 2. Stetco, A., Zeng, X.-J., Keane, J.: Fuzzy C-means++: fuzzy C-means with effective seeding initialization. Expert Syst. Appl. 42(21), 7541–7548 (2015) 3. Bezdek, J.C., Ehrlich, R., Full, W.: FCM: the fuzzy c-means clustering algorithm. Comput. Geosci. 10(2–3), 191–203 (1984) 4. Cannon, R.L., Dave, J.V., Bezdek, J.C.: Efficient implementation of the fuzzy c-means clustering algorithms. IEEE Trans. Pattern Anal. Mach. Intell. 2, 248–255 (1986) 5. Tolias, Y.A., Panas, S.M.: On applying spatial constraints in fuzzy image clustering using a fuzzy rule-based system. IEEE Signal Process. Lett. 5(10), 245–247 (1998) 6. Kamel, M.S., Selim, S.Z.: New algorithms for solving the fuzzy clustering problem. Pattern Recognit. 27(3), 421–428 (1994) 7. Cheng, T.W., Goldgof, D.B., Hall, L.O.: Fast fuzzy clustering. Fuzzy Sets Syst. 93(1), 49–56 (1998) 8. Hore, P., Lawrence, O.H., Dmitry, B.G.: Single pass fuzzy c means. In: 2007 IEEE International Fuzzy Systems Conference. IEEE (2007) 9. Hung, M.-C., Yang, D.-L.: An efficient fuzzy c-means clustering algorithm. In: Proceedings 2001 IEEE International Conference on Data Mining. IEEE (2001) 10. Arthur, D., Sergei, V.: k-means++: The Advantages of Careful Seeding. Stanford (2006) 11. Zadeh, L.A.: Fuzzy sets. Inf. Control 8(3), 338–353 (1965) 12. https://archive.ics.uci.edu/ml/datasets/iris. Accessed 20 Nov 2021 13. https://archive.ics.uci.edu/ml/datasets/wine. Accessed on 20 Nov 2021

Information and Communication Technologies

Kubernetes for Fog Computing Limitations and Research Scope R. Leena Sri1 and Divya Vetriveeran2(B) 1 2

Thiagarajar College of Engineering, Madurai, India [email protected] CHRIST (Deemed to be University), Bangalore, India [email protected]

Abstract. With the advances in communications, Internet of Everything has become the order of the day. Every application and its services are connected to the internet and the latency aware applications are greatly dependent on Fog Infrastructure with the cloud as a backbone. With these technologies, orchestration plays an important role in coordinating the services of an application. With multiple services contributing to a single application, the services may be deployed distributed in multiple server. Proper coordination with effective communication between the modules can improve the performance of the application. This paper deals with the need for orchestration, challenges, and tools with respect to edge/fog computing. Our proposed research solution in the area of intelligent pod scheduling is highlighted with the possible areas of research in Microservices for Fog infrastructure.

1

Introduction

Applications today revolve around the internet and the devices around us are also being connected to it. It is estimated that there are 50 billion connected devices in 2020 [1]. As the world moves towards Internet of Things (IoT), so has the need for Quality of service and Quality of Experience. As the data and applications grow exponentially, access to the application is also being done remotely from a distributed environment. Almost all application providers prefer their application to be fully or partially decentralized. This not only helps in fault handling but also helps diversify the system and promote better control and supervision over the system. With the large use of IoT systems almost everywhere in any tech industry, a large application is split into smaller runtime modules for easy deployments. These modules or business logic have to be distributed with proper interaction between them and satisfy the networking, binding, and state management primitives. Alongside IoT, there has also been growth in the computing paradigm Edge/Fog Computing [2]. With this paradigm in use for multiple Industrial IoT applications, the modules deployed in production are stateless with openstandards, and interoperability. Each module may be developed using different c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 351–361, 2022. https://doi.org/10.1007/978-3-030-96299-9_34

352

R. Leena Sri and D. Vetriveeran

technologies. The application or service may run on different components of a large system to support resource isolation, scaling, configuration management, fault handling, and so on. For any distributed system, the 4 major pillars are Lifecycle, Networking, Binding, and state [3]. The properties of the pillars are elaborated in Table 1. Table 1. Pillars of distributed applications Pillar

Properties

Lifecycle

Deployment/rollback, Placement/Scheduling, Configuration, Resource/failure isolation, Auto/manual Scaling, Hybrid workloads

Networking Service discovery and failover, Dynamic traffic routing, Retry, timeout and circuit breaking, Security, rate limiting, encryption, Observability and tracing Binding

API connectors, Protocol conversions, Message transformations, Filtering, light message routing, Point to point or sun interactions

State

Workflow management, Temporal scheduling, Distributed caching, Idempotency, Transactions, Application states

With these requirements on an application, almost all industries have migrated their monolithic application to a distributed environment and some have even made it to Microservices architecture. The use of containers is now being extensively considered by service providers for easy deployment and monitoring. Dockers and cluster management using Kubernetes are being widely used in both research and production industry. A comparison of the two major tools Swarm and Kubernetes has been made. The advantages of the use of Kubernetes have been analyzed with results from various researchers along with scope and directions for further improvements in the area. The paper is structured as follows. Section 2 gives the need for container orchestration followed by the use of Microservices in production in Sect. 3. The comparison of the two orchestrator engines is given in Sect. 4 followed by its use in edge infrastructures in Sect. 5. Section 6 gives the analysis on the major researches in the area using kubernetes and the parameters taken up for improvement. Section 7 gives our proposed research solution for intelligent scheduling. The research directions are discussed in Sect. 8.

Kubernetes for Fog Computing - Limitations and Research Scope

353

Table 2. Docker swarm and Kubernetes Property

Docker Swarm

Kubernetes

1. Deployment Deployment is a YAML based container specification

Deployment is a combination of pods

2. Installation

Easy installation and flexible node addition

Manual Installation and varies with OS

3. Navigation

Command Line Interface based navigation

Command Line Interface based navigation

4. Application monitoring

Open API or Reimann based monitoring

Heapster/Grafana/Influx based monitoring

5. Scalability

Faster deployment and scaling on This is an all-in-one framework for demand distributed systems and is slow in deployment and scaling

6. Availability

Ensured by replication

Has a more efficient failure handling, load balancing and replica management system

7. Networking

Overlay network based communication

IP based flat networking

8. Overall working

Limited functionality, fault handling, and manual scaling. Lightweight, easy to learn, and integrates smoothly with Docker images

Complex installation, complex toolset. Huge community to learn and support, modular and OS friendly, efficient organisation and powerful

2

Need for Container Orchestration

With almost all companies hosting their application on cloud using DevOps, the collaborative efforts and continuous development and testing have been made easier using Microservices. It has helped in service isolation and running them distributively over the internet. To reduce the dependencies and effective application performance, orchestration plays an important role in every production especially in resource-constrained environments. The applications with mobile nodes, the orchestrator takes the responsibility of addition and deletion of a node in the cluster, job scheduling to nodes for quick responses to the services and container placements for dependency aware computations. The major role of the orchestration engine is the role mapping and have high availability of the master, workers, and the unscheduled pods in case of failures or dynamic loads. The engine must also ensure that there are no frequent service migrations as it may lead to performance issues, especially in edge/fog environments.

3

Microservices in Industry

Are Microservices only at the research and academia level? Or are they being used in real-world industries? The answer is a big Yes! The major industries that

354

R. Leena Sri and D. Vetriveeran

influence research and technologies like Amazon, Cloud Foundry Diego, Google, IBM, RedHat, and so on have already been using Kubernetes for deployments and moving their workloads to the cloud. Kubernetes lies on the portability layer of an application to automate the deployment and orchestration of the services in an application. The impact of usage in the industry is high such that 78% of the major companies have been using Kubernetes by the year 2019, and it is no more a new technology to be experimented with for deployment. The features such as improved productivity, stable application execution, ability to handle large computing resources, cost optimization, Effective migration to the cloud, and faster time to the market have made this technology into the business arena.

4

Docker Swarm and Kubernetes

The two major container orchestration tools in the industry are Docker swarm and Kubernetes. There is always a debate on the comparison of these technologies. An overview of the comparison of the two technologies is given in Table 2. Though the two technologies are beyond compare in their own ways, larger communities make use of Kubernetes for its full-fledged support within a single framework and large cluster set up at ease. Background on Kubernetes: Modern applications are being spread across machines which can be servers, virtual or physical machines, or even on cloud platforms. Manual administration of the application services and manual intervention may not be suitable for dynamic applications. Thus, automated cluster management is highly necessary to deploy and scale the Microservices. This can be realized using Kubernetes. The generic architecture of Kubernetes is given in Fig. 1.

Fig. 1. Kubernetes architecture

Kubernetes for Fog Computing - Limitations and Research Scope

355

The master node receives its instructions via the APIs which takes inputs from internal components as well as external users. Any request (pod) received is pushed on to the scheduler which is assigned to healthy nodes and if none are found, the pod is put to a pending state and assigned on availability. The etc.d cluster maintains the backup of all the clusters in the application and the number of backup nodes can be defined by the application developer. The controller is the decision maker of the system which acts based on the state changes of the cluster nodes. This master node communicates with the worker nodes which are accessible to the users. The kubelet is the major component here, which has the node configurations which is highly essential for the master node to decide on the task allocation. The dockers work well with Kubernetes and any container repository for that matter is connected to the container runtime where the images are pulled from. The IP address for each node is assigned by the kube-proxy which is populated into the local iptables for efficient routing. Of all the components, the pods are the components that make a container a part of the cluster. The scalability is ensured by the pods and it serves as a wrapper of the business logic. The pods are not only created on demand but there are also standby pods that are invoked in case of failures or during load balancing. Not only containers but a VM can also be deployed in Kubernetes as a pod. A sample pod creation snippet is given in Fig. 2.

Fig. 2. Pod creation

Kubectl is the default CLI tool for Kubernetes which has similar commands as that of Dockers and is most commonly used. Some of the other famous Kubernetes monitoring tools include cAdvisor, Twistlock, Falco, Kubespray, and Kubeshell.

5

Kubernetes for Edge

As the edge networks are resource constrained, having heavy models or services running affects the performance of the application. The most common

356

R. Leena Sri and D. Vetriveeran

edge device experimented on is the Raspberry Pi 3 which is used by almost all academic researchers. KubeEdge [4] has been specially designed for these infrastructures. Though it is in the early research stage, it has been quite useful in container orchestration at the edge. The cloud API communicates with the Kubernetes API at the edge. The edge is responsible for container deployment and the MQTT takes care of the seamless communication. The other orchestrator is the K3S [5] which has been modified for edge orchestration. This tool eliminates unnecessary functionalities making it lightweight with its thin wrapper for shell command support. Since it need not be built from the scratch it is tightly coupled and the users cannot change all its internal workings. MicroK8s is a single-package Kubernetes used by researchers for edge/fog and IoT orchestration. This has an edge over the other tools for prototyping, and testing.

6

Kubernetes and Distributed Computing

In several cases of internet applications, there is always a need for resource autoscaling. Based on the changes in the application dynamics, the VMs or any nodes have to be allocated on demand and have to be well-orchestrated with the existing framework. Table 3. Orchestration research Literature Purpose

Work

[6]

Auto scaling to ensure QoS

Significant improvement in the Default Kubernetes auto-scaler of K8s

[7]

Cloud orchestration

Automated distribution and federation of services for a cloud service provider

[8]

Availability Management

Though various failure recovery scenarios can be modeled, repair actions were not sufficient for high availability

[9]

Edge IoT based construction application using virtual clusters

Kubernetes based webserver management for application utility

[10]

Management of cloud infrastructure services

Recovery rate of Kubernetes was much faster than that of a VM

Provisioning an entire VM with its own set of libraries and OS can be a heavy task when compared to a container. Thus in the case of large latency aware applications, containers are preferred to VMs. Threshold-based scaling is highly prevalent in the industry which is decided based on the CPU utilization

Kubernetes for Fog Computing - Limitations and Research Scope

357

[11]. It helps in both the addition and removal of instances based on demand. Some of the prominent work using Kubernetes to have better service over the internet are given in Table 3. The creation and deletion of services can also cause latency in the case of a very large cluster. The work by the authors [12] shows the following cluster latency comparison as given in Fig. 3.

Fig. 3. Latency comparison

As the figure describes, the time taken for retrieval on deletion which may also be due to failure is high. Similarly, On the change of computational node and the availability of DNS for networking also increases. If this is the case for two worker node, the latency can be high for large clusters with higher number of worker nodes. Given the geographical distribution of the services, there has been research by the authors [13] have worked upon the placement policies for better application performance. Their system is designed to take up 529 requests per second in the worst case. They have achieved a throughput of s 50x103 operations per seconds The other parameters of the work are summed up as follows, 1. 2. 3. 4.

Average Average Average Average

Pods CPU Utilization: 86%. CPU limit: 80.32%. pods: 1.62. node utility: 1.34.

The results on waiting time for large container scheduling by [14] is given as in Fig. 4. The work involves the replacement of the scheduling strategy with the default kubelet methodology. This shows that there is still room for improvement in the default parameters and strategies of kubernetes. The work also touches on the aspects of power consumption and the changes in the number of active nodes with respect to the load in the application or the number of containers spawned.

358

R. Leena Sri and D. Vetriveeran

Fig. 4. Waiting time comparison

7

Proposed Research Solution

Based on our previous works on scheduling in Fog Infrastructure [15], and the available literature for containerization, our suggestions on the intelligent scheduling is as follows. The default scheduler is limited to the properties of the CPU utility of the nodes and the topology or the bandwidth is not considered for scheduling. Similarly, the use of unscheduled pods across the nodes is not considered making it less optimum in terms of power consumption. This leaves behind the option of a global optimum since they are chosen at random and the properties are not fully exploited. This can be handled as follows, • Priority based scheduling considering more pod parameters can be done in addition to the default scheduling. • Dynamic scheduler considering the present pod configuration, incoming requests, and the unscheduled pods. This can be topology aware so that the pods are scheduled based on requirements and not at random. This way of scheduling can induce more flexibility and control over the system making it customizable based on the application needs. The services can be split rather than making application modules that can be deployed on the edge or fog for better performance and Quality of Experience.

8

Research Directions

The area of scalability has been explored widely in kubernetes but some areas that can still be improved concerning orchestration for distributed systems. Table 4 shows the areas of research that can be explored with respect to orchestration tools.

Kubernetes for Fog Computing - Limitations and Research Scope

359

Table 4. Kubernetes and research Tool

Research scope

Kubernetes [16, 17]

Availability, QoS, utilization and throughput

Docker Swarm [18, 19] Scalability, availability, utilization and throughput Mesos [20, 21]

Scalability and availability

Aurora [22]

Scalability and availability

Orchestration: In addition to all these, the first research challenge that arise is the element modeling for orchestration and deciding on the deployment to handle the diverse environment. This becomes complicated as the environment extends to IoT, Fog and Cloud elements with heterogeneous nodes. This is the major research scope of AWS Greengrass and Azure IoT [23,24]. Resource Discovery: The next challenge to be taken care is the discovery of the newly added services to the system and immediate task allocation for high availability. This has to be worked upon to achieve the least possible latency. Path Discovery: As the resources are added and the topology becomes dynamic especially in the mobile environment, the path discovery and routing become a challenge. The IoT and Fog make the environment versatile and the information changes rapidly. The challenge is to design a system keeping account of the priority, availability, and mobility of the system. Interoperability: As the degree of heterogeneity increases, interoperability has to be ensured among all the components in the architecture. The authors [25] suggest “‘translators, standard interfaces and ontologies”’ with scope for improvement in the area. The overhead due to the introduction of any broker elements also have to be handled to ensure the Quality of Experience to the user. Latency and Resilience: As of any system, these are the two most important factors that are beyond compromise. High interaction between the components or frequent role transfer has to be managed in case of a large application with numerous Microservices involved. Optimization: As the world moves towards Internet of Everything, the applications become geographically distributed with multiple dashboards for remote access. The resources, services and the data utility have to be optimized for an industry standard application. Service migration, decisions on data transfer, and scheme managements have to be standardized and managed efficiently. Security and Privacy: When an application involves Microservices, orchestration, mobility, and heterogeneity the need for privacy preservation becomes an integral part of it. As the number of containers and the interaction between them becomes essential, the security breaches also increase. This has become the most important research aspect and is also a need of the hour.

360

9

R. Leena Sri and D. Vetriveeran

Conclusion

To sum up, Edge/Fog has been considered an ideal platform for IoT applications and the use of containers has been increasing in production for better application development and deployment. In this work, we have summed up the use of Microservices, it’s the research trends that can be taken up by the researchers. The tools and the experimental setup suggestions are also given for ease of study in the field. The research directions can help academicians or researchers to look for opportunities in choosing the right parameters for research. We propose to carry forward our research in the direction of the proposed scheduling solution for latency aware applications with huge data transfer and container requirements.

References 1. Byers, C., Wetterwald, P.: Ubiquity symposium: the internet of things: fog computing: distributing data and intelligence for resiliency and scale necessary for IoT. Ubiquity 11, 1–12 (2015) 2. Pham, Q.-V., et al.: A survey of multi-access edge computing in 5G and beyond: Fundamentals, technology integration, and state-of-the-art. IEEE Access 8, 116974–117017 (2020) 3. Ibryam, B.: The Evolution of Distributed Systems on Kubernetes (2021). https:// www.infoq.com/articles/distributed-systems-kubernetes/. Accessed 24 Mar 2021 4. Kubeedge: A kubernetes native edge computing framework (2019). https:// kubeedge.io/en/ 5. k3s, “5 less than k8s,” (2019). https://github.com/k3s-io/k3s. Accessed May 2019 6. Taherizadeh, S., Grobelnik, M.: Key influencing factors of the Kubernetes autoscaler for computing-intensive microservice-native cloud-based applications. Adv. Eng. Softw. 140, 102734 (2020) 7. Kim, D., Muhammad, H., Kim, E., Helal, S., Lee, C.: Key influencing factors of the Kubernetes auto-scaler for computing-intensive microservice-native cloudbased applications. Appl. Sci. 9(1), 191 (2019) 8. Vayghan, L.A., Saied, M.A., Toeroe, M., Khendek, F.: Kubernetes as an availability manager for microservice applications. arXiv preprint arXiv:1901.04946 (2019) 9. Kochovski, P., Stankovski, V.: Supporting smart construction with dependable edge computing infrastructures and applications. Autom. Constr. 85, 182–192 (2018) 10. Kang, H., Le, M., Tao, S.: Container and microservice driven design for cloud infrastructure DevOps. In: 2016 IEEE International Conference on Cloud Engineering (IC2E), pp. 202–211. IEEE (2016) 11. Dube, P., Gandhi, A., Karve, A., Kochut, A., Zhang, L.: Scaling a cloud infrastructure. US Patent 9,300,553, 29 March 2016 12. Yaguache, F.R., Ahola, K.: Enabling edge computing using container orchestration and software defined wide area networks. In: 9th International Conference on Computer Science, Engineering and Applications, CCSEA 2019, pp. 353–372. AIRCC Publishing Corporation (2019) 13. Rossi, F., Cardellini, V., Presti, F.L., Nardelli, M.: Geo-distributed efficient deployment of containers with Kubernetes. Comput. Commun. 159, 161–174 (2020)

Kubernetes for Fog Computing - Limitations and Research Scope

361

14. Menouer, T.: KCSS: Kubernetes container scheduling strategy. J. Supercomput. 77(5), 4267–4293 (2021). https://doi.org/10.1007/s11227-020-03427-3 15. Divya, V., Sri, L.R.: Fault tolerant resource allocation in fog environment using game theory-based reinforcement learning. Concurr. Comput. Pract. Experience 33(16), 1–22 (2021) 16. Vlasov, Y., Illiashenko, O., Uzun, D., Haimanov, O.: Prototyping tools for IoT systems based on virtualization techniques. In: 2018 IEEE 9th International Conference on Dependable Systems, Services and Technologies (DESSERT), pp. 87–92. IEEE (2018) 17. Zhong, Z., Buyya, R.: A cost-efficient container orchestration strategy in Kubernetes-based cloud computing infrastructures with heterogeneous resources. ACM Trans. Internet Technol. (TOIT) 20(2), 1–24 (2020) 18. Seiber, C., Nowlin, D., Landowski, B., Tolentino, M.E.: Tracking hazardous aerial plumes using IoT-enabled drone swarms. In: 2018 IEEE 4th World Forum on Internet of Things (WF-IoT), pp. 377–382. IEEE (2018) 19. Serhani, M.A., El-Kassabi, H.T., Shuaib, K., Navaz, A.N., Benatallah, B., Beheshti, A.: Self-adapting cloud services orchestration for fulfilling intensive sensory datadriven IoT workflows. Future Gener. Comput. Syst. 108, 583–597 (2020) 20. L´ opez-Huguet, S., Natanael, I., Brito, A., Blanquer, I.: Vertical elasticity on marathon and Chronos Mesos frameworks. J. Parallel Distrib. Comput. 133, 179– 192 (2019) 21. Herrera, J., Molt´ o, G.: Toward bio-inspired auto-scaling algorithms: an elasticity approach for container orchestration platforms. IEEE Access 8, 52139–52150 (2020) 22. Truyen, E., Van Landuyt, D., Preuveneers, D., Lagaisse, B., Joosen, W.: A comprehensive feature comparison study of open-source container orchestration frameworks. Appl. Sci. 9(5), 931 (2019) 23. A. Inc.: AWS Greengrass (2018). https://aws.amazon.com/greengrass. Accessed 07 Feb 2018 24. M. Inc., Microsoft Azure - IoT Edge (2018). https://azure.microsoft.com/enus/ services/iot-edge. Accessed Feb 2018 25. Wen, Z., Yang, R., Garraghan, P., Lin, T., Xu, J., Rovatsos, M.: Fog orchestration for internet of things services. IEEE Internet Comput. 21(2), 16–24 (2017)

Design and Simulation of 2.4 GHz Microstrip Parallel Coupled Line Low Pass Filter for Wireless Communication System Shamsuddeen Yusuf1 , Shuaibu Musa Adam2,3(B) , Adamu Idris2 , David Afolabi4 , Vijayakumar Nanjappan5 , and Ka Lok Man6,7,8,9,10 1 Department of Electrical Engineering, Kano University of Science and Technology, Wudil,

Kano, Nigeria 2 Department of Physics, Federal University Dutsin-Ma, Dutsin-Ma, Nigeria

[email protected]

3 Faculty of Science and Computing, Al-Istiqama University Sumaila, Kano, Nigeria 4 Design Technology and Computer Science Department, United World College, Changshu,

China 5 Center for Ubiquitous Computing, University of Oulu, Oulu, Finland 6 School of Advanced Technology, Xi’an Jiaotong-Liverpool University, Suzhou 215123, China 7 Swinburne University of Technology Sarawak, Kuching, Malaysia 8 imec-DistriNet, KU Leuven, Leuven, Belgium 9 Kazimieras Simonaviˇcius University, Vilnius, Lithuania 10 Faculty of Informatics, Vytautas Magnus University, Vilnius, Lithuania

Abstract. A low pass filter only allows signals below its cut-off frequency to pass while attenuating other signals with frequencies higher than those of the filter. Several interesting techniques were proposed by researchers to design low pass filter. However, majority of those filters present difficulties of integration with other elements of electronics gadgets, high cost, high power consumption, large size and low-frequency application. Consequently, the current study focused on design and simulation of a parallel coupled-line microstrip low pass filter. Computer Simulation Technology (CST) microwave software was used for the design and simulation of the filter. Results Analyses were made and the resulting frequency responses were plotted using a sigma plot. It was concluded that the proposed microstrip filter presents solutions to the issues observed in the former designs. Keywords: Microstrip · Coupled line · Low pass filter · Insertion loss · Return loss

1 Introduction A microwave low pass filter is a two-port, passive, reciprocal and linear device that shunts unwanted signal frequencies while allowing the desired frequency to pass. In general, the electrical performance of a filter is described in terms of its frequency selectivity, insertion loss, return loss and group delay variation in the passband. Filters are © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 362–370, 2022. https://doi.org/10.1007/978-3-030-96299-9_35

Design and Simulation of 2.4 GHz Microstrip Parallel Coupled Line

363

required to have small insertion loss, large return loss for good impedance matching with interconnecting components and high-frequency selectivity to prevent interference [1]. If a filter has frequency selectivity, and guard band between channels can be determined to be small which indicates that the frequency can be used efficiently. Also, a small group delay and amplitude variation of the filter in the passband is required for minimum signal degradation [2]. Microwave low-pass filters (LPFs) have been widely used as a key building block to suppress all the unwanted high-frequency harmonics and inter-modulation in various wireless communication systems [3]. Due to unexpected frequency-distributed behaviour, the conventional transmission-line based LPFs, such as stepped-impedance and open-stub filters, always suffer from gradual cut-off attenuation skirt and narrow upper-stop bandwidth [4]. To sharpen this roll-off skirt, it is a usual approach to raise the order of a low-pass filter to a great extent. Unfortunately, this enlarges the overall circuit size and accumulates in-band insertion loss. In [5], a microstrip line with extremely high impedance is formed by etching out an aperture on the ground plane underneath the strip conductor. This approach allows effectively widening the upper stopband of stepped-impedance low-pass filters due to an enlarged impedance ratio. But, it still suffers from poor rejection near the cut-off frequency [6]. In recent years, various types of miniaturised and performance-improved low-pass have been explored in [7, 8] based on the original idea of [9]. These low-pass filters are basically constructed by a hairpin- or C-shaped folded unit in geometry. Because an additional coupling route is created between the two non-adjacent sections, attenuation poles are excited to improve the upper stopband performances. High impedance transmission lines and tightly-coupled twin lines are connected in parallel to generate three attenuation poles in the stopband [10]. A stepped impedance hairpin resonator is presented in [11] to make up a class of elliptic-function LPFs with a wide stopband. By allocating an attenuation pole close to the cut-off frequency, a sharp roll-off skirt was achieved in [12]. However, only one attenuation pole was excited in the stopband [13] all three attenuation poles are far away from the desired low passband, so the roll-off skirt near the cut-off frequency has not been increased so much [14]. In this letter, for the first time, a class of coupled-line hairpin filters is designed on a microstrip to have a low-pass frequency response with an expanded upper-stopband. It was designed by centrally tap-connecting the two coupled-line strip conductors and adjusting the coupling strength between them, three attenuation poles with controllable spacing can be excited in the first harmonic passband to widen and deepen the upper stopband. In analysis, a set of closed-form design equations are firstly established based on the well-known coupled-line theory [15] to explain the mechanism for exciting and reallocating these attenuation poles. Finally, two LPFs with a single and two asymmetrical units are optimally designed with the 3 decibel (dB) cut-off frequency at 2.5 Gigahertz (GHz). Measured results are provided to confirm the predicted ones for the single-unit LPF and experimentally demonstrate a wide upper-stopband in the range 3.2–11.8 GHz with the insertion loss higher than 20 dB for the two-unit LPF [16].

364

S. Yusuf et al.

2 Materials and Method 2.1 Setting of Simulation Parameters The design specification of the filter is shown in Table 1. The specification of the dielectric material was obtained from the Rogers Corporation [16]. Table 1. Filter specification Filter specification

Value

Cut-off frequency

2.4 GHZ

Load impedance

50 

Source impedance

50 

Insertion loss

−30 dB

Return loss

Greater than −0.5 dB

Order

3

Microstrip coupled line low pass filter is composed of three basic layers; the ground layer, the substrate and the patch which is the top layer. At each port of the low pass filter, there is a waveguide. As shown in Fig. 1, the ground and the upper layers are made up of conducting material. While the middle layer is the substrate, an insulating material, between the parallel-coupled lines forming the capacitor; and vacuum was used as the dielectric material. The filter was designed by following various steps using CST microwave studio.

Fig. 1. Proposed microstrip low pass filter

1. Ground layer: It is made up of a conducting material i.e. annealed copper which formed the base of the filter with a very thin thickness typically 0.035 mm. The

Design and Simulation of 2.4 GHz Microstrip Parallel Coupled Line

365

dimension of the ground plane is 90 mm by 30 mm. The ground layer is shown in Fig. 2. 2. Substrate layer: The proposed structure is printed over a low-cost FR-4 material which is readily available. The substrate with a dielectric constant of 4.3, loss tangent of 0.016 and thickness of 1.52 mm is considered material forming the substrate. The substrate layer is quite thicker than the ground plane. It has the same dimension as that of the ground plane. The substrate layer is shown in Fig. 3.

Fig. 2. Ground layer of the filter

Fig. 3. Substrate layer of the filter

The specification of dielectric material used as a substrate is given in Table 2. Table 2. Specification of dielectric material used as a substrate Parameter

Value

Dielectric constant (εr )

4.3

Substrate height

1.52 mm

loss tangent (δ)

0.016

Copper thickness

0.035 mm

366

S. Yusuf et al.

3. Transmission lines: The transmission lines form the upper layer of the low pass filter. It is also made up of annealed copper of thickness 0.035 mm. It consists of a rectangular bar laying horizontally and vertically on top of the substrate. The horizontal patch ensures the inductive or resistive effect, while the vertical layer which is space at one point ensures the capacitive effect and thus is responsible for rejecting other harmonics above the fundamental cut-off frequency. The horizontal transmission lines run throughout the length of the substrate (i.e. 90 mm). They are located 12 mm to 15 mm of the substrate width, hence they are 3 mm width. The vertical ones which are attached with the horizontal ones are generally of equal length and width i.e. 14 mm and 6.66 mm respectively. But, to ensure the capacitive effect of the low pass filter the vertical annealed copper was spaced between 4.8 mm and 5 mm to provide a space of 0.2 mm for the vacuum serving dielectric material. The vertical bars were grounded at 13.5 mm below the midpoint of the vertical patch with a cylindrical annealed copper material with inner and outer radius 0.15 mm and 3 mm respectively with a thickness that runs to ground layer of the filter. The transmission line layer is shown in Fig. 4.

Fig. 4. Transmission line layer of the filter

2.2 Analytical Relation for Calculating Characteristics Impedance and Effective Dielectric Constant If thickness of the copper track ‘t’ and the width ‘W ’ of the copper track is less than the height ‘h’ of the insulating material, the effective dielectric constant ‘εff ’ is found to using Eq. 1. Whereas, if the width of the track is greater than the height of the insulator, ‘εff ’ is found using Eq. 2. εff =

1 εr + 1 εr − 1 + [ 2 2 1+

12h W

  W 2 + 0.04 1 − h

Where εr is the relative dielectric constant.

⎡ 1 εr + 1 εr − 1 ⎣  εff = + 2 2 1+

(1)

⎤ ⎦ 12h W

(2)

Design and Simulation of 2.4 GHz Microstrip Parallel Coupled Line

367

Thus, the characteristics impedance becomes: 120π

 ZO = √ w εff h + 1.393 + 0.667 ln wh 1.444

(3)

2.3 Impedance Matching Impedance matching is necessary especially in communication to minimise loss of signal strength and ensure all the power generated is merely transmitted. The filter is no exception. CST microwave studio has the advantage of its built-in analytical calculator, which is used in determining various design parameters. 1. CST microwave studio calculator method: The calculator performs several functions but, most important for this design is calculating the wavelength, effective dielectric constant and the input impedance to ensure impedance matching and other important design parameters given in Table 3. Table 3. Proposed low pass filter design parameters Parameter

Value

Height of the substrate (h)

1.52 mm

Width of the transmission line (w)

3.00 mm

Dielectric constant of the substrate (εr )

4.30

Cut-off frequency

2.4 GHz

This calculator is easier to use in calculating the input and output impedance of the filter. The only parameters needed are the height of the substrate (h), the width of the transmission line (w) and lastly the dielectric constant of the substrate (ε_r). In some CST microwave calculator even the cut-off frequency may be of utmost importance. The values of Table 3 were inputted to the calculator in which the input and output impedance of the filter was found to be 49.80 . This is displayed in Fig. 5. 2. Analytical method for impedance calculation: This is given by the relationship shown in Eq. 6. First, we determined the effective dielectric constant from the information given in Table 2. εff =

1 εr + 1 εr − 1 + [ 2 2 1+

(4) 12h W

⎤ ⎡ 1 4.3 + 1 4.3 − 1 ⎣ ⎦ = 3.27  + εff = 2 2 12(1.52) 1+ 3

(5)

368

S. Yusuf et al.

Fig. 5. CST Microwave studio impedance calculation

The dielectric constant value obtained from Eq. (5) coincides with the value of effective permittivity calculated by the CST microwave calculator. ZO = √ ZO = √

3.27



εff

w h

120π

 + 1.393 + 0.667 ln wh 1.444 120π

3 1.52

+ 1.393 + 0.667 ln



3 1.52 1.444

 = 49.80

(6) (7)

It can be seen from Fig. 5 and Eq. (7) that, the results were found to be almost equal if not exactly equal to the one obtained from the CST microwave studio calculator. Thus, the CST microwave studio calculator and the analytical method proved to be effective.

3 Simulation Results and Analyses 3.1 Insertion (S21 ) and Return (S11 ) Loss The simulated low-pass filter response is shown in Fig. 6. The gain (dB) is plotted on the y-axis against the frequency (GHz) on the x-axis. It is clear that the simulated cut-off frequency was found to be 2.4 GHz. The value of the insertion loss (S21 ) and return loss (S11 ) at 2.4 GHz were found to be −29.941 dB and −0.505 dB respectively. 3.2 Field Monitors As expected, Fig. 7 shows the surface current at different frequency. The field monitors applied at 0.4 GHz, 0.8 GHz and 1.2 GHz show all the signal passing through the filter because, they are within the pass band frequency. The cut-off frequency of the filter as already known is 2.4 GHz and this correspond to what happened in Fig. 7(c) in which the signal is attenuated.

Design and Simulation of 2.4 GHz Microstrip Parallel Coupled Line

369

Reflection coefficient, dB

0

-10

-20

S 1 1 (R L)

-30

S 2 1 (IL)

-40 1.0

1.5

2.0

2.5

3.0

3.5

frequency, GHz

Fig. 6. Frequency response of the low-pass filter

(a)

(b)

(c) Fig. 7. Surface current at different field monitors

4 Conclusions Filters are one of the primary and essential parts of the microwave and communication systems. The microstrip low-pass filter was simulated using CST microwave studio software. In order to predict the performance of the filter, few parameters in the structure were analysed and found to have good relationship with the microwave theory. An optimisation process has been introduced along with the simulation procedure, focusing on the filter dimension in order to improve the response of the filter.

370

S. Yusuf et al.

5 Suggestions for Future Studies Validation of the proposed filter topology via prototype development is beyond the scope of this research study. It is suggested that this should be part of a future study’s objectives. Acknowledgement. This work is partially supported by the AI University Research Centre (AIURC) of the Xi’an Jiaotong-Liverpool University (XJTLU), China and Jiangsu (Provincial) Data Science and Cognitive Computational Engineering Research Centre at XJTLU.

Conflicts of Interest. The authors declare that they have no conflicts of interest.

References 1. Hong, J.S., Lancaster, M.J.: Microstrip Filters for RF/Microwave Applications. Wiley, New York (2001) 2. Zhu, L., Bu, H., Wu, K.: Unified CAD model of microstrip line with back side aperture for multilayer integrated circuit. IEEE MTT-S Int. Dig. 2, 981–984 (2000) 3. Wenzel, R.J.: Small elliptic-function low-pass filters and other applications of microwave C sections. IEEE Trans. Microw. Theory Tech. 18(12), 1150–1158 (1970) 4. Lee, Y.W., et al.: A design of the harmonic rejection coupled line low-pass filter with attenuation poles. In: Proceeding of the Asia-Pacific Microwave Conference, vol. 3, pp. 682–685 (1999) 5. Kuo, J.T., Shen, J.: A compact distributed low-pass filter with wide stopband. In: Proceedings of the Asia-Pacific Microwave Conference, vol. 1, pp. 330–333 (2001) 6. Zysman, G.I., Johnson, A.K.: Coupled transmission line networks in an inhomogeneous dielectric medium. IEEE Trans. Microw. Theory Tech. MTT17(20), 753–759 (1969) 7. Advanced Design System (ADS) 2006a. Agilent Technologies. Palo Alto, CA (2006) 8. Pozar, D.M.: Microwave and RF Design of Wireless Systems, 1st edn. John Wiley & Sons, New York (2001) 9. Coonrod, J.: PCB Fabrication and Material Considerations for the Different Bands of 5G, Rogers Corporation, Chandler, Arizona. Microwave Journal, pp. 14–15 (2018) 10. Richard, J., et al.: A Microwave Filters for Communication Systems. In: Satellite Communications Payload and System, pp. 118–145 (2018) 11. Ojaroudi, N., Ojaroudi, H., Ojaroudi, Y.: Very low profile ultrawideband microstrip band-stop filter. Microw. Opt. Technol. Lett. 56(3), 709–711 (2014) 12. Lan, Y., et al.: Flexible microwave filters on ultra thin liquid crystal polymer substrate. In: IEEE MTT-S Int. Microw. Symp., Phoenix, AZ, USA, pp. 1–3, May (2015) 13. Wael Abd, E.A., Ahmed, B.: Design of low-pass filter using meander inductor and U-form Hi-Lo topology with high compactness factor for L-band applications. Prog. Electromagn. Res. 55, 95–107 (2017) 14. Hsieh, L.H., Chang, K.: Compact elliptic-function low-pass filters using microstrip steppedimpedance hairpin resonators. IEEE Trans. Microwave. Theory Tech. 51(1), 193–219 (2003) 15. Sourabh S., Sonam, Y.C., et al.: Designing and parametric extraction of low pass filter using metamaterials. In: IEEE Students Conference (2020) 16. Teberio, F., et al.: Chirping techniques to maximize the power-handling capability of harmonic waveguide low-pass filters. IEEE Trans. Microw. Theory Techn. 64(9), 2814–2823 (2016)

A New Cascade-Hybrid Recommender System Approach for the Retail Market ˆ Miguel Angelo Rebelo1,2 , Duarte Coelho1,4 , Ivo Pereira1,3,4(B) , and F´ abio Fernandes1 1

E-goi, Av. Men´eres, 840, 4450-190 Matosinhos, Portugal i3s, Rua Alfredo Allen, 208, 4200-135 Porto, Portugal Universidade Fernando Pessoa, Pra¸ca 9 de Abril, 349, 4249-004 Porto, Portugal [email protected] 4 Interdisciplinary Studies Research Center, Rua Dr. Ant´ onio Bernardino de Almeida, 431, 4200-072 Porto, Portugal http://www.e-goi.com, https://www.i3s.up.pt, https://www.ufp.pt, http://www2.isep.ipp.pt/isrc 2

3

Abstract. By carefully recommending selected items to users, recommender systems ought to increase profit from product sales. To achieve this, recommendations need to be relevant, novel and diverse. Many approaches to this problem exist, each with its own advantages and shortcomings. This paper proposes a novel way to combine model, memory and content-based approaches in a cascade-hybrid system, where each approach refines the previous one, sequentially. It is also proposed a straight-forward way to easily incorporate time-awareness into rating matrices. This approach focuses on being intuitive, flexible, robust, auditable and avoid heavy performance costs, as opposed to black-box fashion approaches. Evaluation metrics such as Novelty Score are also formalized and computed, in conjunction with Catalog Coverage and mean recommendation price to better capture the recommender’s performance. Keywords: Cascade-hybrid marketing

1

· Recommender system · Intelligent

Introduction

Recommender systems (RS) have gained momentum as they help customers select products more suitable to their needs [2,8,17]. They use historical data to infer customer interests for intelligent marketing decisions, since past proclivities are often good indicators of future choices [2]. Personalized experiences can be extremely helpful in improving the customer’s overall satisfaction, since it guides each user through a large space of possible actions [8,16,17]. It is important to keep in mind that the primary goal of a RS is to increase product sales (their profit) [2]. By recommending carefully selected items to users, RS bring relevant items to the attention of users. To achieve the broader business goals, the c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 371–380, 2022. https://doi.org/10.1007/978-3-030-96299-9_36

372

ˆ Rebelo et al. M. A.

technical targets of RS are: Relevance, recommended items need to be relevant for the user at hand, or the communication will be ignored or found nettlesome [7]; Novelty, since it should recommend items that the user has not seen in the past (the opposite, in fact, can lead to reduction in sales) [7]; Serendipity, recommendations that truly surprise the user lead to higher sales diversity, having long-term benefits to the merchant [10]; Diversity, the list should contain items of different types to raise the probability of a correct match user-item [2]. The basic models for RS work with two kinds of data, which are user-item interactions and user and/or item variables. Methods that use user-item interactions are referred to as collaborative filtering (CF) methods [20]. Methods that use user and/or item variables, such as age and gender or product descriptions and keywords, are called content/contextual/knowledge-based RS [17]. Some combine these two approaches to create hybrid systems to perform more robustly in a wide variety of settings [5]. The proposed approach fits into a group of hybrid recommenders called cascade hybrids, in which each recommender actively refines the recommendations made by the previous recommender [5,8].

2

Literature Review

CF methods try to leverage the often high correlation of ratings across various users and items to impute the unobserved ones [2,20]. They branch into: memorybased methods, where the ratings are predicted on the basis of neighbors [1]; and model-based methods, which provide recommendations by developing a model from user ratings [2]. Recently, it has been shown that combinations of memorybased and model-based methods provide accurate results [14,21]. In content-based RS, the items’ information is used to make recommendations, by analyzing the items previously rated by a user to build a profile of user interests [16]. They have some advantages in making recommendations for new items with seldom to none interactions [2]. But due to the fact that the community knowledge is not leveraged, these methods provide obvious predictions, with reduced recommendations’ diversity [2,10,16]. Contextual information can also be leveraged to fine-tune recommendations in context-based post-filtering approaches. Such information could include time, location, or social data [2]. The present approach leverages context by the use of a time-decay process or the filtering of unwanted items. The three aforementioned systems, CF, content-based and context-based, exploit different sources of input that may work well in different scenarios. However many opportunities exist for hybridization, where various aspects from different types of systems are combined to achieve the best of all worlds. Ratings can be clustered into two main kinds, which impacts RS: explicit feedback, where the user has to explicitly rate their level of satisfaction in a predefined scale [2]; implicit feedback, as is the case of unary data, where the customer preferences are derived from their past activities [12]. Because only a small fraction of the items are rated frequently (popular items), the distribution of ratings among items often satisfies the long-tail property. This translates into

A New Cascade-Hybrid Recommender System Approach

373

a highly skewed distribution, with important implications: i) most popular items tend to leave little profit for the merchant [4]; ii) RS tend to suggest popular items due to the difficulty of providing robust rating predictions for the others [6]; iii) high-frequency items are not representative of the low-frequency [2].

3

Methods

In the present cascade approach (Fig. 1), the first recommender acts as a powerful filter and provides a rough ranking, eliminating many of the potential items [5]. The second level of recommendation uses this rough ranking to further refine it. The resulting ranking is then presented to the user. This procedure combines various methods in a particular way to serve the final recommendations.

Fig. 1. Diagram of the proposed new cascade-hybrid approach.

3.1

Data Partitioning

This paper focuses on a composite solution to the problem, since it would join the better aspects from different approaches. High and Low interaction customers were divided based on the number of items bought. For higher interaction costumers, a technique based on matrix factorization was chosen. Meanwhile for the lower interaction users, item-item K-nearest neighbors (K-NN) model was a better fit due to the lack of data from the user. The ALS strategy produced acceptable results for users with more than 3 interactions. 3.2

Content-Based Approach: Cosine Similarity

Content-based systems largely operate in the text domain, working with a wide variety of item descriptions and knowledge about users [2]. The present procedure is generically common to that of content-based, with the exception that here the objective is to filter the pre-computed recommendations: i) Descriptions were transformed into a keyword-based vector-space representation; ii) Pairwise cosine similarity was computed between items; iii) The calculated distances are used to filter previous recommendations on items for specific users.

374

ˆ Rebelo et al. M. A.

In the current approach, the cosine distance metric was the one used for item descriptions. Mathematically, it measures the cosine of the angle between two vectors projected in a multi-dimensional space [11]. The process used to get the most similar items is detailed in Algorithm 1. Algorithm 1. get most similar items 1: procedure algorithm 2: user history ← buying history for each user id 3: For each user id : 4: history items ← user history(uid) 5: similarity vectors ← similarity matrix[history items(indeces)] 6: centroid ← similarityvectors 7: recommended items ← user recommends(uid) 8: subcentroid ← centroid[recommended items(indeces)] 9: orderthesubcentroid 10: product indeces ← top-k indeces from subcentroid (most similar to the history) 11: product ids ← product ids[product indeces] 12: return product ids

3.3

Neighbourhood Model-Based Approach: ALS

Through matrix factorization [3,19], a matrix R with rank larger than k, can be approximately expressed as the product of rank-k factors: R ≈ U V T . Matrix factorization is particularly useful for very sparse data and can enhance the quality of recommendations [15]. This methodology was cleverly adapted for implicit feedback problems by Yifan Hu, et al., which proposed an efficient (weighted) ALS method for the factorization process in order to avoid the computational challenge of handling the large number of zero entries [12]. The best parameter space for this specific case was estimated through a cross-validation routine. 3.4

Neighbourhood Memory-Based Approach: Item-KNN

User-based approaches are often harder to scale because of the dynamic nature of users, whereas items usually do not change much [2]. In addiction, the userbased approaches do not perform well for low interaction cases. This is why the item-KNN approach was chosen for the low interaction users. In the item-based approach, to make recommendations for target item B, the user’s own ratings on neighboring (i.e., closely related) items are used [18]. The idea is to leverage the user’s own ratings on similar items to make the predictions [2]. 3.5

Time-Sensitive Recommendations: Time-Decay Approach

Any history data has the time at which an interaction occurred. One can take advantage of this information when computing an implicit rating. To implement a time-decay algorithm, the time-span since the item was bought was used:

A New Cascade-Hybrid Recommender System Approach

ratingui =

375

βbuy + αnui + δpui (elapsedtime + 1)gravity

where nui is the number of items bought and pui is the price of the product i. β, α, δ and gravity stand for the weights attributed to the act of buying, the number of products bought, the price of the product and the time decay element, respectively. The time-decay element (gravity), indicates how fast an item’s rating decays. Because people’s tastes change, old events should count less than new ones [2]. By changing these hyper-parameters, it is feasible to achieve different rating distributions (Fig. 2c). Final computed ratings were re-scaled to fit between 0 and 1 scale, for reduced computational strain.

4 4.1

Results and Discussion Time-Decay Ratings

Our data set from a fashion retailer revealed that most clients buy less than 0.01% of the product range. with 148301 users and 41285 items in total and sparcity = 99.98%. Most popular items are the less expensive ones, which is a common situation (Fig. 2a and 2b). It is not ideal to center the recommendations along these products, as they have lower profit margins [4].

Fig. 2. a) Distribution of buys by price. Lower priced products sell much more. b) ECDF of buys by price. Lower priced products amount to almost all of the buys. c) Computed ratings’ distributions.

As the empirical cumulative distribution (ECDF) of buys by the price of the product shows, lower priced products represent most of the interactions (Fig. 2a and 2b). It is important to notice that the price upper quartile per user is 14.99, with a mean of 15.27 per item bought. Also, extremely costly products are a rarity in this particular data set. The mean number of products bought by each user is 4.51 but there are cases in the hundreds (175 is the max). So, when designing the ratings, this variables have to be carefully weighted, since we ought to obtain highly discriminant implicit preferences without favoring too much the expensive items nor the popular ones. After testing different hyper-parameter

376

ˆ Rebelo et al. M. A.

combinations for the time-decay algorithm, three ratings were chosen (Fig. 2c), each with similar distribution characteristics. By including information about the unitary price, as well as the number of items bought, the preference of the item i by the user u can be more accurately inferred. If u bought 5 units of item i, there is more confidence to assume a higher preference than if he had only bought one. But to much weight in the number of items and the less expensive ones will probably be favored, which is undesirable. To circumvent this issue, the price of each item has to be weighted in. The more expensive the item i, the more probable it is that the user u did some research before the decision to buy that item. This means one can more confidently affirm that the user u has a stronger preference for item i. Furthermore, since the data comes from a fashion retailer, the decay of the ratings with time is of high importance to promote the newest (on the season) products. The hyperparameter spaces tested were: β = [0.9, 0.8, 0.9], α = [0.025, 0.02, 0.01], δ = [0.075, 0.18, 0.09], gravity = [0.2, 0.2, 0.3]. 4.2

Model-Based Approach

The hyper-parameter space for the matrix factorization model was optimized for each rating using RMSE as the metric of choice, in a cross-validation routine. The set of hyper-parameters tested was: Number of latent factors: [20, 40, 60, 80]; λ: [0.01, 0.001]; α: [10, 20, 40, 70]; Max iterations: [10, 25, 40, 50]. 4.3

Memory-Based Approach

As for the low interaction set of customers, the chosen approach was to compute the K-NN across items, due to the lack of user information. This way one can have accurate recommendations, at the sacrifice of serendipity. Testing three different similarity measures (cosine, adjusted-cosine and Pearson), the RMSE was the same for all (0.0077), but the adjusted-cosine metric revealed to be faster to train (2.9520 s) and faster to predict (1.0732 s), which functioned as a deciding factor. The number of k-nearest neighbors was set to 50. 4.4

Similarity Pruning

One key aspect of this cascade approach, or to say its novelty, is the similarity pruning technique used to filter the pre-computed recommendations obtained from other methods. This proved to be valuable, not on decreasing the error, but in giving contextually more relevant and accurate recommendations. The pre-computed latent-model recommendations were filtered using the proposed content-based approach, described in Algorithm 1. Pre-computing double the number of products needed in the final recommendation list, using the latentmodel and the Item-KNN approaches, gave enough diversity and serendipity to the recommendations while Algorithm 1 assured that sub-optimal suggestions were not being presented (i.e., mixing girl and boy recommendations for a client with a history of only buying boy’s clothes).

A New Cascade-Hybrid Recommender System Approach

4.5

377

Evaluation and Performance Considerations

By measuring the conversion rate of users, the direct impact of the RS can be estimated. However, it is often not feasible to use them in bench-marking and research. For that reason, offline evaluations with historical data sets are used [2,9]. Accuracy measures alone can often provide an incomplete picture of the RS’s performance. Metrics such as novelty, coverage, and serendipity, though hard to quantify, are important for the user experience and have important short and long-term impacts on the conversion rates [2,9,10]. The proposed model was evaluated using root mean squared error RMSE (hyper-parameter optimization only) [13], Catalog Coverage (CC), Novelty Score and Scalability measures. The mean recommendation price was also computed, since pricier items have higher profit margins. CC evaluates how the item catalog is being covered by the recommendations, since they should be diverse across users [9]. Let Tu represent the list of top-k items recommended to user u. The CC is defined as the fraction of items that are recommended to at least one user [9]. m Tu | |Uu=1 n The Novelty Score (NS) can be estimated using the rating’s time stamps. All ratings after time t0 were removed from the training data, and then used for scoring purposes. For each item rated after t0 and correctly recommended, the novelty evaluation score is rewarded. In this approach it is assumed that popular items are less likely to be novel, and less credit was given to them.

CC =

|U m Cu | − N S = u=1 nr

m |Uu=1 Pu | 2

Here, Cu represents the number of correct recommendations in the list of items for an user u, Pu is the number of popular items in Cu and nr is the total number of recommendations, as inspired by the works of Mouzhi Ge, et al. [9]. Table 1 shows the evaluation metrics obtained for the proposed model with three different ratings (computed using different time-decay hyper-parameter spaces, detailed above), plus a model using unary ratings (baseline). Due to the nature of this particular data set, the proposed model’s NS is lower compared to the baseline. There is a trade-off between CC and NS, so some compromise is necessary. Meanwhile, CC averaged close to 20% across the proposed models. These matrices largely surpass the baseline, which is a good indicator of the ability to recommend across a diverse set of products from the catalog. Figure 3 demonstrates how CC evolves with the proportion o users that get recommendations. Recommendations using the baseline recommender with the unary ratings hit the plateau much quicker, having a distinct disadvantage when compared to the proposed model using the computed ratings. This and the improved mean recommendation price have important business implications [2,4,7,10].

378

ˆ Rebelo et al. M. A. Table 1. Model evaluation metrics. Model baseline

Catalog coverage Novelty score Median price 7.51%

6.24%

11.51

Model 1 19.80%

1.58%

11.98

Model 2 19.51%

1.53%

13.36

Model 3 19.43%

1.66%

12.79

Fig. 3. CC by proportion of users with recommendations.

RS should also be designed to perform efficiently [22]. For this purpose, some measures of scalability were also collected: Training time was approximately 8 s/200Mentries ; Prediction time averaged 30 s/1000users ; Memory requirements averaged 15 Gb for the same matrix. Both the model-based and memory-based approaches have good performance because they can be easily parallelized. There are very efficient applications using both [12,18]. Still and all, the present approach is using both and applying on top of it a content-based method. By the data was partitioned between high and low interaction users, the size of the matrices for each approach (model and memory-based) was significantly reduced. So, the same original matrix is not being used twice. One has also to account for the transformation of the product descriptions into vector-space representations, the computation of pairwise similarity between items and then the computational strain from the pruning done by Algorithm 1. This slows down the total computation time, as expected. But due to the simple matrix calculations it performs, it is not slower in total than any of the other methods in isolation. One of its advantages is that it has no hyper-parameters to optimize and works with just the product descriptions. Both the model-based and memory-based methods have hyper-parameters that need to be tailor-optimized for the issue at hand.

A New Cascade-Hybrid Recommender System Approach

5

379

Conclusions

The present work describes a novel cascade-hybrid approach for RS that combines time awareness and strengths from the model, memory and content-based approaches in an intuitive and flexible manner, without heavy performance costs. Offline evaluation metrics for RS were also developed on, formalizing a way to compute NS. The model exhibited good evaluation metrics, revealing a good trade-off between CC and NS. Checking the outputs from the same model using different time-decay hyper-parameters, proved that this step has a great influence on final results and needs to be tailor-optimized for the retailer at hand. Though good, there is an ongoing effort to raise NS metrics. This cascade-hybrid approach has one key advantage beyond being highly adaptable: it is highly interpretable and does not function in a black-box fashion. It can adapt to diverse business practices and objectives, and its results are easily auditable. In the content-based similarity pruning using item’s textual descriptions, we are working on giving textual features differential weight based on the importance of product characteristics. There is also a performance cost to this hybrid solution. Acknowledgements. This article is a result of the project “Cria¸ca ˜o de um N´ ucleo de I&D para a gera¸ca ˜o de novo conhecimento nas ´ areas de Inteligˆencia Artificial, Machine Learning, Intelligent Marketing e One-2-One Marketing”, supported by Operational Programme for Competitiveness and Internationalisation (COMPETE 2020), under the PORTUGAL 2020 Partnership Agreement, through the European Regional Development Fund (ERDF), for E-goi.

References 1. Aditya, P.H., Budi, I., Munajat, Q.: A comparative analysis of memory-based and model-based collaborative filtering on the implementation of recommender system for E-commerce in Indonesia: a case study. In: 2016 International Conference on Advanced Computer Science and Information Systems, pp. 303–308. IEEE, Malang, Indonesia (2016) 2. Aggarwal, C.C.: Recommender Systems. Springer International Publishing, Cham (2016) 3. Aggarwal, C.C., Parthasarathy, S.: Mining massively incomplete data sets by conceptual reconstruction. In: Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 227–232. KDD 2001, Association for Computing Machinery, New York, USA (2001) 4. Anderson, C.: The Long Tail: Why the Future of Business Is Selling Less of More. Hyperion, New York (2006) 5. Burke, R.: Hybrid recommender systems: survey and experiments. User Model. User Adapt. Interact. 12(4), 331–370 (2002) 6. Cremonesi, P., Koren, Y., Turrin, R.: Performance of recommender algorithms on top-N recommendation tasks. In: Proceedings of the Fourth ACM Conference on Recommender Systems, pp. 39–46. RecSys 2010, Association for Computing Machinery, Barcelona, Spain (2010)

380

ˆ Rebelo et al. M. A.

7. Fleder, D.M., Hosanagar, K.: Recommender systems and their impact on sales diversity. In: Proceedings of the 8th ACM Conference on Electronic Commerce EC 2007, p. 192. ACM Press, San Diego, California, USA (2007) 8. Funk, P. (ed.): Advances in Case-Based Reasoning: 7th European Conference, ECCBR 2004, Madrid, Spain, August 30–September 2, 2004, Proceedings, vol. 3155. Lecture Notes in Computer Science, Lecture Notes in Artificial Intelligence, Springer, Heidelberg (2004). https://doi.org/10.1007/b99702 9. Ge, M., Delgado-Battenfeld, C., Jannach, D.: Beyond accuracy: evaluating recommender systems by coverage and serendipity. In: The 4th ACM Conference, p. 257. ACM Press, Barcelona, Spain (2010) 10. Good, N., et al.: Combining collaborative filtering with personal agents for better recommendations. In: Combining Collaborative Filtering with Personal Agents for Better Recommendations, pp. 439–446. AAAI 1999/IAAI 1999, American Association for Artificial Intelligence, Orlando, Florida, USA (1999) 11. Han, J., Kamber, M., Pei, J.: Getting to know your data. In: Data Mining, pp. 39–82. Elsevier (2012) 12. Hu, Y., Koren, Y., Volinsky, C.: Collaborative filtering for implicit feedback datasets. In: 2008 Eighth IEEE International Conference on Data Mining, pp. 263–272. IEEE, Pisa, Italy (2008) 13. Hyndman, R.J., Koehler, A.B.: Another look at measures of forecast accuracy. Int. J. Forecast. 22(4), 679–688 (2006) 14. Koren, Y.: Factor in the neighbors: scalable and accurate collaborative filtering. ACM Trans. Knowl. Disc. Data 4(1), 1–24 (2010) 15. Koren, Y., Bell, R., Volinsky, C.: Matrix factorization techniques for recommender systems. Computer 42(8), 30–37 (2009) 16. Lops, P., de Gemmis, M., Semeraro, G.: Content-based recommender systems: state of the art and trends. In: Ricci, F., Rokach, L., Shapira, B., Kantor, P.B. (eds.) Recommender Systems Handbook, pp. 73–105. Springer, Boston (2011). https:// doi.org/10.1007/978-0-387-85820-3 3 17. Lorenzi, F., Ricci, F.: Case-based recommender systems: a unifying view. In: Mobasher, B., Anand, S.S. (eds.) Intelligent Techniques for Web Personalization. Lecture Notes in Computer Science. LNCS, vol. 3169, pp. 89–113. Springer, Heidelberg (2005). https://doi.org/10.1007/11577935 5 18. Sarwar, B., Karypis, G., Konstan, J., Reidl, J.: Item-based collaborative filtering recommendation algorithms. In: Proceedings of the Tenth International Conference on World Wide Web, pp. 285–295. ACM Press, Hong Kong (2001) 19. Sarwar, B., Karypis, G., Konstan, J., Riedl, J.: Application of Dimensionality Reduction in Recommender System - A Case Study. Technical report, Defense Technical Information Center, Fort Belvoir, VA (2000) 20. Srifi, M., Oussous, A., Ait Lahcen, A., Mouline, S.: Recommender systems based on collaborative filtering using review texts—a survey. Information 11(6), 317 (2020) 21. Su, X., Khoshgoftaar, T.M.: A survey of collaborative filtering techniques. Adv. Artif. Intell. 2009, 1–19 (2009) 22. Tak´ acs, G., Pil´ aszy, I., N´emeth, B., Tikk, D.: Scalable collaborative filtering approaches for large recommender systems. J. Mach. Learn. Res. 10, 623–656 (2009)

A Novel Deep Neural Network Based Approach for Alzheimer’s Disease Classification Using Brain Magnetic Resonance Imaging (MRI) Ruhul Amin Hazarika(B) , Debdatta Kandar, and Arnab Kumar Maji(B) Department of Information Technology, NEHU, Shillong 793022, India

Abstract. Alzheimer’s Disease (AD) is a fatal cognitive disorder, where patients experience serious symptoms of dementia. In AD, brain is the primarily affected region. Bio-markers of AD can be obtained from brain images which can help in classification of the disease. Traditional image processing technologies struggle to process the images due to the intricate nature of brain tissues. Deep Neural Network (DNN) is a well-known Machine Learning (ML) approach that helps to take decisions from complex situations. DNN is widely used in the applications of image processing such as classification. In this work, some of the most famous DNN models are implemented for AD classification using brain Magnetic Resonance (MR) images. All of the data utilised in this study are obtained from the “Alzheimer’s Disease Neuroimaging Initiative: ADNIonline data set”. According to the findings of the experiment, all the models have their average performance rate below 90%. It is also observed that, because of simple and effective architectures, LeNet and AlexNet consume the least computational time (64 s and 74 s per epoch respectively) amongst all the implemented models. Hence, to design a model that can perform better and faster, we have combined LeNet and AlexNet parallelly and proposed a hybrid approach for AD classification. It is observed that the hybrid method performs convincingly with an average performance rate of 90.68%. The average computational time taken by the hybrid model is 120 s per epoch which is faster than all other discussed models (except LeNet and AlexNet). Keywords: Magnetic Resonance Imaging (MRI) · Deep Neural Network (DNN) · Machine Learning (ML) · Alzheimer’s Disease (AD) Mild Cognitive Impairment (MCI) · Cognitively Normal (CN)

1

·

Introduction

Alzheimer’s disease (AD) is a top reason of death in dementia [1]. The most typical signs of AD are difficult in remembering, cognitive decline, losing controls in mood/personality, and so on [2]. Because limbic in the nervous system regulates the majority of the symptoms of AD, some of its constituents, for example, c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 381–390, 2022. https://doi.org/10.1007/978-3-030-96299-9_37

382

R. A. Hazarika et al.

hippocampus and amygdala, take the brunt of the damage. [3]. Most AD victims undergo a mild dementia phase called Mild Cognitive Impairment (MCI), before suffering major signs of the disease [4]. In MCI, the person begins to show signs of AD. Hence, MCI is one of the most important variants in the classification of AD. While tracking brain alterations, it has been discovered that the changes may begin before a person fully gets AD [5]. The hippocampus, entorhinal cortex, and other memory-forming regions are among the first areas to be damaged [3]. More cells die on a moderate scale, resulting in brain atrophy. It is possible to classify AD more effectively by utilising brain imaging to extract features from every classes [6]. Figure 1 shows a variety of brain scans for distinct stages of dementia, including Cognitively Normal (CN), MCI, and AD.

Fig. 1. Sample brain MR image of a. CN, b. MCI, and c. AD subject

An Artificial Neural Network (ANN) is a subclass of ML techniques that are influenced by the human brain’s functional structure [7]. ANN creates a structure of interconnected neurons that aids the system in absorbing fresh knowledge from the environment [8]. An ANN with several linked neurons concealed between the input and output layers is known as a Deep Neural Network (DNN). Hidden layers assist the model in accurately training and producing desired outcomes [9]. DNN works well even if the information in the data is difficult to grasp, such as MR scans [10]. Figure 2 depicts a sample DNN architecture. If the incoming

Fig. 2. Sample DNN architecture used in classifying images

A Novel DNN Based Approach for AD Classification Using Brain MRI

383

weights of a cell ‘n’ are v1 , v2 , v3 , ..., vi , then the numerical expression in ‘n’ can be expressed as Eq. 1.  N= vi × xi + a (1) i

‘a’ stands for a bias value. The neuron ‘n’ must make an output decision after estimating ‘N’. The transfer function γ utilised in the network determines the output ‘Z’ of ‘n’, which is given by Eq. 2. ReLu, Softmax, etc. are the examples of well-known activation functions. Z = γ(N )

(2)

The loss function is determined after all the data has been redistributed, possibly totally or in batches. The sum of errors established at the forecasted outcomes is the loss function [11]. Forward Propagation (FP) is the name given to this phenomenon. In Eqs. 3 and 4, two prominent loss functions, MeanSquareError (MSE) and BinaryCrossEntropy (BCE), are explained. n

1 [ri − f (pi ; w)]2 n i=1

(3)

1 ri log[f (pi ; w)] + (1 − ri ) log[(1 − f (pi ; w)] n i=1

(4)

M SE : L(v) = n

BCE : L(v) =

The real parameters are ri , and the projected parameters are f (pi ; w).. Upon obtaining the deficit estimate, the following operation is used for computing gradients of every essential parameters and improve them using a proper method. Back propagation (BP) is the term for this phenomenon. Equations 5 and 6 can be used to represent the Back Propagation process. Gr =

∂L(v) ∂v

vnew = vcurrent − α

∂L(vcurrent ) ∂vcurrent

(5) (6)

The learning rate is denoted by alpha. When BP is carried out at neuron ‘l’ towards ‘n’ through ‘m’ the procedure is represented by Eq. 7. ∂L(v) ˆ ∂n ∂L(v) ∂ m · · = ∂v1 ∂m ˆ ∂w ∂v1

(7)

Researchers have built a number of DNN models for image classification. As our knowledge, very less DNN models are applied in the classification of AD. In this work, we have implemented some popularly used DNN models. By considering execution time as an important factor, we have considered the models having simple and effective architectures. LeNet, AlexNet, VGG 16, VGG 19, Inception V1, Inception V2, and Inception v3 are the models we’ve examined for implementation. Though there are many advanced DNN models available

384

R. A. Hazarika et al.

for image classification, the main motivation of taking the specific models for implementation are, i) LeNet is an old and simple model with effective performance, ii) AlexNet is the winner of 1st-ImageNet Large Scale Visual Recognition Challenge (ILSVRC), iii) our focus is to design a simple and effective approach for AD classification that requires less memory space as well as computational resources. From the performance comparison, it is observed that all these existing models performs with an average rate of below 90%. Since LeNet and AlexNet can perform faster amongst all these models, we have combined these two models parallelly to get a better classification model. It is observed from the implementation results that, the proposed hybrid approach can classify AD more convincingly.

2

Related Study: ANN in Classification of AD

The ANN algorithms are trending in the classification of AD. One of the main reasons for its appeal is that it can learn from its surroundings and increase its forecasting efficiency in subsequent repetitions [12]. The following sections explore few recently proposed works on ANN in the AD classification. A residual DNN-based AD classification approach was presented by Farheen Ramzan, et al. [13]. For accurate training, authors inherited the concept from ResNet model. Proposed model include OneChannelResNet (1CRN), OffTheShelf ResNet (OtS), and FineTuningResNet (FTRN). The batch size used is 32, and SGD based solver is employed in the model. The OtS model outscored the other two ResNet models convincingly, according to the authors. By concatenating Convolutional as well as Recurrent Neural Networks (RNN), Manhua Liu et al. developed an AD classification method [14]. All the 3D images are transformed to a series of 2D slices. A combined method of CNNs and RNNs is applied for training the model about intra- as well as inter-slice information. RNNs’ Gated Recurrent Unit (GRU) is trained to learn and absorb inter-slice information, whereas 2D CNNs are designed to collect the features from each visual slice. Using DenseNet as a reference, Braulio Solano Rojas et al. developed a DNN-based technique for AD classification [15]. From 3D MR data, the authors selected 42 of the most appropriate slices for further processing. The authors employed DenseNet’s Bottleneck-Compressed based model. In addition to the original architecture, the authors introduced a channel option that took into consideration three distinct channels (RGB) from monochromatic MR pictures. To improve imaging feature selection, the M3d-Cam tool is paired with a GuidedGradientweightedClassActivationMapping (Grad-CAM) method. Attention mapping is a technique that aids in the detection of undesirable elements. All undesirable pixels are then eliminated using the proper processing methods. Boo Kyeong Choi, et al. demonstrated a CNN-based technique for categorising AD, MCI, and CN patients using brain scans [16]. At first, the 3D Slicer toolbox is used to separate hippocampus areas in brain images. After that,

A Novel DNN Based Approach for AD Classification Using Brain MRI

385

a homogeneity rectification strategy based on Local-Entropy-Minimization-bi cubic Spline is used to handle the area of segmented areas (LEMS). Finally, a binary neural network-based classifier is used to perform the classifications. The suggested network comprises of an input layer-two ConvolutionalLayers, two MaxPoolingLayers, FlattenLayers, FullyConnectedLayers, and an Output classification layer. For AD classification, Xin Bi et al. [17] introduced a CNN model based on the concept of extreme learning. Two different networks were designed for categorising functional brain networks. In addition, the concept of a boosted Extreme Learning Machine (ELM) is presented. The model is trained on aspects of deep regional connection using ELM. ELM is also used for helping the network to train about the features in its immediate vicinity. The brain network is built using the Pearson correlation (PC) coefficient. ConvolutionalLayers, the ReLuActivationFunction, PoolingLayers, FullyLinkedLayers, and DecisionLayers make up the proposed DNN.

3

Discussion and Experimental Evaluation of Different DNN Models for AD Classification

Data and Tools: We obtained T1-weighted, MagnetizationPreparedRapidGradientEcho (MPRAGE) brain MRIs. More than 100 patients are considered while acquiring the data from online data-set ADNI [18] for training and testing of the models. The images were obtained from 3 different patient categories, are i) CN, ii) MCI, and iii) AD. For accurate training of the models, the DataGenerator (DG) function is used to increase the training images. The parameters used in the DG function are, rotation, mirror reflection, and so on. Table 1 shows how all of the data is distributed. Table 1. Data distribution Subject No. of subjects Training images Testing images Total images CN

50

2000

400

2400

MCI

50

2000

400

2400

AD Total

50

2000

400

2400

150

6000

1200

7200

Experimental Setup: We used a CPU with twelve GB of RAM, 500 GB of SSD storage, two GB graphics (Nvidia), and an i7 processorto evaluate all of the experimental analyses in this work. We used the Python 3.0 toolbox for the experimental implementations. Parameter Settings: For all the implemented models, we have used “softmax” activation function, “StochasticGradientDescent (SGD)” optimizer, and “SparseCategoricalCrossEntropy (SCCE)” loss function. Data are distributed in 32 batches and are trained using 60 epochs.

386

R. A. Hazarika et al.

A short description about all the implemented models are described below. LeNet: Yann LeCun presented LeNet, a deep neural networks architecture, in 1989 [19]. The typical architecture of LeNet consists of several layers such as Input → ConvolutionLayer (CL) → P oolingLayer (P L) → CL → P L → DenseLayer (DL) → DL → Output. Here, Convolution is an operation in which feature maps are being used to overlap tumbling over the image pixels. Equation 8 can be used to represent the convolutional operation.  Ix−a,y−b ∗ Ja,b (8) Convab = C(x, y) = bias + (I ∗ J)xy = bias + a

b

In Eq. 8, ‘C’ represents the convolutional function, ‘I’ input function (pixel value matrix), ‘J’ kernel function, and ‘a’ & ‘b’ represents rows and the columns of the pixel matrices. By discarding less pertinent information, the pooling operation aims to minimise the dimensionality of the matrices. Popularly used pooling operations are, max-pooling (eliminates the minimum weighted features) and the average pooling (takes the average weight of the feature elements in a window). LeNet is well-known for simple yet an effective architecture for classification problems. AlexNet: AlexNet is another DNN based classification model. The architecture of the model is designed by Alex Krizhevesky in 2012 [20]. Layers in this model are Input → CL → P L → CLtoP L → CL → CL → CL → P L → DL → DL → Output. The 1stt algorithm to win Imagenet’sLargeScaleVisualRecognitionChallenge (ILSVRC)-2012 was AlexNet. [21]. VGG-16 & VGG-19: VGG16 is refers to a model that have 16 deep layers. Karen Simonyan and Andrew Zisserman from Oxford University’s Visual Geometry Group (VGG) Lab designed the model’s architecture in 2014 [22]. A VGG 16 model’s typical architecture consists of 13 CLs, 5 PLs, and DLs. VGG 16 finished 2nd in the ILSVRC-2014 [23]. VGG-19 refers to a VGG model variant that also adheres to the VGG-16 architecture having a total of 47 layers, amongst which 19 are deep layers [24]. Inception-V1, V2, & V3: One major issue with the deeper models is that the models are computationally expensive due to the large number of serial convolutional operations [25]. To solve this problem, M. Lin et al. [26] presented a concept called inception module in the year of 2014. Module allows multiple convolutional operations to be performed in parallel. C. Szegedy et al. designed architecture of the Inception-V1 (GoogleNet) model using the concept of Inception Module (IM) as a reference [27]. The model comprises of 22 layers having many IMs (each module consists of 1 × 1, 3 × 3 & 5 × 5 CLs and a 3 × 3 PL). Inception-V1 is also the winner of ILSVRC-2014 [28]. Inception-V1 uses the density filters, such as 5 × 5, that can cause the input dimensions to decompose by a large margin, potentially resulting in the loss of crucial information [29]. Later, the Inception-V2 model is presented, in which the 5 × 5 convolutions are replaced with two 3 × 3 convolutions [30]. Another change

A Novel DNN Based Approach for AD Classification Using Brain MRI

387

made to this model is that the ntimesn factorization is changed to ntimes1 & 1timesn, that makes the model computationally faster. The Inception-V3 is nothing but an updated variant of the Inception concepts. Label smoothing, factorising of 7 × 7 convolutions, using RMSprop optimizer, and other major changes introduced in the 3rd version [30]. Inception-V3 finished first runner-up in the ILSVRC-2015 [31]. For performance evaluation, we have considered some of the popular performance analysis parameters such as Accuracy, Precision, Recall, and F1-score. The average of all these performance measures are calculated. Table 2 shows the efficiency of the DNN models that have been discussed. Table 2. Performance comparison table Models

Performance (Average) Average time required per epoch

LeNet

0.7850

64 s

AlexNet

0.6925

74 s

VGG-16

0.7825

120 s

VGG-19

0.8475

230 s

Inception-V1 0.8125

210 s

Inception-V2 0.8200

175 s

Inception-V3 0.8225

187 s

From the performance comparison it can be observed that, none of the discussed models could perform convincingly. It is also observed that LeNet and AlexNet took least implementation time than all other discussed models. To improve the classification performance in least possible implementation time, we have proposed to combine LeNet and AlexNet together. Details of the proposed hybrid approach is discussed in Sect. 3.1. 3.1

Proposed Hybrid DNN Based Model for AD Classification

For the proposed model, we considered all of LeNet’s and AlexNet’s layers and combined them parallelly. The proposed model’s architecture is depicted in Fig. 3. In Fig. 3, “+” denotes the concatenation of the layers. Performance of the hybrid model is presented in Table 3. It can be observed from Table 3 that

Fig. 3. Block diagram of the proposed architecture

388

R. A. Hazarika et al. Table 3. Evaluation of the proposed model’s performance

Model

Category

Proposed hybrid DNN model

CN vs MCI 0.9075 MCI vs AD 0.8950 CN vs AD 0.9325

Average

Accuracy Recall

0.9117

0.8950 0.9000 0.9275

Precision F1 score Average performance

Average time per epoch

0.8850 0.8900 0.9350

0.9075 0.9034

0.9100 0.8950 0.9100

0.8994 0.8950 0.9261

116 s 119 s 112 s

0.9050

0.9068

116 s

Fig. 4. Performance comparison graph

the proposed model achieves a convincing performance of around 91% accuracy, 0.91 recall, 0.90 precision, and 0.91 F1 Score. The model’s average performance is approximately 91%. It can also be seen that, while the model’s average computational time is longer than that of LeNet and AlexNet, the model still outperforms all other models discussed. Figure 4 depicts the average performance comparison graph for all of the models discussed in this work. In Fig. 4, x-axis shows the models name and y-axis is the average performance achieved by the models. The performance comparison graph shows that the proposed hybrid model with an average performance rate of 90.68%, outperforms all other discussed models convincingly. The next highest performance is obtained by the model Inception V3 with an average performance rate of 82.25%.

4

Conclusion and Future Work

In this work, few of the popularly used DNN models are examined and implemented for AD classification using brain MR images. Classification performance as well as the average computational time required are compared amongst the models. According to the performance evaluation, VGG-19 achieved the highest classification performance among the discussed models. One disadvantage of the VGG-19 model is that it requires more memory space and execution time. The LeNet and AlexNet models consume the least amount of execution time. Considering computational time as an important factor, we have proposed a hybrid model for improving the classification performance. LeNet and AlexNet are combined in parallel in this hybrid model. According to the performance evaluation of the proposed model, the hybrid approach outperforms all of the discussed models also in a convincing executing time.

A Novel DNN Based Approach for AD Classification Using Brain MRI

389

Though the hybrid approach achieves the highest performance in AD classification amongst the discussed models, the model still can be improved by introducing some advanced DNN parameters, such as depth-wise convolution, dense block, etc. In our future work, more recently developed DNN models can be added for the analysis purposes, and more data can be acquired for better train of the models from different sources. Moreover, since human brain changes over aging, data distribution can be done age-wise to obtain a better AD classification model.

References 1. Association, A., et al.: 2018 Alzheimer’s disease facts and figures. Alzheimer’s Dement. 14(3), 367–429 (2018) 2. Korolev, I.O.: Alzheimer’s disease: a clinical and basic science review. Med. Stud. Res. J. 4(1), 24–33 (2014) 3. Moon, S.W., Lee, B., Choi, Y.C.: Changes in the hippocampal volume and shape in early-onset mild cognitive impairment. Psych. Investig. 15(5), 531–537 (2018). https://doi.org/10.30773/pi.2018.02.12 4. Gauthier, S., et al.: Mild cognitive impairment. The Lancet 367(9518), 1262–1270 (2006). https://doi.org/10.1016/S0140-6736(06)68542-5 5. Beason-Held, L.L., Goh, J.O., An, Y., Kraut, M.A., O’Brien, R.J., Ferrucci, L., Resnick, S.M.: Changes in brain function occur years before the onset of cognitive impairment. J. Neurosci. 33(46), 18008–18014 (2013). https://doi.org/10.1523/ JNEUROSCI.1402-13.2013 6. Oh, K., Chung, Y.-C., Kim, K.W., Kim, W.-S., Oh, I.-S.: Classification and visualization of Alzheimer’s disease using volumetric convolutional neural network and transfer learning. Sci. Rep. 9(1), 1–16 (2019). https://doi.org/10.1038/s41598-01954548-6 7. Wang, S.C.: Artificial neural network. In: Interdisciplinary Computing in Java Programming. The Springer International Series in Engineering and Computer Science, vol. 743. Springer, Boston, MA (2003). https://doi.org/10.1007/978-14615-0377-4 5 8. Pagel, J.F., Kirshtein, P.: Machine Dreaming and Consciousness. Academic Press, Cambridge (2017) 9. Raghavan, V.V., Gudivada, V.N., Govindaraju, V., Rao, C.R.: Cognitive Computing: Theory and Applications. Elsevier, Amsterdam (2016) 10. Lundervold, A.S., Lundervold, A.: An overview of deep learning in medical imaging focusing on MRI. Zeitschrift f¨ ur Medizinische Physik 29(2), 102–127 (2019). https://doi.org/10.1016/j.zemedi.2018.11.002 11. Mebsout, I.: Deep Learning’s mathematics. https://towardsdatascience.com/deeplearnings-mathematics-f52b3c4d2576. Accessed 23 May 2021 12. Dumitru, C., Maria, V.: Advantages and disadvantages of using neural networks for predictions. Ovidius Univ. Ann. Ser. Econ. Sci. 13(1), 444–449 (2013) 13. Ramzan, F., et al.: A deep learning approach for automated diagnosis and multiclass classification of Alzheimer’s disease stages using resting-state FMRI and residual neural networks. J. Med. Syst. 44(2), 1–16 (2020). https://doi.org/10.1007/ s10916-019-1475-2

390

R. A. Hazarika et al.

14. Liu, M., Cheng, D., Yan, W.: Alzheimer’s disease neuroimaging I, classification of Alzheimer’s disease by combination of convolutional and recurrent neural networks using FDGPET images. Front. Neuroinform. 12, 35 (2018). https://doi.org/10. 3389/fninf.2018.00035 15. Solano-Rojas, B., Villal´ on-Fonseca, R.: A low-cost three-dimensional DenseNet neural network for Alzheimer’s disease early discovery. Sensors 21(4), 1302 (2021). https://doi.org/10.3390/s21041302 16. Choi, B.-K., et al.: Convolutional neural network-based MR image analysis for Alzheimer’s disease classification. Current Med. Imaging 16(1), 27–35 (2020). https://doi.org/10.2174/1573405615666191021123854 17. Bi, X., Zhao, X., Huang, H., Chen, D., Ma, Y.: Functional brain network classification for Alzheimer’s disease detection with deep features and extreme learning machine. Cogn. Comput. 12(3), 513–527 (2020). https://doi.org/10.1007/s12559019-09688-2 18. ADNI, Alzheimer’s disease Neuroimaging Initiative: ADNI. http://adni.loni.usc. edu/data-samples/access-data. Accessed 13 Sep 2020 19. Albawi, S., Mohammed, T.A., Al-Zawi, S.: Understanding of a convolutional neural network. In: 2017 International Conference on Engineering and Technology (ICET), pp. 1–6. IEEE (2017) 20. Alom, M.Z., et al.: The history began from AlexNet: a comprehensive survey on deep learning approaches (2018). arXiv preprint arXiv:1803.01164 21. Nagata, F., et al.: Orientation detection using a CNN designed by transfer learning of AlexNet. In: Proceedings of the 8th IIAE International Conference on Industrial Application Engineering 2020, Vol. 5, pp. 26–30 (2020). https://doi.org/10.12792/ iciae2020.051 22. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition (2014). arXiv preprint arXiv:1409.1556 23. Mehra, R., et al.: Breast cancer histology images classification: training from scratch or transfer learning? ICT Express 4(4), 247–254 (2018). https://doi.org/ 10.1016/j.icte.2018.10.007 24. Kwasigroch, A., Mikolajczyk, A., Grochowski, M.: Deep neural networks approach to skin lesions classification-a comparative analysis. In: 2017 22nd International Conference on Methods and Models in Automation and Robotics (MMAR), pp. 1069–1074. IEEE (2017) 25. G. for Geeks, ML — Inception Network V1. https://www.geeksforgeeks.org/mlinception-network-v1/. Accessed 28 May 2021 26. Lin, M., Chen, Q., Yan, S.: Network in network (2013). arXiv preprint arXiv:1312.4400 27. Szegedy, C., et al.: Going deeper with convolutions. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1–9 (2015) 28. Ajit, A., Acharya, K., Samanta, A.: A review of convolutional neural networks. In: 2020 International Conference on Emerging Trends in Information Technology and Engineering (IC-ETITE), pp. 1–5. IEEE (2020) 29. Khan, A., Sohail, A., Zahoora, U., Qureshi, A.S.: A survey of the recent architectures of deep convolutional neural networks. Artif. Intell. Rev. 53(8), 5455–5516 (2020). https://doi.org/10.1007/s10462-020-09825-6 30. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2818–2826 (2016) 31. Tsang, S.-H.: Review: inception-v3-1st runner up (image classification). In: ILSVRC 2015, linea]. Disponible en (2018). https://bit.ly/2MKWI5k

Classification of Cognitive Ability from Multichannel EEG Signals Using Support Vector Machine Nilima Salankar(B) School of Computer Science, University of Petroleum and Energy Studies, Dehradun, India

Abstract. Nowadays stress is an important factor to deal with. In today’s scenario, everybody wants to excel in their domain area and failure to do so lead to severe consequences. In this study, authors have done an extensive analysis of brain signals captured while performing mental arithmetic test in sequences. Placement of 10–20 electrodes are done at frontal, parietal, temporal, central and occipital region. Hjorth Parameters activity, mobility and complexity have been extracted from the signals converted in frequency domain by discrete Fourier transform by using variable window size. Statistical test Mann Whitney U test with p < 0.05 has been carried out to decide the relevant basis in the discrete and real world correlation. Electroencephalogram (EEG) signals in the frequency range of 0– 30 Hz has been studied in the bands delta, theta, alpha, beta and gamma. Support vector machine (SVM) and hyper parameterized version of SVM has been used for the classification purpose. The regions which has exhibit prominent impact of cognitive task in both the male and female is in frontal and temporal region with high activation of delta waves while performing cognitive task. In absence of cognitive activity alpha waves towards lower limit has been found in poor performer while good performer has achieved alpha wave in higher domain. The accuracy achieved by hyper parameter tuned SVM is in range for 90–94% in the 12 classification sets. Keywords: EEG · Cognitive · SVM · Arithmetic test · Classification

1 Introduction Cognitive activity often leads to stress in certain category of population, lots of research has been carried out in the area of identification of related area of brain which shows the active involvement while performing cognitive task. It’s been proven track that there are certain areas of brain which are more active while performing the cognitive task. The performance of the performer strongly depends upon the preparation of the candidate which is strong in this domain. One of the authors have addressed that trained data and untrained data results into activation of different band in brain region [1] due to which known activity gives rise to the theta band and lower the alpha band in parietal and parietal occipital region. In another work, Zammouri et al. [2] has discussed about the band values theta and alpha variate to the lower side with the level of difficulty © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 391–401, 2022. https://doi.org/10.1007/978-3-030-96299-9_38

392

N. Salankar

in performing cognitive task, inversely proportional relationship has been presented. Further, Fischer et al. [3] have done the comparative EEG study between talented group of people and adolescent handicapped and fetched the results which lie in the area of different hemisphere symmetry. Afterwards, Ghali et al. [4] have studied the different age category while performing cognitive task of three level easy, medium and hard and concluded the non-dependent between bright and age of the subject. Cognitive capability has been explored from the mathematical perspective and correlation and anticorrelation has been establish between the various entities adopted while performing the cognitive task [5]. Zhban et al. discussed the reduction in the fatigue rate while not involving subjects in the same level of activity [6]. In [7], Kim and Seo have studied the underlying EEG signals while subjects performing the mental task which included the different processes like responding to visual stimulus, understanding, hand gesture control as well hitting the key. Further Osborne et al. [8] have discussed the impact of arithmetic activity in various types of subjects. Harmony et al. [9] addressed the difference in frequency capturing during different kind of activities which has shown correlation with speech, storage, long term and short term memory usage. In [10], Stroop color test and mental arithmetic have used to induced the stress in subjects and beta and alpha waves has shown the impact of the induced activities. Further, So et al. has discussed that various activities have strong impact on theta waves and on frontal region [11]. The impact of emotional task on alpha waves in temporal region whereas cognitive and cognitive task on beta waves in parietal region has been shown in [12]. Observing the EEG signal notify about the functional state of users brain and gives an update in the form neurofeedback which is very vital for brain functional understanding. Functional activity or concentration deficits are termed as attention deficit hyperactivity disorder or in some cases even though concentration is good but deficit can be found in the form of learning difficulties and comprehension of arithmetic in autism spectrum disorder [13]. In this paper, analysis of invoked EEG signals have been analyzed in the male and female category which further subcategorize to good performer and bad performer. The motivation for carrying this research is to identify the capability of stress management in case of good and bad performer. The contributions of this work are as: (i) An efficient classifier has been designed to distinguish between good and bad performer in terms of cognitive activity (mental arithmetic test) which is beneficial for the determination of stress managing capability of performer at early stage. (ii) Identification of the two correlated lobes temporal and frontal which is highly affected region for analysis of impact of cognitive activity on the good and bad performer which has not reported till time as per the authors best knowledge. Rest of the paper is structured as Methodology is discussed in Sect. 2 followed by Results in Sect. 3 and Discussion and Conclusion is presented in Sect. 4.

2 Methodology 2.1 Material and Methods The dataset used in this study (Goldberger et al., 2000) contains EEG recordings of subject before and during the cognitive activity (mental arithmetic test). The data is captured using Neurocom EEG-23 channel system (Ukraine, XAI-MEDICA). Electrodes used in

Classification of Cognitive Ability from Multichannel EEG Signals

393

this study are of category silver chloride and 10–20 standards have been used for the placement of electrodes on the scalp. Referencing used in experimentation is interconnected ear reference electrodes. The duration of the actual cognitive task recordings is 60 s. The data acquisition protocol is as shown in Fig. 1.

Fig. 1. Data acquisition (Goldberger et al., 2000)

Data is preprocessed by using Independent Component Analysis (ICA) to eliminate the artifacts (eyes, muscle, and cardiac overlapping of the cardiac pulsation). The cognitive task which subjects have performed includes serial subtraction of two numbers. Each trial started with the communication orally 4-digit (minuend) and 2-digit (subtrahend) numbers (e.g. 3141 and 42). The more details of the dataset can be found in (Goldberger et al., 2000). The captured data is labelled in 2 category good performer 24 subject with mean number of operations per 4 min = 21, standard deviation = 7.4 and bad performer 12 subjects with mean number of operation per 4 min = 7 with standard deviation 3.6. Subjects involved in this study are male and female with age category of 16–26. Channels Selected for analysis are frontal position FP1, FP2, F3, F4, F7, and F8, Temporal position T3, T4, T5, T6 Central position C3, C4; Parietal position P3, P4; Occipital position O1, O2. The purpose for the selection of channels which spread the entire skull is to investigate the difference in affected regions from the perspective of electrode position, in between good performer and bad performer, good performer female vs bad performer female, good performer male vs bad performer male, baseline vs good performer, baseline vs bad performer, baseline vs good performer female, baseline vs bad performer male, baseline vs good performer male, baseline vs bad performer male. The entire methodology followed in this work is as shown in Fig. 2. The EEG data used in this work hold total sample size is 1516/S thus total 1516 ∗ 60 samples have been analyzed with the help of variable window size.

EEG Data( 1516 data point s/sec )

Prep roces sing

Wind ow Size (20 sec)

Fourier Transform Lobes wise (Frontal, Temporal, Central, Parietal, Occipital)

Bands (Delta, Theta, Alpha, Beta and Gamma)

Hajorth Feature Extracon (Acvity, mobility and Complexity )

Fig. 2. Complete methodology

Stascal Test

SVM Classifier/Hype r parameter tuning

394

N. Salankar

The statistical significance of the extracted features have been tested with Mann whitely U test with p < 0.05. The features which has not exhibit any statistical significance with respect to combination class has been eliminated from the classification process. The classifier used in this study is Support vector machine for binary classification as well as binary multiclass classification. Support Vector Machine has used for classification, with hyper parameter tuning and model selection through cross validation. Hyper parameter optimization was done through grid search that works by searching exhaustively through a defined set of hyper parameter values. The best estimator, i.e. the combination of parameter values that gave the best cross validation accuracy, was finally used for training purposes. Grid search was preferred over random search, as even though random search has decreased process time, it doesn’t guarantee to find the optimal combination of hyper parameters. RBF kernel gave the best results with tuned values of regularization parameter C and kernel parameter γ. Effect of region wise channel selection was studied. And the performance metrics was calculated using accuracy and sensitivity. 2.2 Discrete Fourier Transform This transformation is useful to get insights of the details of the signals in frequency domain as time series of the signal is very complicated to fetch the details of the signal. To identify which ranges of frequencies are available inside the signal discrete Fourier transformation is very useful for analysis of the signal. It is a dot product between the proposed signal and sinusoidal wave’s particularly cosine on the real axis and sine on the imaginary axis. Change in the length of the signal doesn’t change the magnitudes except the length. Fourier coefficient is computed with the help of magnitude away from the center of the plane and given by the Eq. 1. Equation 2 depicts the conversion of real wave from time domain series to frequency domain.  √ 2 (1) Magnitude of Fourier Coefficient = a + b2 Where, a = Coordinate on Real axis b = Coordinate on Imaginary axis Fourier transform =

k=N −1 k=1

e

jπ ft

(2)

Where, ejπ ft = Euler  s function

3 Results The results extracted included the identification of prominent features among the three activity, mobility and complexity of Hjorth parameter as shown in Eqs. 3–5, which are

Classification of Cognitive Ability from Multichannel EEG Signals

395

significant to use to feed to the classifier. The box plot as shown in Fig. 3 indicates the statistical variance between the activity and mobility parameter fetched for male and female irrespective of the performance at frontal region. There is no overlapping in any quartile region which is good indication of significant usage of parameter for the classification purpose. In Fig. 4 the parameters are shown at temporal region. In this also there is significant variability has retrieved between male and female category. In Fig. 5, impact of baseline and cognitive task performed by the subject in frontal and temporal region. All these results are retrieved from the alpha and beta regions only. Activity = var(y(t))  Mobility =  Complexity =

var( dy(t) dt var(y(t))

Mobility( dy(t) dt (t) Mobility(y(t))

(3)

(4)

(5)

Where, y(t) = signal.

Fig. 3. Hjorth parameter for frontal region

Fig. 4. Hjorth parameter for temporal region

As shown in Table 1, 12 different subsets have tested for the understanding of behavior with respect to lobe and wave activation in response to cognitive task mental arithmetic in series. Binary and Multiclass binary classification is done. Sets 1–8 are binary classification are 1) Good Performer vs. Bad performer which includes both the male and female category. 2) Good Performer Male Vs. Bad Performer Male 3) Baseline Vs.

396

N. Salankar

Fig. 5. Hjorth parameter for frontal and temporal region

Good Performer Male 4) Baseline Vs. Bad Performer Male 5) Good Performer Female Vs. Bad Performer Female 6) Baseline Vs. Good Performer 7) Baseline Vs. Good Performer Female 8) Baseline Vs. Bad Performer Female. Sets 9–12 are multiclass binary classification which includes the combination of good as well bad performer in both the category and compared against the activation of wave and lobes in baseline activity. 9) Baseline Vs. Good Performer and Bad Performer 10) Baseline Vs. Good Performer Male and Bad Performer Male 11) Baseline Vs. Good Performer Female and Bad Performer Female 12) Baseline Vs. Good Performer Male and Good Performer Female. Classification is done by using the Baseline vs. Good Performer Male and Good Performer Female three types of refined version of SVM. Performance of the classifier has computed by accuracy and sensitivity parameters. Sensitivity talk about the probability of prediction of correct and FN in denominator signifies false negative, whereas accuracy deals with total computation of true positive true negative false positive and false negative as shown in Eqs. 6 and 7. TP ∗ 100% TP + FN

(6)

TP + TN ∗ 100% TP + FN + TN + FP

(7)

Sensitivity = Accuracy =

Table 1. Classification results Sets

Classifier

Good SVM Performer Vs. Bad SVM performer (Hyper parameter tuning) SVM (Lobes wise)

Frequency Acc range Beta; Alpha Beta; Alpha

F(Beta; Alpha) T(Beta; Alpha)

Sen

Sets

Classifier

Frequency Acc range

80.23 90.45 Good Performer 84.67 85.67 Female Vs. Bad Performer Female

SVM

Beta; Alpha

79.67 78.67

SVM Beta; (Hyper Alpha parameter tuning)

88.45 89.67

92.34 84.67

SVM (Lobes wise)

93.56 88.67

F(Beta; Alpha)

Sen

T(Beta; Alpha)

(continued)

Classification of Cognitive Ability from Multichannel EEG Signals

397

Table 1. (continued) Sets

Classifier

Frequency Acc range

Good Performer Male Vs. Bad Performer Male

SVM

Beta; Alpha

SVM Beta; (Hyper Alpha parameter tuning) SVM (Lobes wise)

Baseline SVM Vs. Good Performer SVM Male (Hyper parameter tuning) SVM (Lobes wise) Baseline SVM Vs. Bad Performer SVM Male (Hyper parameter tuning)

Baseline Vs. Good Performer and Bad Performer

F(Beta; Alpha)

Sets

Beta; Alpha Beta; Alpha

F(Beta; Alpha)

93.23 90

Beta; Alpha

F(Beta; Alpha)

SVM

Beta; Alpha

78.45 Baseline SVM Vs. Good 89.56 78.89 Performer SVM Female (Hyper parameter tuning) 91

89.67

F(Beta; Alpha) T(Beta; Alpha)

SVM (Lobes wise)

88.45 90

Baseline SVM Vs. Bad 93.67 92.67 Performer SVM Female (Hyper parameter tuning) 93

78.67

T(Beta; Alpha)

SVM Beta; (Hyper Alpha parameter tuning)

SVM (Lobes wise)

79

T(Beta; Alpha) Beta; Alpha

Classifier

79.67 84.67 Baseline SVM Vs. Good 87.89 79.89 Performer SVM (Hyper parameter tuning)

T(Beta; Alpha)

SVM (Lobes wise)

SVM (Lobes wise)

Sen

78.89 89.67 Baseline Vs. Good 89.56 78.89 Performer Female and Bad Performer Female 91.45 89.43

Frequency Acc range

Sen

Beta; Alpha

78.56 89.67

Beta; Alpha

89.94 89.78

F(Beta; Alpha)

91.45 90.78

T(Beta; Alpha) Beta; Alpha

80.67 87.67

Beta; Alpha

91.89 88.67

F(Beta; Alpha)

92.45 78.67

T(Beta; Alpha) Beta; Alpha

67

78.89

Beta; Alpha

78.67 89.7

SVM (Lobes wise)

F(Beta; Alpha)

89.67 87.67

SVM

Beta; Alpha

75.56 89.67

SVM Beta; (Hyper Alpha parameter tuning)

89.78 88.67

SVM (Lobes wise)

92.67 78.65

T(Beta; Alpha)

F(Beta; Alpha) T(Beta; Alpha)

(continued)

398

N. Salankar Table 1. (continued)

Sets

Classifier

Frequency Acc range

Baseline Vs. Good Performer Male and Bad Performer Male

SVM

Beta; Alpha

SVM Beta; (Hyper Alpha parameter tuning) SVM (Lobes wise)

F(Beta; Alpha)

Sen

76

Sets

78.78 Baseline Vs. Good 89.45 89.67 Performer Male and Good Performer Female 91.78 89.78

T(Beta; Alpha)

Classifier

Frequency Acc range

Sen

SVM

Beta; Alpha

89.45

79

SVM Beta; (Hyper Alpha parameter tuning)

78.98 85

SVM (Lobes wise)

89.56 89.56

F(Beta; Alpha) T(Beta; Alpha)

4 Discussion and Conclusion Long time involvement in an activity results implication either in positive or negative way. Mental fatigue is the result of long time arithmetic effect as in [29] but in our study the dataset used is designed from very short time silent arithmetic activity and bad performance is not because of the mental fatigue to the extent. The most popular algorithm for the classification purpose SVM has reported classification accuracy 91%. Timely identification of mental stress and ability to cope up with mental stress can saves the life as it is identified as one of the major contributing factor which leads to several disease related to heart attack and stroke [21] and for that purpose arithmetic task activities have been performed. Once acquainted with stress markers by analysing EEG signals it’s very easy for the analysis of brain state and timely cure can be possible [10]. The results extracted have a significant role in understanding the cognitive ability of the performer. The beta waves which gets prominently visible while performing the work with concentration or focused indication, which is shown in both the category of male and female as well as good and bad performer, which gives the intuition of focused behavior of the subject category. But the significant difference has found in the peak of alpha waves in the baseline condition. It has shown high peak in the case of good performer both male and female but has shown significantly less involvement in the resting state which gives an intuition that bad performers are slightly less involved in the case of resting which is indication of stressed behavior and ultimately has shown impact in the performance. The region which is strongly involved in case of all categories are frontal and temporal while very less significance has been observed into the central and occipital region which has intuition that mostly cognitive activity arises at the frontal and temporal lobes. As capturing EEG signals from all the lobes is very tedious task correlation between the activated lobes has been evaluated and results are very interesting and depicted by the high and low range of alpha waves in frontal and temporal regions is directly correlated with good and bad performer. High state of relaxation of state of mind leads to good performer whereas low state of relaxation of state of mind leads

Classification of Cognitive Ability from Multichannel EEG Signals

399

to bad performer. The classification accuracy and sensitivity is studied by using three versions of SVM. In SVM hyper-parameter tuning version the accuracy achieved for all the 12 subsets lies in the range 91%–92.34% at frontal temporal region which is aligned with result obtained in [12] and very promising results are achieved whereas long mental arithmetic has achieved only. In case of SVM without tuning the accuracy achieved for all the subsets lies in the range of 67.00%–84.25%. The less accuracy is shown at the subset baseline vs. bad performer female whereas the baseline vs bad performer male has achieved the accuracy 88.45% which gives the insight that bad performer female are more restless and prone to stress activation easily as compared to male participants. This study has given a clear insight of identification of stress level without being actually exposing a subject to stress environment. So the anxiety before an examination or any important task is as harmful as anxiety during the actual execution. So the appearance of a peak in alpha level during the rest condition is best way to determine the stress level in case of cognitive ability of the subject.

References 1. Grabner, R.H., De Smedt, B.: Oscillatory EEG correlates of arithmetic strategies: a training study. Front. Psychol. (2012). https://doi.org/10.3389/fpsyg.2012.00428 2. Zammouri, A., Ait Moussa, A., Mebrouk, Y.: Brain-computer interface for workload estimation: assessment of mental efforts in learning processes. Expert Syst. Appl. 112 (2018). https://doi.org/10.1016/j.eswa.2018.06.027 3. Fischer, D.G., Hunt, D., Randhawa, B.S.: Spontaneous EEG correlates of intellectual functioning in talented and handicapped adolescents. Percept. Mot. Skills 54(3), 751–762 (1982). https://doi.org/10.2466/pms.1982.54.3.751 4. Ghali, R., BenAbdessalem, H., Frasson, C., Nkambou, R.: Identifying brain characteristics of bright students. J. Intell. Learn. Syst. Appl. 10(3), 93103 (2018). https://doi.org/10.4236/ jilsa.2018.103006 5. Mohanavelu, K., Vishnupriya, R., Poonguzhali, S., Adalarasu, K., Nathiya, N.: Mathematical models for predicting cognitive workload. Int. J. Pure Appl. Math. (2018) 6. Zhban, E.S., Likhanov, M.V., Zakharov, I.M., Bezrukova, E.M., Malykh, S.B.: The role of mathematical and trait anxiety in mental fatigue: an EEG investigation. Psychol. Russ. State Art (2018). https://doi.org/10.11621/pir.2018.0406 7. Kim, M.S., Seo, H.D.: Analysis of EEG signals during mental tasks. Proc. Int. Conf. Math. Eng. Tech. Med. Biol. Sci. (2003) 8. Osborne, P.G., Chou, T.S., Shen, T.W.: Characterization of the psychological, physiological and EEG profile of acute betel quid intoxication in naïve subjects. PLoS ONE (2011). https:// doi.org/10.1371/journal.pone.0023874 9. Harmony, T., et al.: Do specific EEG frequencies indicate different processes during mental calculation? Neurosci. Lett. (1999). https://doi.org/10.1016/S0304-3940(99)00244-X 10. Jun, G., Smitha, K.G.: EEG based stress level identification (2017). https://doi.org/10.1109/ SMC.2016.7844738 11. So, W.K.Y., Wong, S.W.H., Mak, J.N., Chan, R.H.M.: An evaluation of mental workload with frontal EEG. PLoS ONE (2017). https://doi.org/10.1371/journal.pone.0174949 12. Ray, W.J., Cole, H.W.: EEG alpha activity reflects attentional demands, and beta activity reflects emotional and cognitive processes. Science 228(4700), 750–752 (1985). https://doi. org/10.1126/science.3992243

400

N. Salankar

13. Wang, Q., Sourina, O.: Real-time mental arithmetic task recognition from EEG signals. IEEE Trans. Neural Syst. Rehabil. Eng. (2013). https://doi.org/10.1109/TNSRE.2012.2236576 14. Li, Y., et al.: Abnormal EEG complexity in patients with schizophrenia and depression. Clin. Neurophysiol. (2008). https://doi.org/10.1016/j.clinph.2008.01.104 15. Trejo, L.J., et al.: EEG-based estimation of mental fatigue: convergent evidence for a threestate model. In: Schmorrow, D.D., Reeves, L.M. (eds.) FAC 2007. LNCS (LNAI), vol. 4565, pp. 201–211. Springer, Heidelberg (2007). https://doi.org/10.1007/978-3-540-73216-7_23 16. Matsuoka, H., et al.: Neuropsychological EEG activation in patients with epilepsy. Brain (2000). https://doi.org/10.1093/brain/123.2.318 17. Fulmare, N.S., Chakrabarti, P., Yadav, D.: Understanding and estimation of emotional expression using acoustic analysis of natural speech. Int. J. Nat. Lang. Comput. (2013). https://doi. org/10.5121/ijnlc.2013.2503 18. Salankar, N., Chaurasia, S., Prasad, A.: Modelling of human emotion using analysis of natural speech using refinement approach (2017). https://doi.org/10.1109/NGCT.2016.7877407 19. Jawad Khan, M., Hong, M.J., Hong, K.S.: Decoding of four movement directions using hybrid NIRS-EEG brain-computer interface. Front. Hum. Neurosci. 8 (2014). https://doi.org/ 10.3389/fnhum.2014.00244 20. Willoughby, J.O., et al.: Mental tasks induce gamma EEG with reduced responsiveness in primary generalized epilepsies. Epilepsia (2003). https://doi.org/10.1046/j.1528-1157.2003. 20103.x 21. Alshargie, F.M., Tang, T.B., Badruddin, N., Kiguchi, M.: Mental stress quantification using EEG signals. In: Ibrahim, F., Usman, J., Mohktar, M.S., Ahmad, M.Y. (eds.) International Conference for Innovation in Biomedical Engineering and Life Sciences. IP, vol. 56, pp. 15– 19. Springer, Singapore (2016). https://doi.org/10.1007/978-981-10-0266-3_4 22. AlShargie, F., Kiguchi, M., Badruddin, N., Dass, S.C., Hani, A.F.M., Tang, T.B.: Mental stress assessment using simultaneous measurement of EEG and fNIRS. Biomed. Opt. Express (2016). https://doi.org/10.1364/boe.7.003882 23. Harrison, A.H., Noseworthy, M.D., Reilly, J.P., Guan, W., Connolly, J.F.: EEG and fMRI agree: mental arithmetic is the easiest form of imagery to detect. Conscious. Cogn. (2017). https://doi.org/10.1016/j.concog.2016.10.006 24. Salankar, N., Mishra, A.: Statistical feature selection approach for classification of emotions from speech. SSRN Electron. J. (2020). https://doi.org/10.2139/ssrn.3527262 25. Ryu, K., Myung, R.: Evaluation of mental workload with a combined measure based on physiological indices during a dual task of tracking and mental arithmetic. Int. J. Ind. Ergon. (2005). https://doi.org/10.1016/j.ergon.2005.04.005 26. Sammer, G., et al.: Relationship between regional hemodynamic activity and simultaneously recorded EEG-theta associated with mental arithmetic-induced workload. Hum. Brain Mapp. (2007). https://doi.org/10.1002/hbm.20309 27. Kristeva-Feige, R., Fritsch, C., Timmer, J., Lücking, C.H.: Effects of attention and precision of exerted force on beta range EEG-EMG synchronization during a maintained motor contraction task. Clin. Neurophysiol. (2002). https://doi.org/10.1016/S1388-2457(01)00722-2 28. Miwakeichi, F., MartínezMontes, E., Valdés-Sosa, P.A., Nishiyama, N., Mizuhara, H., Yamaguchi, Y.: Decomposing EEG data into space-time-frequency components using Parallel Factor Analysis. Neuroimage (2004). https://doi.org/10.1016/j.neuroimage.2004.03.039 29. Zhang, C., Yu, X.: Estimating mental fatigue based on electroencephalogram and heart rate variability. Polish J. Med. Phys. Eng. (2010). https://doi.org/10.2478/v10013-010-0007-7

Classification of Cognitive Ability from Multichannel EEG Signals

401

30. Katahira, K., Yamazaki, Y., Yamaoka, C., Ozaki, H., Nakagawa, S., Nagata, N.: EEG correlates of the flow state: A combination of increased frontal theta and moderate frontocentral alpha rhythm in the mental arithmetic task. Front. Psychol. (2018). https://doi.org/10.3389/fpsyg. 2018.00300 31. Salankar, N., Mishra, P., Garg, L.: Emotion recognition from EEG signals using empirical mode decomposition and second-order difference plot. Biomed. Signal Process. Control (2021). https://doi.org/10.1016/j.bspc.2020.102389

Heterogeneous DBSCAN for Emergency Call Management: A Case Study of COVID-19 Calls Based on Hospitals Distribution in Saudi Arabia Naila Aziza Houacine(B) , Lydia Sonia Bendimerad, and Habiba Drias LRIA, USTHB, BP 32 El-Alia Bab-Ezzouar, Algiers, Algeria {nhouacine,lbendimerad,hdrias}@usthb.dz

Abstract. Unsupervised machine learning and especially unshaped clustering approaches like Density-Based Spatial Clustering of Applications with Noise (DBSCAN) remain at the center of researchers’ attention. Many DBSCAN improvements have been recently proposed, but none of them involve the data heterogeneity that we may find in real-life problems. To tackle this issue, in this paper, a novel DBSCAN-based approach is proposed and applied to real data of Saudi Arabia (KSA), in order to cluster the COVID-19 calls based on hospitals distribution for the resolution of the ambulances dispatching problem and COVID-19 emergency calls covering. The designed approach is called Heterogeneous DBSCAN (HDBSCAN). It considers different data types: Statically distributed hospitals over Saudi Arabia and dynamic COVID-19 incoming calls. The experiments show that the proposed HDBSCAN greatly speeds up the original DBSCAN and provides an efficient classification for the application problem. Keywords: Heterogeneous data · Real-time clustering · Heterogeneous DBSCAN · DBSCAN · Emergency calls · Saudi Arabia COVID-19

1

Introduction

In the field of machine learning, talking about “Clustering” is synonymous with unsupervised learning methods. Clustering is the process of grouping a set of data objects into multiple groups called clusters. Objects within a cluster have high similarities but are very dissimilar to objects in other clusters. This method is widely used in several fields such as image processing, Geography, Molecular biology, Social Networks and Multimedia, Telecommunication Industry, and so on [2]. Partitioning, Hierarchical, Density-based, and Grid-based methods are clustering techniques that allow finding shaped and unshaped (arbitrary shape) clusters in a static dataset. Density-based methods are characterized by arbitrary shape clusters forming dense regions of objects in space that are separated by low-density regions and their ability to filter out outliers. These techniques c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 402–411, 2022. https://doi.org/10.1007/978-3-030-96299-9_39

HDBSCAN for Emergency Call Management

403

are DBSCAN (Density-Based Spatial Clustering of Applications with Noise), OPTICS (Ordering points to identify the clustering structure), and DENCLUE (DENsity-based CLUstEring). Real-life problems pose obstacles for algorithms dealing with simple and static data because the real world is based on dependent and different data that evolve over time (Dynamic). This prompted us to propose a new clustering approach that takes into account the diversity and heterogeneity of realtime data and its dynamicity. Also, Density-Based Spatial Clustering of Applications with Noise approach is a widely used algorithm in literature and has been improved in several ways to overcome its limitations, among the most recent ones we found ECR-DBSCAN [5], 3W-DBSCAN [11], Quantum DBSCAN [10], Improved DBSCAN [8], Block-DBSCAN [3], a clustering learning approach based on DBSCAN and Density Peak Algorithm [7]. That’s what inspired us to propose the improved DBSCAN method, namely Heterogeneous DBSCAN and adapt it to a dynamic problem. Ambulance dispatching and covering problems are real-time processing problems. It is defined as a set of bases receiving a random number of emergency calls during the day. These calls should be answered in a brief time, without dispatching ambulances to far away located calls. The idea of this paper is to study the Global Positioning System (GPS) of static bases and dynamic incoming calls of COVID-19 in order to group a finite number of these emergency calls in clusters containing a minimum number of bases and thus delimit the ambulance dispatch space and avoid wasted time. This study will compare the results of the original DBSCAN algorithm and its proposed improvement applied to Saudi Arabia Data since it’s a country that receives so many visitors during the pelting period. The remainder of this paper is organized as follows. Section 2 presents some definitions of the classical DBSCAN and its Algorithm. The proposed Heterogeneous DBSCAN is introduced and described in Sect. 3 as well as its application in real-time. Section 4 presents the experimental study, including dataset presentation and comparison between the proposed approach and the original DBSCAN. Finally, Sect. 5 concludes the paper.

2

Original DBSCAN

Density-Based Spatial Clustering of Applications with Noise [4], commonly called DBSCAN, is widely used to cluster a lot of data in arbitrary shapes and sizes. It is based on two main parameters, which are the minimum number of points in a cluster MinPts and the distance that group data points in a cluster epsilon (). In the following, some key definitions of DBSCAN are presented [6]: Definition 1. (−neighborhood): In a search-space D, it represents a collection of points within a radius  centered at a point q, and it is defined as N (q) = {p ∈ D|Distance(p, q) ≤ }. Definition 2. (Density): represents the number of points in the –neighborhood of a point q.

404

N. A. Houacine et al.

Definition 3. (Core Point): In a search-space D, a core point corresponds to each data point with a density higher than M inP ts. Definition 4. (directly density-reachable): A point p is directly densityreachable from a point q, if and only if q is a core point and p belongs to the –neighborhood of q. Definition 5. (density-reachable): p is density-reachable from q, if there is a chain of points p1 , ..., pn , such that i ∈ [1, n], p1 = q, pn = p, and where pi+1 is directly density-reachable from pi . The original version of DBSCAN is summarized in Algorithm 1. The  − neighborhood of each unvisited point p of the dataset is calculated either to mark it as a Noise if it has a density lower than MinPts or to create a new cluster and launches the ExpandCluster() function described in Algorithm 2. This function allows adding the directly density-reachable or density-reachable points from P to the current cluster. Algorithm 1. Original DBSCAN Input: D: Dataset , (, MinPts) : Empirical Parameters. begin Mark all the points of D as Unvisited. for each Unvisited point P of D do Mark P as visited; Pts Neighbors = −nighborhood (D, P, ); if Size ( Pts Neighbors) < MinPts then Mark P as NOISE; else C = Create Cluster (); ExpandCluster (D, P, Pts Neighbors, C, , MinPts);

3 3.1

Proposed Approach Heterogeneous-DBSCAN

The following contribution consists of a Heterogeneous version of the Original DBSCAN. As its name suggests, this algorithm is intended to form clusters on a dataset D composed of different types of points. In our case, we consider twopoint types, let’s refer to them as the ReferencePoints and the RegularPoints. In addition to the two parameters that DBSCAN is equipped with, the Heterogeneous version needs one extra parameter, MaxReferenceRate. This parameter designates the maximum number of ReferencePoints per cluster, such as each cluster must have at least one ReferencePoint and one RegularPoint and a maximum of MaxReferenceRate ReferencePoints.

HDBSCAN for Emergency Call Management

405

Algorithm 2. ExpandCluster Input: D: Dataset, P: a point of D, Pts Neighbors: list of P’s neighbors, C: the current cluster, (, MinPts): Empirical Parameters. begin Add P to Cluster C; for each point P’ in Pts Neighbors do if P’ is Unvisited then Mark P’ as visited; Pts Neighbors’ =  − nighborhood (D, P’, ); if Size ( Pts Neighbors’) ≥ MinPts then Pts Neighbors = Pts Neighbors ∪ Pts Neighbors’ ; if P’ is Unclassified then Add P’ to Cluster C;

As shown in Algorithm 3, the dataset D is explored, such as for each unvisited RegularPoint P, Its  − neighborhood (considering both regular and reference points) is retrieved and saved in the list Pts Neighbors. If this list contains less than MinPts points or only one type of points, then P is marked as NOISE. Otherwise, P is assigned to a cluster, and for that, the ExpandHeteroCluster() function is executed. Algorithm 3. Heterogeneous-DBSCAN Input: D: Dataset , (, MinPts, MaxReferenceRate) : Empirical Parameters. begin Mark all the points of D as Unvisited. for each Unvisited RegularPoints P of D do Mark P as visited; Pts Neighbors = −nighborhood (D, P, ); if Size ( Pts Neighbors) < MinPts or nbrRegularPoints ( Pts Neighbors) = 0 or nbrReferencePoints ( Pts Neighbors) = 0 then Mark P as NOISE; else C = Create Cluster (); ExpandHeteroCluster (D, P, Pts Neighbors, C, , MinPts, MaxReferenceRate);

The ExpandHeteroCluster() algorithm browses the  − N eighborhood of the RegularPoint P. It checks if each neighbor is a RegularPoint otherwise, it’s a ReferencePoint and should verify that the Rate of ReferencePoints in the current cluster C does not exceed maxReferenceRate. If this condition is verified and the point is unvisited, then the neighbors’ list is enriched by the neighbors of P ’. Finally, if the current point P ’ is unclassified, it joins cluster C, and the reference rate of C is updated. Algorithm 4 presents details of this function.

406

N. A. Houacine et al.

Algorithm 4. ExpandHeteroCluster Input: D: Dataset, P: a RegularPoint of D, Pts Neighbors: list of P’s neighbors, C: the current cluster, (, MinPts, MaxReferenceRate): Empirical Parameters. begin Add P to Cluster C; for each point P’ in Pts Neighbors do if P’ is a RegularPoint or ReferencePointsRate ( C) < MaxReferenceRate then if P’ is Unvisited then Mark P’ as visited; Pts Neighbors’ =  − nighborhood (D, P’, ); if Size ( Pts Neighbors’) ≥ MinPts then Pts Neighbors = Pts Neighbors ∪ Pts Neighbors’ ; if P’ is Unclassified then Add P’ to Cluster C; if P’ is a ReferencePoint then Update ReferencePoints Rate in Cluster C;

3.2

Dynamicity Modeling for DBSCAN and HDBSCAN

In order to tackle real-world dynamic problems, the data should be clusterized in real-time. Thus DBSCAN and HDBSCAN are applied in a context defined by a random number of incoming data in several units of time t where each time DBSCAN or DHDBSCAN should be launched on the data of the current period as shown in Algorithm 5. Algorithm 5. Dynamic Clustering Approach Input: (MinPts, , maxReferenceRate) : Empirical parameters, Dataset : List of data per time unit, Unit times : total number of time units. begin for each unit of time t in Unit times do Update Dataset: Incoming data (P)= Incoming data (P-1) ∪ Dataset (P); Launch Clustering Approach(Empirical parameters, Incoming data (P));

4 4.1

Experiments and Results Dataset

The dataset used for the experiments of the proposed method is based on the real case study of COVID-19 daily cases in Saudi Arabia. From the data provided by the KSA in [9], we considered the day with the highest number of new cases (June 17, 2020) to build our dataset. Hence our dataset contains 4919 instances,

HDBSCAN for Emergency Call Management

407

Fig. 1. Sample of distribution of incoming calls over 24 h.

where each line has the form . In the perspective of getting effective experiment results, we generated 5 datasets where the positions of the emergency calls are randomly generated and distributed along the 24 h of the day. So they receive a certain number of calls at each unit of time t, in our case t = 10 min, which makes 144 time units in one day. Figure 1 shows one distribution (dataset) from the 5 generated datasets. In addition to the calls, our dataset contains the positions of the hospitals [1] where the emergency vehicles belong. As illustrated in Fig. 2, the KSA possesses 278 available hospitals spread over 13 regions. But considering the regions of the 4919 COVID-19 cases on June 17, 2020, it shows that the cases are not equally split into the different regions. We can observe from Fig. 3 that the new cases submerge “Eastern”, “Mecca”, and “Riyadh”. This is where a heterogeneous clustering approach may effectively divide the KSA area by grouping together the emergency calls and the adequate hospitals, forming well-balanced heterogeneous regions. The formation of the clusters requires a distance calculation formula, the most suitable one when dealing with Longitude (γ) and Latitude (δ) is the spherical distance defined in Eq. (1). Note that R is the radius of earth and is equal to 6371 km.   δi − δj 2 γi − γj 2 ) + cos(δi ). cos(δj ). sin( ) .R (1) sin( S = arcsin 2 2

4.2

Results

Comparison Metrics: As in our case, when the ground truth is not available, the clustering quality should be evaluated through intrinsic methods that estimate how well the clusters are separated and coherent. The silhouette coefficient

408

N. A. Houacine et al.

Fig. 2. Number of hospitals per Region.

Fig. 3. Number of cases in Saudi Arabia, June 17, 2020.

is one of those metrics. Its formula is given in Eq. (2) where for a dataset D, for each point p ∈ D, the mean distance between p and the other points of its own cluster are computed as a(p), and the minimum mean distance of p to the other clusters as b(p) [6]. Sglobal =

|D|  i=1

s(pi ) where s(pi ) =

(b(pi) − a(pi)) max{a(pi ), b(pi )}

(2)

An s(p) value is bounded by −1 and 1. Contrary to an s(p) close to −1, an s(p) close to 1 reflects a good clustering, and a near zero s(p) signifies that p is at the border of two potential clusters. Hence, the larger the Sglobal is, the better the clustering quality. Additionally to the silhouette measure, we compare the three approaches based on their execution time (s), the number of outliers (Noise), and the number of clusters formed. Results’ Comparison: To treat the data either with DBSCAN algorithm or Heterogeneous-DBSCAN incites to vary the parameters MinPts and  as well as MaxReferenceRate for HDBSCAN. In this case of studies, the parameters MinPts and , respectively, represent the number of emergency calls and the distance between these calls for DBSCAN. However, for HDBSCAN, which considers 2 types of data (Hospitals and Calls), MinPts represent the minimum number of hospitals and calls in an –neighborhood, as for the parameter MaxReferenceRate, it represents the maximum rate of hospitals compared to the calls’ rate considered in the –neighborhood. The best parameters found for DBSCAN are MinPts = 2 and  = 30 km, and for HDBSCAN they are MinPts = 2,  = 50 km, and MaxReferenceRate = 3%. Figures 4, 5, 6, and 7 depicts the comparative results of the two DBSCAN versions through the dynamic emergency calls arrival. It’s drawn from the average performances achieved on the 5 datasets and on the basis of each of the four used metrics.

HDBSCAN for Emergency Call Management

409

Fig. 4. Evolution of the number of clusters within the 144 times unit.

Fig. 5. Evolution of the number of outliers within the 144 times unit.

Fig. 6. Evolution of the silhouette metric within the 144 times unit.

Fig. 7. Evolution of the execution time within the 144 times unit.

Figure 4 shows that HDBSCAN remains consistent in the number of clusters formed through the arrival of new emergency calls. Unlike Original DBSCAN, whose number of clusters varies abnormally and exceeds 130 clusters for the 32-th times unit. The results of Figs. 5 and 6 are closely related, such as with the large number of calls classified as NOISE by the original DBSCAN, the performed silhouette reached high values. Indeed, a high silhouette score is attractive, but not at the expense of a high number of outliers, especially for our application of dispatching and covering problems. Thus, the HDBSCAN’s outcomes point out much less noise with an acceptable positive silhouette score. From Fig. 7, we can observe that the proposed HDBSCAN compared to the original DBSCAN, has a significantly faster clustering capability, despite considering 2 kinds of points to cluster. Figures 8, 9, and 10 exhibit the resulting clustering of the hospitals and emergency calls according to the Heterogeneous-DBSCAN policy at 1, 75, and 144 units of time, respectively. The build clusters show a balance between the calls rate and the hospitals’ rate, in general. We can distinguish the formation of one large cluster when the number of calls increases (cluster n◦ 1 Fig. 9 and 10), which is much more prominent than the other clusters. This is caused by the

410

N. A. Houacine et al. Capon:

Capon:

Fig. 8. Distribution of heterogeneous clusters at time unit = 1.

Fig. 9. Distribution of heterogeneous clusters at time unit = 75.

Capon:

Fig. 10. Distribution of heterogeneous clusters at time unit = 144.

elevated population density of those regions, which implied the emergence of a high COVID-19 emergency calls density. However, as shown in the KSA’s map, the HDBSCAN has grouped a sufficient number of hospitals to face that calls density. Therefore, in each cluster, an adequate number of hospitals are made available in so that each cluster can independently manage the dispatching of their resources to cover their own COVID-19 emergency calls. Also, we can remark that the south-east (El-hufuf ) region of KSA does not have any health infrastructure or hospital that may relieve the load on cluster n◦ 1.

5

Conclusion

The present paper proposes an improved version of a well-known density-based clustering algorithm (DBSCAN) in order to help rescue teams responding to COVID-19 emergency calls. The proposed approach called HDBSCAN is based

HDBSCAN for Emergency Call Management

411

on different types of data points. It works on clustering incoming emergency calls into clusters containing a number of hospitals proportional to the number of calls in the cluster (according to a MaxReferenceRate). In addition to this contribution, the current paper compares the Original DBSCAN and HDBSCAN algorithms on real data of Saudi Arabia on a dynamic framework. The dynamicity here consists of clustering the incoming COVID-19 calls at different time units representing the flow of time during the 24 h of the day. The results obtained by the proposed HDBSCAN approach are very encouraging compared to those of original DBSCAN in this context. Acknowledgements. We would like to express our special thanks of gratitude to Prince Mohammad Bin Fahd Center for Futuristic Studies for the support of this work.

References 1. Accredited health service providers mar2021 (2021). https://data.gov.sa/Data/en/ dataset/accredited-health-service-providers mar2021 2. Bhattacharjee, P., Mitra, P.: A survey of density based clustering algorithms. Front. Comput. Sci. 15(1), 151308 (2020). https://doi.org/10.1007/s11704-019-9059-3 3. Chen, Y., et al.: BLOCK-DBSCAN: fast clustering for large scale data. Pattern Recogn. 109, 107624 (2021). https://doi.org/10.1016/j.patcog.2020.107624 4. Ester, M., Kriegel, H.P., Sander, J., Xu, X.: A density-based algorithm for discovering clusters in large spatial databases with noise. In: KDD, pp. 226-231 (1996) 5. Giri, K., Biswas, T.K., Sarkar, P.: ECR-DBSCAN: an improved DBSCAN based on computational geometry. Mach. Learn. App. 6, 100148 (2021). https://doi.org/ 10.1016/j.mlwa.2021.100148 6. Han, J., Kamber, M., Jian, P.: Data Mining: Concepts and Techniques. Morgan Kaufmann Publishers, San Francisco (2012) 7. Li, M., Bi, X., Wang, L., Han, X.: A method of two-stage clustering learning based on improved DBSCAN and density peak algorithm. Comput. Commun. 167, 75–84 (2021). https://doi.org/10.1016/j.comcom.2020.12.019 8. Li, S.S.: An improved DBSCAN algorithm based on the neighbor similarity and fast nearest neighbor query. IEEE Access 8, 47468–47476 (2020). https://doi.org/ 10.1109/access.2020.2972034 9. Rami, K.: Saudi Arabia coronavirus disease (COVID-19) situation (2021). https://datasource.kapsarc.org/explore/dataset/saudi-arabia-coronavirusdisease-covid-19-situation/export/?disjunctive.daily cumulative&disjunctive. indicator&disjunctive.event&disjunctive.city en&disjunctive.region en&sort=region en 10. Xie, X., Duan, L., Qiu, T., Li, J.: Quantum algorithm for MMNG-based DBSCAN. Sci. Rep. 11(1), 15559 (2021) 11. Yu, H., Chen, L., Yao, J., Wang, X.: A three-way clustering method based on an improved DBSCAN algorithm. Phys. A Statist. Mech. App. 535, 122289 (2019). https://doi.org/10.1016/j.physa.2019.122289

A Survey on the Quality of Service and Metaheuristic Based Resolution Methods for Multi-cloud IoT Service Selection Ahmed Zebouchi(B) and Youcef Aklouf RIIMA, USTHB, BP 32 El-Alia Bab-Ezzouar, Algiers, Algeria {azebouchi,yaklouf}@usthb.dz Abstract. The Internet-of-Things (IoT) generate increasingly significant amount of data that needs to be stored and analysed. The use of IoT devices as a service makes it more accessible and exploitable, this could be achieved using of cloud computing. Multi-cloud service composition and selection are required to fulfill increasingly complicated user requests for services. A service request is made from a cloud broker to cloud providers (CP) to deliver the required Quality of Service (QoS). Selecting services and optimizing service compositions to satisfy functional and non-functional conflicting requirements across various cloud service providers is an non-deterministic polynomial-time hardness problem (NP-hard). Multiobjective (MO) metaheuristics are known to be performant to solve such a problem. This study examines how to select IoT services to achieve the best performances on the eight selected QoS across multiple CP. The experiment results reveal that among the 18 compared algorithms, the parallel NSGAII provides the most efficient and optimal outcomes. Keywords: Internet of things · Multi-objective metaheuristics Multi-cloud · Cloud brokering · Quality of service · Service composition · Pareto front

1

·

Introduction

The IoT is a widely-adopted domain application. Nowadays, the internet of things is becoming the trademark technology to provide specific services and enable an intelligent world. The world will see a tripling of internet-connected devices in the next decade to 30 billion by 2030 [9]. The IoT system should then exploit the advantage of cloud computing using service oriented architecture [4]. Therefore, IoT can take advantage of the cloud paradigm’s virtually unlimited capabilities and resources to compensate for its technological constraints. Cloud service providers are increasing and offering a wide range of services. Despite this, a given cloud provider likely won’t be capable of meeting all of the consumer’s requirements [28]. The multi-cloud c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 412–424, 2022. https://doi.org/10.1007/978-3-030-96299-9_40

A Survey on the QoS and MO Metaheuristics for IoT

413

composition/selection is crucial to fulfilling a complicated user request, or when the requested service is unavailable. The collaboration of several service providers models the Broker-based service selection; this formulates what so-called Multiclouds IoT Environment. What makes the request fulfillment more complex is that users generally focus on some specific Quality of Service, like the cost of the service or response time, with little or no consideration for the service location or energy efficiency. On the other hand, the cloud broker focuses on cloud profit. These QoS are objectives that the cloud broker goals to minimize or maximize. Some of these objectives are conflicting, as in cost minimization and response time minimization. This process is known for its complexity and is classified as an NP-hard problem. Such problems are usually approached using metaheuristics, such as bio inspired Pareto-based metaheuristics used for multiobjective problems. Due to the worldwide popularity and success that metaheuristics have known and the increasing publication counts of these studies. In this survey, we review eighteen MO metaheuristic algorithms. We consider a need for a survey to review and summarize the most appealing current studies on IoT service selection optimization problems. Hence, as a first contribution, we have focused on studying distinguished metaheuristic algorithms that we term as the “new generation” metaheuristic algorithms. Also, through the studied papers, we noticed that most of the researches on that issue has considered QoS optimization. Each study has only focused on a few QoS, ignoring other important ones. In our research, we realized the necessity of gathering all QoS considered in service selection to effectively compare the existing solving methods and ensure the best service level agreement, which constitutes the second contribution of this paper. The rest of the paper is presented as follows. The next Section is dedicated to raising the QoS that are considered in IoT service selection. In Sect. 3, we expose a brief description of MO metaheuristic algorithms functioning. In Sect. 4, we reveal the experimental results of applying the algorithms on IoT service selection. Finally, Sect. 6 concludes the paper.

2

Quality of Services in IoT

Our work covers the QoS found in the most recent and relevant state-of-the-art researches. We give a brief overview of each one below. 2.1

Energy Consumption

The required energy for IoT service computation is of high importance in the QoS measurement. As mentioned in [11,24,26], it is considered a major issue in cloud computing. Thus, service providers aim to complete the job with minimal energy consumption efficiently. Assuming E(Sij ) is the energy consumption used by service i to perform the job by cloud provider j and N is the number of services included in the service composition. The service N provider’s total energy usage is computed within the equation: Eglobal = i=1 E(Sij )

414

2.2

A. Zebouchi and Y. Aklouf

Cost

With the rise of CP numbers and each one offers a tremendous number of services, the cost of some services became a key element of the QoS. Thus, most recent researches as in [24,32–34] focused on minimizing cloud prices. The calculation ofthe global cost of a cloud service composition can be formulated as N Cglobal = i=1 C(Sij ). Where C is the cost of service i from provider j. 2.3

Profit

Profit is a monetary amount that cloud providers can be rewarded with. The execution cost of the service is the fee that a tenant needs to pay for invoking some operations. The broker is expected to profit from the task of finding the best solution. The broker’s profit is, therefore, considered as a second goal. Profit is N considered in [3,10,31] and calculated by Pglobal = i=1 P (Sij ) − C(Sij ). Where P (Sij ) is the consumer price of service i of the service provider j and C(Sij ) is the service provider value of service i provided by cloud j. 2.4

Availability

Availability represents the uptime of cloud service during a specified time interval. In [1,14,25], authors judged the Availability criteria as an impacting parameter that must be taken into account. The Availability A(Sij ) of service i in cloud j is calculated as A = t/ts where t and ts represent the uptime and the total time of service. As the value of A approaches Nto 1, availability increases. The overall availability is calculated as Aglobal = i=1 A(Sij ). 2.5

Reliability

The Reliability calculates the level of assertion free of any cloud service software or hardware fault. Several authors [1,14,23,34] took into account this requirement for its and its importance. The total reliability is calculated with relevance N Rglobal = i=1 R(Sij ), where R(Sij ) function returns the reliability of service i in cloud j, as the value of R approaches 1, the reliability increases. 2.6

Response Time

Response time is defined as the time required to send a request and receive the response from the service. Numerous articles, for instance [1,14,25,32–34] have been considering the customer query response time for their service selection problem. Assuming L(Sij ) is the latency difference between customer request time and service provider response time and ET (Sij ) is the execution time of service i in cloud j. When a customer accepts a request from the service provider, the service provider must spend ET (Sij ) time implementing the service request. Thus, the goal is to minimize request response time (RT). The wording is given N in the next equation : RTglobal = i=1 L(Sij ) + ET (Sij ).

A Survey on the QoS and MO Metaheuristics for IoT

2.7

415

Number of Clouds Involved in the Composition

The number of clouds included in the service composition (θ), including several services from a large number of clouds, increases the inter-cloud communication, which raises, in turn, the energy consumption and communication costs. As mentioned in [2] is calculated following |compositon|



θ :=

(α · Cloud(Si ))

(α : boolean; 0 if Cloud already added, else 1).

i=1

2.8

Number of Clouds Involved Composition Plan

The number of cloud composition plans involved in the composition (γ),has been considered as an effective QoS measure in [2]. Where each cloud has its own predefined composition plan list that is already developed and optimized, inciting composition to use a minimal number of cloud composition plans enhances the response time and lowers the costs. The calculation of this QoS measure is explained by |compositon|

γ :=



(α · C plan(Si ))

(α : boolean; 0 if C plan already added, else 1).

i=1

3

Multiobjective Metaheuristics

The multiobjective optimization with metaheuristics algorithms tested in this project are the most used and suitable for the MO IoT service selection problem. The algorithms are briefly described in the following sections. Multiobjective optimization’s primary goal is to identify a set of solutions that are as close to Pareto Front as possible. A solution here is a composition of IoT services from multiple service providers which satisfies the client request. On the other hand, the second goal is the solution set must be diverse and dispersed along the entire optimal Pareto Front, in order to portray an accurate estimation. 3.1

Indicator-Based Evolutionary Algorithm

IBEA [35] is an indicator-based metaheuristic. This evolutionary algorithm’s fitness assignment scheme is based on a pairwise comparison of solutions in a population using a binary quality indicator. A binary tournament among randomly chosen individuals is used as a reproduction selection strategy. The replacement technique involves deleting the worst individuals one by one and adjusting the fitness values of the remaining solutions at each suppression. This process is repeated until the required population size is met. In terms of convergence, IBEA surpasses NSGAII. The computational Cost of the quality indicator value is IBEA’s major weakness.

416

3.2

A. Zebouchi and Y. Aklouf

Parallel Speed-Constrained Multiobjective Particle Swarm Optimization

pSMPSO [17] is a parallelized version of SMPSO which is the finest algorithm in the context of the problems due to the velocity constriction mechanism. The technique uses a standard dominance method that can be enhanced to generate a population that converges more quickly in a multiobjective problem [15]. In [27], the authors developed a parallel version of the algorithm using the masterslave paradigm to parallelize the SMPSO evaluation operator. The algorithm evaluates all of the solutions right after they have been created, and it does so in parallel. The master process, which is implemented on the main execution thread, assigns the solutions to be evaluated to each thread. The master process then waits until all slave threads have completed their tasks and reported the fitness values. 3.3

dual MO Particle Swarm Optimization Algorithm

dMOPSO is commonly used to solve MO problems. The usage of the NSGAII crowding distance to filter out leader solutions is one of its primary characteristics. The algorithm is the most effective method in terms of the quality of the approximations to the Pareto front found. On the other hand, decomposition approaches may worsen the algorithm’s performance, especially when working with increasingly complicated situations [16]. 3.4

Multiobjective Cellular Genetic Algorithm

MOCell is a multiobjective optimization technique based on cellular genetics. The non-dominated individuals found throughout the search are saved in an external archive by the algorithm. When it comes to problems with more than two objectives, it’s a highly competitive method in terms of convergence and hypervolume measures. The feedback of individuals from the archive to the population is the most notable characteristic of MOCell in comparison to other existing cellular techniques [20]. However, it is possible to find a variant of the algorithm that outperforms MOCell. Taking into account synchrony, archive feedback, and replacement, Nebro et al. [19] offered several alternative versions: Synchronous algorithms can be more effective in terms of hit rate [22]. – sMOCell1; It is the basic synchronous MOCell algorithm. – sMOCell2; It consists of the original MOCell with archive feedback through parent selection. An asynchronous version can be obtained by updating the cells in sequential order, using a unique population. Asynchronous algorithms can be more efficient (faster) than synchronous ones. – aMOCell1; Is the asynchronous version of MOCell

A Survey on the QoS and MO Metaheuristics for IoT

417

– aMOCell2; It is a configuration of aMOCell1 where the worst neighbour of each generation is replaced. – aMOCell3; it is a configuration of asynchronous MOCell, archive feedback through parent selection. 3.5

Non-dominated Sorting Genetic Algorithm II

NSGAII proposed in [5], is a fast and elitist multiobjective genetic algorithm that can keep a wider variety of solutions and converge more quickly in the nondominated front. The present population is sorted by the number of solutions that dominate each solution in NSGAII. As a result, a series of non-dominated fronts with individuals of the same rank emerge. The algorithm then sorts each of these fronts based on the distance between consecutive solutions, favouring solutions in lightly populated regions of the search space before adding them into the population of the following iteration. The algorithm uses selection, crossover, and mutation to create a child population, then combined with the parent population before the next iteration begins. It is noticeable that NSGAII has problems with convergence closer to Pareto to the right front in some scenarios. 3.6

Parallel Non-dominated Sorting Genetic Algorithm II

Many approaches proposed to parallelize the NSGAII algorithm [6,12,23], the most used and applicable, is the master-slave approach, which is the calculation of the objective formulate of each individual in the generation in parallel. The master thread executes the NSGAII core code, and the agents evaluate the objective functions of each new individual. Random selection of operators provides overall better results than the adaptive version in bi-objective problems, while the latter outperforms the original NSGAII in three-Objective problems. 3.7

Non-dominated Sorting Genetic Algorithm II Steady State

A steady-state GA is an alternative to a generational GA, in which there is only one population. In this manner, new individuals are instantly absorbed into the evolutionary cycle, which allows the parents and offspring to dwell in the same community. A steady-state variant of the NSGAII Durillo et al. proposed in [7] converges better to the optimal Pareto front and spread, and it can be simply implemented using a one-generation offspring population. The ranking and crowding algorithms must be applied each time a new individual is created, significantly increasing the algorithm’s computing complexity. 3.8

Non-dominated Sorting Genetic Algorithm II Random

NSGAIIr [21] is an extension of NSGAII that employs three different variation operators: SBX, polynomial mutation, and DE’s variation operator. When a new solution product is introduced, these operators are chosen at random. The key

418

A. Zebouchi and Y. Aklouf

difference from the original NSGAII is the selection method used by the parents, and the way offspring are produced. One of the three variation operators is chosen based on this value. The algorithm then operates like the original NSGAII once the offspring has been formed. 3.9

Adaptative Non-dominated Sorting Genetic Algorithm II

Adaptative NSGAII (ANSGAII) [21] uses the variation operators in the same manner that NSGAII Random does, but a more adaptive approach that considers their contribution. In other words, each operator selection probability is altered based on the operator’s success in the previous iteration. The adaptive strategy for operator selection that is considered is based on the one used in AMALGAM. Because it incorporates an adaptive mechanism, ANSGAII outperforms the original algorithms in various experiments. ANSGAII can find solutions for each occurrence that all fall within the first non-dominated front. 3.10

Multiobjective Evolutionary Algorithm

MOEAD [13] is an approach that combines Evolutionary Algorithms and the traditional mathematical Decomposition method. This hybridization makes applying the single objective optimizer to each sub-problem associated with a solution, resulting in a more dispersed solution. The neighbourhood structure introduced in MOEAD allows each sub-problem to be optimized using only the information from its neighbours, greatly speeding up algorithm convergence. However, when dealing with scaled problems, such as scaling challenges, its fundamental flaw is the decrease of diversity and solution distribution. 3.11

Parallel Multiobjective Evolutionary Algorithm

The parallel version of MOEAD is proposed in [18], the authors chose the strategy of distributing the population across multiple concurrent threads. This way, each thread might work on a different segment of the population at the same time. In most of the examined cases, the results show that there are no significant differences between the fronts calculated by MOEAD and pMOEAD in terms of quality metrics. 3.12

Optimized Multi-Objective Particle Swarm Optimization

Optimized MOPSO [29] is a multiobjective particle swarm optimization technique. Its primary characteristics include using an external archive based on the NSGAII crowding distance to filter out leader solutions. In a binary tournament, the leaders are picked based on their crowding value, and only those with the highest levels of overcrowding are maintained. According to the findings in [30], OMOPSO outperforms the other MO algorithms. The idea of dominance is used to limit the number of solutions saved in this archive.

A Survey on the QoS and MO Metaheuristics for IoT

3.13

419

Multiobjective Random Search

Multiobjective random search (MORS) [8] random search is a single solution search approach. The solutions are generated by randomly adding, removing, or shifting one element of the solution to the preceding solution. For non-dominance evaluation, the new solution is compared to the current Pareto optimal solutions, and the Pareto optimal solution set is updated accordingly. The algorithm works better for high dimensional, highly constrained nonlinear optimization problems; it is challenging to outperform other MO algorithms when using it for manyobjective problems.

4 4.1

Experimental Evaluation Dataset

As we could not find workloads fitting our purposes in the literature, we created a generator of random multi-cloud environment’s configurations. We provide the generator parameters, including the number of CP and the number of services in the Multi-cloud environment. The script generates an XML file for each CP, having randomly generated values, including QoS of each Service, Number of composition plans, Number of available services in CP, Number of predefined composition plans for every cloud, selected services in a CP, and selected services from available services to be included in a composition plan of the CP. 4.2

Results

To effectively evaluate and compare the performance of each algorithm on our dataset. All algorithms were run 50 times on the generated dataset of 100 cloud providers and 1000 services for eight objective function evaluations, assuring us that each algorithm performs the same amount of work. All experiments are conducted on the same machine with Intel Core i7 2.60 GHz. Processor, 8 GB RAM, and Windows 10. The comparison of the algorithm performances is made based on their quality metrics and execution time. The used metrics for the evaluation of the convergence and diversity of the Pareto solution set are: EPSILON, HV, and the Spread, such as the algorithms with a smaller value of EPSILON, smaller value spread, and higher value of HV are the better-computed fronts. Table 1 shows the normalized EPSILON, HV and Spread each have its Mean Standard deviation and Median IQR of the obtained values from the runs of the 18 Multiobjective algorithms considering the 8 QoS.

420

A. Zebouchi and Y. Aklouf Table 1. Evaluation results. NSGAII

MOCell

aMOCell1 aMOCell2 aMOCell3 sMOCell1 sMOCell2 pNSGAII

ANSGAII

EPSILON.MeanAndStandardDeviation 4,37E+03 6,45E+03 6,45E+03 5,81E+03 6,22E+03 7,27E+03 6,19E+03 2,87E+03 6,33E+03 EPSILON.MedianAndIQR

4,37E+03 6,45E+03 6,45E+03 5,81E+03 6,22E+03 7,27E+03 6,19E+03 2,87E+03 6,33E+03

HV.MeanAndStandardDeviation

2,98E+05 3,50E+06 4,64E+05 2,39E+05 1,11E+05 4,34E+06 2,07E+05 3,00E+05 8,21E+05

HV.MedianAndIQR

2,98E+05 3,50E+06 4,64E+05 2,39E+05 1,11E+05 4,34E+06 2,07E+05 3,00E+05 8,21E+05

SPREAD.MeanAndStandardDeviation 3,72E+03 4,68E+03 4,90E+03 4,17E+03 4,04E+03 4,09E+03 3,48E+03 3,51E+03 3,92E+03 SPREAD.MedianAndIQR

3,72E+03 4,68E+03 4,90E+03 4,17E+03 4,04E+03 4,09E+03 3,48E+03 3,51E+03 3,92E+03 NSGAIIr NSGAIIss dMOPSO pSMPSO

OMOPSO MOEAD

pMOEAD IBEA

EPSILON.MeanAndStandardDeviation 4,82E+03 2,65E+03 4,36E+03 4,54E+03 5,76E+03 8,80E+03 6,95E+03 5,17E+03

MORS 7,73E+03

EPSILON.MedianAndIQR

4,82E+03 2,65E+03 4,36E+03 4,54E+03 5,76E+03 8,80E+03 6,95E+03 5,17E+03

7,73E+03

HV.MeanAndStandardDeviation

7,72E+05 5,49E+05 6,29E+05 3,79E+05 3,08E+05 5,47E+05 4,91E+05 6,03E+05

1,68E+05

HV.MedianAndIQR

7,72E+05 5,49E+05 6,29E+05 3,79E+05 3,08E+05 5,47E+05 4,91E+05 6,03E+05

1,68E+05

SPREAD.MeanAndStandardDeviation 3,81E+03 3,91E+03 1,37E+02 4,60E+03 3,97E+03 6,96E+03 1,24E+02 5,23E+03

3,92E+03

SPREAD.MedianAndIQR

3,92E+03

3,81E+03 3,91E+03 1,37E+02 4,60E+03 3,97E+03 6,96E+03 1,24E+02 5,23E+03

Fig. 1. Multiobjective metaheuristics ranks comparison.

Fig. 2. Multiobjective metaheuristics execution time comparison.

For each metric, the mean is calculated from both the Mean Std and Median IQR of that metric, which have very similar values, the algorithms are then sorted according to each metric from the worst to the best algorithm, the higher the metric, the better the algorithm. The sum slope represents the sum of the ranks of each metric rank. The higher mean value of HV and the lower mean value of SPREAD show a good approximation to the algorithm’s Pareto Front. As shown in Fig. 1 and Table 1 pNSGAII has a higher level of 36 which suggests that pNSGAII produce the best computed non-dominated front. Figure 2, shows the means execution time for each algorithm; it shows that OMOPSO performed the worst execution time, while pNSGAII overpassed pMOEAD by slightly two seconds which makes pSNGAII having the best execution time among all the tested algorithms.

5

Discussion

We can deduce several facts from the previous results. The outcomes of MOCell and its derivatives were inconsistent, and it did not outperform many other algorithms. sMOCell only brings a minor improvement to the solution quality in comparison with other synchronous algorithms. The asynchronous versions

A Survey on the QoS and MO Metaheuristics for IoT

421

aMOCell 1 and 2, performed better in terms of execution time. In contrast to aMOCell3, which offered a better Pareto front and increased execution time a bit compared to the original algorithm. We remark that OMOPSO algorithm had the slowest execution time, four times that of the second MO algorithm. Still, it showed potential in producing high-quality distributions. MOEAD and pMOEAD had the best execution time; dMOPSO, IBEA, PSMPSO, and MORS all produce non-satisfying results in terms of execution time, and did not perform better solution quality. The pNSGAII variants outperformed all other MO algorithms in terms of metrics and execution time, the algorithm provided the best Pareto front quality and improved execution time over the original algorithm. The number of objectives made it more challenging for other algorithms to meet the pNSGAII quality metrics in a reasonable amount of time thanks to pNSGAII low computational needs. Mutation operators and appropriate population distribution were two important reasons for this superiority since they prevented the algorithm from becoming trapped in local optimum. The crowding distance operator for pNSGAII played great during the selection and the population diversity between generations was well conserved. Therefore, the pNSGAII is the most suitable for improving multiobjective multi-cloud service selection with additional clouds and objectives.

6

Conclusion

This paper has collected and presented the most targeted QoS for multi-cloud Computing applied to IoT. Then, we proposed a survey on the most important Multiobjective optimization methods that are based on metaheuristics. We also implemented and compared their performances based on four well-known metrics. Thus, we found out that the pNSGAII algorithm outperforms other approaches and that it is the most adequate for the multi-cloud IoT service selection issue. We hope that the findings in this paper will spark some thoughts for new research works that offer more efficient methods in this field. In future, we will extend this study to include other MO Algorithms such as games theory and Deep Reinforcement Learning based approaches. Additionally, we plan to work on the amelioration of the best Metaheuristic found, namely pNSGAII.

References 1. Asghari, P., Rahmani, A.M., Javadi, H.H.S.: Privacy-aware cloud service composition based on QoS optimization in internet of things. J. Amb. Intell. Human. Comput. (2020). https://doi.org/10.1007/s12652-020-01723-7 2. Baker, T., Asim, M., Tawfik, H., Aldawsari, B., Buyya, R.: An energy-aware service composition algorithm for multiple cloud-based IoT applications. J. Netw. Comput. App. 89, 96–108 (2017). https://doi.org/10.1016/j.jnca.2017.03.008

422

A. Zebouchi and Y. Aklouf

3. Chauhan, S.S., Pilli, E.S., Joshi, R., Singh, G., Govil, M.: Brokering in interconnected cloud computing environments: a survey. J. Parallel Distrib. Comput. 133, 193–209 (2019). https://doi.org/10.1016/j.jpdc.2018.08.001 4. Choudhary, G., Jain, A.: Internet of things: a survey on architecture, technologies, protocols and challenges. In: 2016 International Conference on Recent Advances and Innovations in Engineering (ICRAIE). IEEE (2016). https://doi.org/10.1109/ icraie.2016.7939537 5. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multi-objective genetic algorithm: NSGA-II. IEEE Trans. Evol. Comput. 6(2), 182–197 (2002). https://doi.org/10.1109/4235.996017 6. Durillo, J.J., Nebro, A.J., Luna, F., Alba, E.: A study of master-slave approaches to parallelize NSGA-II. In: 2008 IEEE International Symposium on Parallel and Distributed Processing. IEEE (2008). https://doi.org/10.1109/ipdps.2008.4536375 7. Durillo, J.J., Nebro, A.J., Luna, F., Alba, E.: On the effect of the steady-state selection scheme in multi-objective genetic algorithms. In: Ehrgott, M., Fonseca, C.M., Gandibleux, X., Hao, J.K., Sevaux, M. (eds.) Evolutionary Multi-criterion Optimization. EMO 2009. Lecture Notes in Computer Science. LNCS, vol. 5467, pp. 183–197. Springer, Berlin, Heidelberg (2009). https://doi.org/10.1007/978-3642-01020-0 18 8. Feng, J., Shen, W.Z., Xu, C.: Multi-objective random search algorithm for simultaneously optimizing wind farm layout and number of turbines. J. Phys. Conf. Ser. 753, 032011 (2016). https://doi.org/10.1088/1742-6596/753/3/032011 9. Hatton, M.: The IoT in 2030: 24 billion connected things generating $1.5 trillion (2020). https://alhena.io/the-iot-in-2030-24-billion-connected-things-generating1-5-trillion/. Accessed 23 Sep 2021 10. Kumrai, T., Ota, K., Dong, M., Kishigami, J., Sung, D.K.: Multi-objective optimization in cloud brokering systems for connected internet of things. IEEE Internet Things J. 4(2), 404–413 (2017). https://doi.org/10.1109/jiot.2016.2565562 11. Lakhdari, A., Bouguettaya, A., Mistry, S., Neiat, A.G.G.: Composing energy services in a crowdsourced IoT environment. IEEE Trans. Serv. Comput. 99, 1 (2020). https://doi.org/10.1109/tsc.2020.2980258 12. Lancinskas, A., Zilinskas, J.: Approaches to parallelize pareto ranking in NSGAII algorithm. In: Wyrzykowski, R., Dongarra, J., Karczewski, K., Wasniewski, J. (eds.) Parallel Processing and Applied Mathematics. PPAM 2011. Lecture Notes in Computer Science, vol. 7204, pp. 371–380. Springer, Heidelberg (2012). https:// doi.org/10.1007/978-3-642-31500-8 38 13. Li, H., Zhang, Q.: Multi-objective optimization problems with complicated pareto sets, MOEA/d and NSGA-II. IEEE Trans. Evol. Comput. 13(2), 284–302 (2009). https://doi.org/10.1109/tevc.2008.925798 14. Liu, J., et al.: A cooperative evolution for QoS-driven IoT service composition. Automatika 54(4), 438–447 (2013). https://doi.org/10.7305/automatika.54-4.417 15. Maltese, J., Ombuki-Berman, B.M., Engelbrecht, A.P.: A scalability study of many-objective optimization algorithms. IEEE Trans. Evol. Comput. 22(1), 79–96 (2018). https://doi.org/10.1109/tevc.2016.2639360 16. Mart´ınez, S.Z., Coello, C.A.C.: A multi-objective particle swarm optimizer based on decomposition. In: Proceedings of the 13th Annual Conference on Genetic and Evolutionary Computation - GECCO 2011. ACM Press (2011). https://doi.org/ 10.1145/2001576.2001587 17. Nebro, A.J., Durillo, J., Garc´ıa-Nieto, J., Coello, C., Luna, F., Alba, E.: SMPSO: a new PSO metaheuristic for multi-objective optimization (2009)

A Survey on the QoS and MO Metaheuristics for IoT

423

18. Nebro, A.J., Durillo, J.J.: A study of the parallelization of the multi-objective metaheuristic MOEA/D. In: Blum, C., Battiti, R. (eds.) Learning and Intelligent Optimization. LION 2010. Lecture Notes in Computer Science. LNCS, vol. 6073, pp. 303–317. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-138003 32 19. Nebro, A.J., Durillo, J.J., Luna, F., Dorronsoro, B., Alba, E.: Design issues in a multi-objective cellular genetic algorithm. In: Obayashi, S., Deb, K., Poloni, C., Hiroyasu, T., Murata, T. (eds.) Evolutionary Multi-Criterion Optimization. EMO 2007. LNCS, vol. 4403, pp. 126–140. Springer, Heidelberg (2007). https://doi.org/ 10.1007/978-3-540-70928-2 13 20. Nebro, A.J., Durillo, J.J., Luna, F., Dorronsoro, B., Alba, E.: MOCell: a cellular genetic algorithm for multi-objective optimization. Int. J. Intell. Syst. 24(7), 726– 746 (2009). https://doi.org/10.1002/int.20358 21. Nebro, A.J., Durillo, J.J., Machin, M., Coello Coello, C.A., Dorronsoro, B.: A study of the combination of variation operators in the NSGA-II Algorithm. In: Advances in Artificial Intelligence. CAEPIA 2013. LNCS, vol. 8109, pp. 269–278. Springer, Heidelberg (2013). https://doi.org/10.1007/978-3-642-40643-0 28 22. Olariu, S., Zomaya, A.Y. (eds.): Handbook of Bioinspired Algorithms and Applications. Chapman and Hall/CRC, Boco Raton (2005). https://doi.org/10.1201/ 9781420035063 23. de Oliveira, L.B., Marcelino, C.G., Milanes, A., Almeida, P.E.M., Carvalho, L.M.: A successful parallel implementation of NSGA-II on GPU for the energy dispatch problem on hydroelectric power plants. In: 2016 IEEE Congress on Evolutionary Computation (CEC). IEEE, July 2016. https://doi.org/10.1109/cec.2016.7744337 24. Pang, B., Hao, F., Park, D.S., Maio, C.D.: A multi-criteria multi-cloud service composition in mobile edge computing. Sustainability 12(18), 7661 (2020). https:// doi.org/10.3390/su12187661 25. Singh, M., Baranwal, G., Tripathi, A.K.: QoS-aware selection of IoT-based service. Arabian J. Sci. Eng. 45(12), 10033–10050 (2020). https://doi.org/10.1007/s13369020-04601-8 26. Sun, M., Zhou, Z., Wang, J., Du, C., Gaaloul, W.: Energy-efficient IoT service composition for concurrent timed applications. Future Gener. Comput. Syst. 100, 1017–1030 (2019). https://doi.org/10.1016/j.future.2019.05.070 27. Toutouh, J., Alba, E.: Parallel multi-objective metaheuristics for smart communications in vehicular networks. Soft Comput. 21(8), 1949–1961 (2015). https://doi. org/10.1007/s00500-015-1891-2 28. Vakili, M., Jahangiri, N., Sharifi, M.: Cloud service selection using cloud service brokers: approaches and challenges. Front. Comput. Sci. 13(3), 599–617 (2018). https://doi.org/10.1007/s11704-017-6124-7 29. Wang, H., Qian, F.: Improved PSO-based multi-objective optimization using inertia weight and acceleration coefficients dynamic changing, crowding and mutation. In: 2008 7th World Congress on Intelligent Control and Automation. IEEE (2008). https://doi.org/10.1109/wcica.2008.4593644 30. Wang, H., Qian, F.: Improved PSO-based multi-objective optimization using inertia weight and acceleration coefficients dynamic changing, crowding and mutation. In: 2008 7th World Congress on Intelligent Control and Automation, pp. 4479–4484 (2008). https://doi.org/10.1109/WCICA.2008.4593644 31. Wang, W., Niu, D., Li, B., Liang, B.: Dynamic cloud resource reservation via cloud brokerage. In: 2013 IEEE 33rd International Conference on Distributed Computing Systems. IEEE, July 2013. https://doi.org/10.1109/icdcs.2013.20

424

A. Zebouchi and Y. Aklouf

32. Yang, C., Peng, T., Lan, S., Shen, W., Wang, L.: Towards IoT-enabled dynamic service optimal selection in multiple manufacturing clouds. J. Manuf. Syst. 56, 213–226 (2020). https://doi.org/10.1016/j.jmsy.2020.06.004 33. Zhang, M., Liu, L., Liu, S.: Genetic algorithm based QoS-aware service composition in multi-cloud. In: 2015 IEEE Conference on Collaboration and Internet Computing (CIC). IEEE, October 2015. https://doi.org/10.1109/cic.2015.23 34. Zhang, X., Geng, J., Ma, J., Liu, H., Niu, S.: A QoS-driven service selection optimization algorithm for internet of things, September 2020. https://doi.org/ 10.21203/rs.3.rs-69961/v1 35. Zitzler, E., K¨ unzli, S.: Indicator-based selection in multi-objective search. In: Yao, X., et al. (eds.) Parallel Problem Solving from Nature - PPSN VIII. PPSN 2004. LNCS, vol. 3242, pp. 832–842. Springer, Heidelberg (2004). https://doi.org/10. 1007/978-3-540-30217-9 84

Critical Success Factors for Information Technology and Operational Technology Convergence Within the Energy Sector Thabani Dhlamini and Tendani Mawela(B) University of Pretoria, Hatfield, South Africa [email protected], [email protected]

Abstract. Smart grids involve the incorporation of ICTs within electrical power systems. Smart grids are touted to offer a variety of opportunities for public electricity utilities. Governments and electricity utilities around the world are taking advantage of various technology advancements. This has given rise to complex and integrated electrical power systems. With these technology advancements, a new breed of devices, technologies, smart grids and systems allow the electricity utilities to change the way their electricity is delivered, maintained and managed for high reliability. Smart grids have introduced new infrastructure into the grid that integrates Information Technology (IT) and Operational Technology (OT) and other support systems to maximize business value. Traditionally IT and OT are managed as independent network systems and are subsequently supported separately. This study explored the convergence of the IT and OT functions for smart grid implementations in support of a business transformation strategy. It investigated the enablers and barriers towards IT and OT convergence in the energy sector. The study adopted a case study approach and collected data through a survey with the IT and OT representatives within a public electricity utility. The study offers several critical success factors for consideration by IT and OT functions within organizations for smart grid implementation. Keywords: Smart grid · Information technology · Operational technology · Convergence · Government · Electricity utility

1 Introduction ICTs can influence business transformation through its impact on human resources, organizational structure, processes, strategy, services, products and how things are done [1]. Technology is an enabler for transformation across industries supporting organizational growth, cost reduction and gaining a competitive advantage. Electricity utilities worldwide are taking advantage of technology advancements in communication, computing and power systems algorithms that have led to continuous growth and interconnectivity. Smart grids are touted to offer various opportunities for electricity utilities [2]. Technology can be classified as either Operational Technology or Information Technology [3]. Information Technology (IT) and Operational Technology (OT) both support business © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 425–434, 2022. https://doi.org/10.1007/978-3-030-96299-9_41

426

T. Dhlamini and T. Mawela

operations and processes through the provision of technology, as a partner to businesses that range from simple to complex solutions [3]. OT is defined as “hardware and software that detects changes, monitors or controls assets or processes; such technologies generate real-time asset data with embedded software” [4]. IT on the other hand, provides the back-office information systems related to revenue, billing, accounting, workforce management records and other related information systems services to enable organizations to function efficiently [5]. Electricity utilities are increasingly facing financial pressure and are seeking ways to balance the demand and supply of electricity. In doing so, they have employed a complementary approach by investing in technologies that will lead to smart generation of electricity in the power grids [6]. Technology advancements have resulted in a new breed of devices, technologies, and systems that allow the electricity utilities to change how their electricity is delivered, maintained and managed for high reliability [7]. The technology-enabled power grids are termed smart grids and can benefit electricity utilities with the demand side management, efficiency, improved plant availability and reliability, amongst other key metrics [6]. Smarts grids are made up of networks, which include IT and OT machines sharing data between them [8]. It is also said that a smart power grid ensures grid reliability by coordinating multiple efforts from different role players and multiple organizations, to balance efforts from grid operators and transmission operators in a complex interconnected system [9]. The premise of the smart grid is on the intelligent network that can be monitored and controlled in real-time [10]. The convergence of IT and OT can be considered to support smart grids as a part of an IT-enabled business transformation strategy within the energy sector. This paper focused on understanding the opportunities for IT and OT convergence in one public energy utility to draw lessons and critical success factors for smart grid implementation towards integrated electric power systems. The need for further studies to understand IT and OT convergence better has been previously emphasized by scholars [11]. This study investigated the IT and OT convergence within the energy sector, which requires collaborative efforts between IT and OT departments. Its premise is on the case study of a public utility that has adopted a strategy of implementing smarts grids to improve operations and deliver reliable energy to its customers through its energy distribution network.

2 Informing Literature 2.1 Information Technology and Operational Technology in Utilities Information technology is made up of information systems that have various types of technology that can run, process, store and transmit data through network connectivity using computers and other equipment as means of communication [12]. IT enables an organization to be creative and to solve business problems by using new knowledge that is stored and retrieved in the IT systems [13]. However, with regard to security functions, IT is concerned with the confidentiality of data, how data is stored in a secured manner, and the integrity of such data. IT systems must ensure the availability of the stored and processed data through information management systems [14]. The IT function in an

Critical Success Factors for Information Technology

427

organization mainly maintains the corporate network and the flow of information in those networks. Operational Technology is defined as the activities and operations of the electric grid in real-time to ensure high reliability [15]. The electric grid is maintained by field workers who are technical resources in the OT space, and they work through control centres to open and close circuits in real-time observing safety measures. OT components have a mix of hardware and software responsible for changes in physical processes, which might detect or cause changes to devices in the network systems to deliver the aimed responses and functions into the physical world, with an intention to ensure constant availability and reliability [14]. The OT system is hence made up of software and hardware components similar to what is found in the IT environment. The difference is how these components influence or change physical processes, which are devices in the OT network. The intention is to detect any physical failures or performance issues of the devices to ensure reliability and availability; hence physical changes can be made on the OT network to change course and avoid failures in the network. It is noted that the OT environment has management and supervisory systems that depend on IT for their smooth functioning [16]. The assets that collect data in OT are not classified as IT infrastructure assets due to their specialized nature [17]. 2.2 IT and OT Convergence Opportunities and Challenges Smart grid implementations may be enhanced through better integration, collaboration and convergence of the IT and OT functions within organizations. The convergence of IT and OT is driven by a changing technology landscape that is increasingly data- driven with remote operational tools operating in the two environments. Traditionally OT had a physical separation from IT, and they were relatively isolated and secure as standalone entities [18]. The trend in many different industries indicates an important need for IT and OT convergence. The literature confirms the need for shared goals in the two environments and emphasizes IT and OT convergence as it benefits the organization to optimize business processes, which leads to better decision making due to the availability of timely information. This will then result in improved productivity and high system availability and reliability [19]. It has been highlighted that the key challenge which leads to disjointedness in OT and IT convergence is the difference in priorities that exists in these two environments [18]. Furthermore, it is noted that there are no priorities that are more important than the other because they essentially differ due to the nature of operations in these two environments [18]. The challenges of OT and IT integration are due to the fact that different departments have to manage applications in their respective departments within the same organization [20]. Hence, it is not easy to integrate the systems. Previous studies highlight the challenges of governance and security with the integration of IT and OT and the flow of data from engineering asset management to corporate information systems [4]. Furthermore, the challenges can be categorized into people, processes and technical, where people challenges will typically include roles and responsibilities [4]. Process challenges include governance standards, enterprise-level architecture, standardized platform, open data and communication standards. Technical challenges are

428

T. Dhlamini and T. Mawela

identified as the consolidation of hardware, network, application, information and data tiers when integrating OT and IT [4]. It is also highlighted that the blending of IT and OT systems enable industrial applications to keep up with the level of growing network complexity and sub-networks interconnections [21]. OT and IT differ immensely in their approach, requirements and objectives, and as such, they are designed to run independently in many organizations. They are connected through data sharing, where OT data are transferred to IT for further processing, leading to some analysis and reporting via business applications [22]. The main issue with IT and OT convergence is the different priorities that hinder progress in the IT and OT convergence space where OT’s main goal is on operations while IT is business oriented [22]. The literature also argues that there are opportunities for OT and IT to collaborate better [23]. 2.3 IT and OT Convergence to Support Smart Grids Smart grids are power system grids in the electricity utility that are enabled by information manipulation on factors of the network [16, 17]. The extant literature argues that smarts grids can better manage the power grid by using the information on the predictive maintenance to identify asset failures, asset design failures and maintenance scheduling [16, 17]. For smart grids to work, there is a need for an exchange of data and information between OT and IT; therefore, it would not be possible without the convergence of IT and OT. Smart grids require components from OT, which are sensors, controllers, other devices and assets in the OT environment. It needs the IT network with the internet to transmit and process information. With IT and OT convergence, the operational environment is thus connected and accessible to other parties that previously connected only to the IT environment through the internet and cloud technologies [18]. In the context of smart grids, electricity utilities adopt technology solutions that enable the convergence of IT and OT, such as automation, system protection and other control capabilities, which are real-time with increased reliability [24]. The literature points to the demand for increased reliability, which forces the electricity utility to upgrade their electric grid technology to ensure the practicality of digitization, automation and for these to be possible, data transfer between IT and OT becomes important and a necessity [24]. With the high connectivity required because of greater accessibility, a variety of vulnerabilities have emerged in the IT and OT convergence, which must be addressed [24]. The convergence of IT and OT will be accelerated by introducing smart machines that will see production data being used in various management activities, including energy management, stock control and product replacements [25]. This is underpinned by a common language and standards as a basic requirement to ensure interoperability in the two environments [25]. A smart-grid is a form of a smart machine. It is argued that smart grid technologies will introduce efficiencies in the business with the improved availability in many elements of the power grid, such as plant availability and key metrics that allow for efficiency in the demand side management [6]. Research has further elaborated on services offered by a smart enabled grid in terms of system availability for long-term operations, maintenance, minimization of power disruption, disturbances, security and safety related to quick turn-around times for diagnosis and problem solving [6].

Critical Success Factors for Information Technology

429

3 Research Methods 3.1 Case Background This study is based on a single case study of an electricity utility organization. The organization provides electricity to different categories of customers including residential customers, industrial customers and municipalities. The organization operates 30 power stations with a generation capacity of 45 561 MW with a network of 38 1594 kms of low, medium and high voltage lines. It has adopted a turnaround plan with five focus areas, namely: debt relief, revenue management, business separation, cost initiatives and operational stability. Part of the strategy includes smart grid expansion. The Information Technology department has a Smart Metering, Smart Grid and Smart Electricity strategy. This strategy aims to optimize the business (grid operations, asset management, customer service, future energy products), achieve grid resilience and sustainability as well as enable business model evolution. 3.2 Research Approach The study was exploratory, followed a quantitative approach and adopted an online selfadministered survey questionnaire. Through purposeful sampling, the data was collected from practitioners in the IT and OT departments of the organization. A total of thirty respondents completed the online questionnaire. The sample included eighteen people from IT and twelve from the OT department. Additionally, two semi-structured interviews were held with managers from IT and OT to understand their perspectives on the convergence of IT and OT. The questionnaire and interview guides were reviewed by colleagues to ensure clarity of the questions. The interviews were also recorded and notes taken during the interview sessions. The data from the questionnaires was analysed via a descriptive approach and a thematic analysis was adopted for the textual interview data. Additionally, documents from the IT and OT departments, including strategy documents, annual reports, standards and operational procedures, were reviewed to further understand the context of the organization.

4 Findings and Discussion 4.1 IT Enabled Business Transformation It has been highlighted that corporate executives are facing a challenge with IT-enabled business transformation in that there should be an alignment between the business strategy, Information and Communication Technology (ICT) and organizational transformation [26]. This study investigated the IT-enabled business transformation through the implementation of a smart grid. The majority of the study respondents highlighted the need to reorganize the IT operation structurally to enable technology transformation. Also, two-thirds of respondents believed that the organization currently has the correct skills set to support the business transformation.

430

T. Dhlamini and T. Mawela

4.2 Smart Grid Strategy and Implementation This section indicates the IT and OT respondents’ understanding and views on the smart grid strategy of the organization. Over ninety percent of the respondents were aware of the organization’s smart grid strategy, fifty-seven percent supported the strategy, while thirty-three percent were neutral. However, ten percent did not agree with the strategy. Twenty-eight respondents, said there are some new opportunities and products that can be realized through smart grids. In addition, smart grids will enable business transformation in the organization. About ninety percent of the respondents acknowledge that overall, they believe the organization is ready to implement the smart grid strategy. Although the majority of respondents agreed, the IT senior manager said: “Although the smart grid strategy is there, there is no synergy between IT and OT to implement smart grids successfully, it is still siloed in the approach. There is a dependency on IT and OT to make it a success”. The respondents also highlighted that the implementation of smart grids will impact business processes and human resources, and it requires a change and alignment of the current organizational structure to support the implementation. In an interview, the OT senior manager indicated that the process can be impacted negatively, resulting in slow decision making when IT is not integrated with OT. The senior manager for IT said, “not much of an impact at the moment because IT and OT are still separate departments, but for smart grid strategy implementation, there will be a huge impact on business processes as IT will be required to collaborate with OT ”. In the literature [27] it is highlighted that: “Collaboration can occur in a variety of ways of sharing resources; for instance, sharing information, people, machinery, finance or even managerial expertise; The greater the degree of sharing between two independent units/entities then the closer are the two units/entities, in terms of operation, values and mental models…”. For this study, this collaboration is seen as a form of IT and OT convergence in order to facilitate and coordinate the implementation of digital and smart grids strategies. 4.3 Perceptions on IT and OT Convergence The study found that twelve IT staff and eight OT staff indicated that they understand the concept of IT/OT convergence. Another aspect of the assessment was to investigate whether there was a clear vision that is mapped out for IT and OT convergence in the organization. Just over half of the respondents (11 IT and 5 OT staff) agreed that there was a clear vision. Scholars argue that a clear vision and strategy of the future for an energy utility is critical to ensure that these types of organizations [10]: “empower, provide opportunities, and fulfil customer needs while remaining profitable, sustainable, and vital organizations”. In the interviews with OT and IT senior managers, both agreed that the smart grid strategy supports the vision and strategy of the organization. The respondents were also requested to reflect on how IT operates. The results indicated that 18 of the respondents agreed or strongly agreed that IT is currently an isolated function. Six respondents disagreed or strongly disagreed, and the remaining 6 respondents were neutral in their stance. In an interview with the senior manager for IT, he described IT as “siloed due to the maturity level in aligning IT and OT ”. Still, he continued to say IT services all business areas. In an interview with the OT senior

Critical Success Factors for Information Technology

431

manager he said, “IT works in isolation and should be working hand-in-hand with OT.” He also gave an example of the slow delivery of IT services. On the contrary to the above mentioned, 19 respondents believed that OT functions effectively when it is separate from IT. Half of the respondents indicated that they do not see a need for IT and OT to be integrated. Eighteen of the respondents indicated there are some cases they believed that IT and OT should exist independently. Also, there are still OT projects that are implemented independent of the IT department. It has been noted that an asset-intensive organization will have IT and OT that are disconnected [28]. Therefore, teams within this environment will have different skills in which they hardly communicate with each other, creating separate information silos. Interestingly sixty percent of the respondents either agreed or strongly agreed that the roles are clearly defined for IT and OT collaboration efforts. However, the respondents foresaw a significant change in the organizational culture to accelerate IT and OT convergence. The interview indicated that both OT and IT foresaw a major culture shift that will impact the organizational structure for successful IT and OT convergence. In addition to the above mentioned, training including cross-functional training of IT and OT to support convergence is needed. The senior manager for IT said, “there is going to be an impact on people and structure of the organization; re-training and re-skilling will be required; there are new opportunities for different types of jobs that don’t exist today”. The OT senior manager said “proper training on new technologies is needed and the challenge would be on older people that are computer illiterate, even with operating the network manual on ways of doing things would change as automation takes over”. Another aspect for this research to investigate was on the governance of projects, whether they align resources, skills and investments in the two departments of IT and OT. The majority of respondents agreed that this was in place. 4.4 Enablers, Barriers and Critical Success Factors for IT and OT Convergence A number of benefits were cited in the study regarding the move towards better collaboration of IT and OT in utilities. These are the automation of work processes, including automatic dispatching of resources, as well as creation of opportunities for self-healing technologies to reduce maintenance costs. In addition, there may be improved restoration times and the use of data analytics to optimize the network environment. The study found various enablers for IT and OT convergence, and these included: integrated IT operation into the business to enable technology transformation. Also, there is a need to ensure cross-training between IT and OT divisions. This should be supported by efforts to retrain or reskill staff members. Additionally, adjustments to the organizational structure are needed to support convergence and strategy implementation. These should be underpinned by a clear and shared strategy driven by metrics that track both the IT and OT space. The challenges and barriers towards IT and OT successful convergence are that currently, the IT and OT units operate in silos, with different practices and maintain parallel technology solutions and networks. Also, there are separate governance structures for IT and OT, which makes integration challenging. There is also resistance to change noted by some of the IT and OT staff members. Externally there is pressure and requirements related to government regulation which influence how the organization operates. There

432

T. Dhlamini and T. Mawela

are also potential job losses that may result from the convergence of IT and OT units, which is a concern for some members of the organization. However, it is noted that there is a need for IT and OT integration in utilities [31]. Risks related to cybersecurity are another bottleneck. The “ introduction of IT technologies in OT environments significantly increases the security risk and novel security frameworks need to be developed to meet industrial requirements” [21]. OT connectivity to the internet has brought about cyber security challenges that were only associated with the IT environment in the past [18]. Potential cyberattacks are a threat to devices that control, monitor and protect the power grid and should there be a successful invasion in the IT and OT environment, major power disruption can be realized and also result in damaged equipment. Hence, cybersecurity interventions with appropriate controls are required to protect the power grid [29]. The study highlights that organizations consider the following critical success factors that may support the integration and convergence of the IT and OT units. First, there is need for clarity in the roles of IT and OT. Where necessary, there may be a change to the roles to include OT and IT components in one role. Also, it is suggested that IT and OT units should report to one head or manager or alternatively, they should be consolidated into one department. Additionally, IT and OT support groups should be reporting to a Chief Information Officer (CIO) or Chief Technology Officer (CTO) [30], thereby merging IT and OT groups. Another consideration is the skills of the team members. There should be multi-skilled resources that work in both IT and OT.

5 Conclusion, Limitations and Future Research 5.1 Concluding Remarks Technology advancement has boosted enterprise IT with resource planning, customer relationship management, and decision support systems due to its role in managing the value chain, cost reduction and customer service improvement. The advancement of IT has been rapidly changing the ways organizations operate their businesses. As such, the implementation of the smart grid represents an opportunity for the business transformation of the organization. The convergence of IT and OT is required to enable this business transformation through emerging technology solutions such as smart grids. The study highlighted various enablers, barriers and critical success factors that electricity utilities may consider in advancing strategies related to smart grid implementation towards driving business growth. 5.2 Contribution, Limitations and Future Research This study investigated the convergence of IT and OT to enable business transformation, specifically in the energy sector. The study offered insights into the motivation and constraints behind IT and OT convergence. It focused on one organization as a case study, which is noted as a limitation. Thus, future studies may seek to solicit the views of additional entities and stakeholders in the energy sector on their experiences of driving IT and OT convergence in support of business transformation. Future research can also expand on the entities that provide energy to include other areas such as oil and gas.

Critical Success Factors for Information Technology

433

Acknowledgements. This work is based on the research supported in part by the National Research Foundation of South Africa (Grant Numbers 127495).

References 1. Brocke, J., Schmiedel, T., Simons, A., Schmid, A.M., Petry, M., Baeck, C.: From local IT needs to global process transformation: Hilti’s customer service program. Bus. Process. Manage. J. 22(3), 594–613 (2016) 2. Tuballa, M.L., Abundo, M.L.: A review of the development of smart grid technologies. Renew. Sustain. Energy Rev. 59, 710–725 (2016) 3. Lim, F.P.: A research analysis on the convergence of information and operational technologies in business. J. Next-gener. Converg. Inf. Serv. Technol. 5(1), 45–58 (2016) 4. Kuusk, A., Koronios, A., Gao, J.: Overcoming integration challenges in organizations with operational technology. Melbourne, ACIS (2013) 5. Saha, S.K.: Smart Grid Training Program Smart Grid Building Blocks. TATA Power Delhi Distribution Limited, Delhi (2011) 6. Dayabhai, S., Diamandis, P.: The role of virtualization in a smart-grid enabled substation automation system. Consolidated Power Projects (CONCO), Johannesburg (2015) 7. Sodha, N., Wadhwa, K., Tripathi, M.K., Jain, A.: Future Control Center with Advent of Smart grid. Power Grid Corporation of India Limited, 13–15 November (2013) 8. Vernotte, A., Välja, M., Korman, M., Björkman, G., Ekstedt, M., Lagerström, R.: Load balancing of renewable energy: a cyber security analysis. Energy Inform. 1(1), 1–41 (2018). https://doi.org/10.1186/s42162-018-0010-x 9. Le Blanc, K., Ashok, A., Franklin, L., Scholtz, J., Andersen, E., Cassiadoro, M.: Characterizing cyber tools for monitoring power grid systems: what information is available and who needs it? In: Canada, IEEE International Conference on Systems, Man, and Cybernetics (SMC) (2017) 10. Agüero, J.R., Khodaei, A., Masiello, R.: The utility and grid of the future challenges, needs, and trends. IEEE Power Energy Mag. (2016) 11. Kuusk, A., Gao, J.: Factors for successfully integrating operational and information technologies. In: Proceedings of PICMET 2015: Management of the Technology Age, Australia, pp. 1513–1523 (2015) 12. Amuna, Y.M.A., Shobaki, M.J.A., Naser, S.S.A.: The role of knowledge-based computerized management information systems in the administrative decision-making process. Int. J. Inf. Technol. Elect. Eng. 6(2), 1–9 (2017) 13. Olszak, C.M., Bartu´s, T., Lorek, P.: A comprehensive framework of information system design to provide organizational creativity support. Inf. Manage. 55, 94–108 (2018) 14. Theron, P.: Through-Life cyber resilience in future smart manufacturing environments. Res. Programme. Procedia Manuf. 16, 193–207 (2018) 15. Gauci, A., Michelin, S., Salles, M.: Addressing the challenge of cyber security maintenance through patch management. CIRED-Open Access Proc. J. 1, 2599–2601 (2017) 16. Haider, A.: Information and Operational Technologies Nexus: An Integrated Governance Model. IEEE (2011) 17. Haider, A.: IT enabled engineering asset management: a governance perspective. J. Organ. Knowl. Manage. 2011, 1–12 (2011). https://doi.org/10.5171/2011.348417 18. Murray, G., Johnstone, M.N., Valli, C.: The convergence of IT and OT in critical infrastructure. In Valli, C. (Ed.) The Proceedings of 15th Australian Information Security Management Conference, 5–6 December, 2017, Edith Cowan University, Perth, Western Australia, pp. 149– 155 (2017)

434

T. Dhlamini and T. Mawela

19. Lydon, B.: IT and OT convergence. International Society of Automation (2017). https://www. isa.org/intech-home/2017/may-june/columns/it-and-ot-convergence 20. Lu, Y., Ju, F.: Smart manufacturing systems based on cyber-physical manufacturing services (CPMS). IFAC PapersOnLine 50(1), 15883–15889 (2017) 21. Steiner, W., et al.: Next generation real-time networks based on IT technologies. In: IEEE 21st international Conference on Emerging Technologies and Factory Automation (ETFA), Austria, TTTech Computertechnik AG, pp. 1–8 (2016) 22. Ivankovi´c, I., Rubeša, R., Kekelj, A., Kuzle, I.: SCADA Maintenance and Refurbishment with Security Issue in Modern IT and OT Environment. Valletta, Innovation and Networks Executive Agency (2018) 23. Felser, M., Rentschler, M., Kleineberg, O.: Coexistence standardization of operation technology and information technology. Proc. IEEE 107(6), 962–976 (2019). https://doi.org/10. 1109/JPROC.2019.2901314 24. Pierpoint, L., Ledesma, R.: Cyber Threat and Vulnerability Analysis of the U.S. Electric Sector. Idaho National Laboratory, United States (2017) 25. Beudert, R., Juergensen, L., Weiland, J.: Understanding smart machines: how they will shape the future. Schneider-Electric, white paper (2015) 26. Margherita, A., Petti, C.: ICT-enabled and process-based change: an integrative roadmap. Bus. Process Manage. 16(3), 473–491 (2010) 27. Ahmed, P., Simintiras, A.: Conceptualizing business process re-engineering. Bus. Process Re-eng. Manage. J. 2(2), 73–92 (1996) 28. Lara, P., Sánchez, M., Villalobos, J.: OT modeling: the enterprise beyond IT. Bus. Inf. Syst. Eng. 61(4), 399–411 (2018). https://doi.org/10.1007/s12599-018-0543-3 29. Hawk, C., Kaushiva, A.: Cybersecurity and the smarter grid. Electr. J. 27(8), 84–95 (2014) 30. Gray, G.: Digital Transformation: Aligning Information Technology and Operations. Electric Power Research Institute, Palo Alto (2019) 31. Garimella, P.K.: IT-OT integration challenges in utilities. In: 2018 IEEE 3rd International Conference on Computing, Communication and Security (ICCCS) IEEE, pp. 199–204 (2018)

A New Structured Model for ICT Competencies Assessment Through Data Warehousing Software Vladimir Dobrynin1 , Michele Mastroianni2(B) , and Olga Sheveleva1 1 2

Dubna State University, Dubna, Moscow Oblast, Russia [email protected], [email protected] Universit´ a della Campania Luigi Vanvitelli, Caserta, Italy [email protected]

Abstract. The development of the digital economy is one of the priority areas for most countries, so the requirements for ICT specialists are changing and increasing, as a consequence of these transformation. In this perspective, the strengthening, measuring and assessment of digital competences are becoming crucial to ensure quality and security of the digital products and services implemented. This paper proposes MMACK, a new Meta-model for assessing ICT competencies and knowledge. This work also explores the feasibility of the implementation of such a model using Data Warehousing software. Keywords: Competence model OLAP · Hypercube

1

· Cybersecurity · Data Warehousing ·

Introduction

At the end of the 20th century and the beginning of the 21st century, the educational system is being modified in Europe and the United States. As a result, the National Skills Standard Board is being approved in the United States [23], and the Bologna Declaration on the Formation of a Single European Educational Space comes into force in Europe [4]. This declaration was also signed by Russia in 2003 [2]. One of the main principles of the Bologna Declaration and the National Skills Standard Board is a competence-based approach to assessing learning outcomes. At the initial stages, the competence-based practice-oriented approach was considered as the prerogative of secondary vocational educational institutions; now it has actively penetrated into higher education as well. The peculiarity of the competence-based approach is that in the learning process not only ready-made knowledge is acquired, but the conditions for the origin of this knowledge are traced. The methodology is based on learning through activity. It is also important that each competence at each level should have a clear assessment system with clear results. Both the learning process and the assessment process should be as transparent and understandable as possible to all participants. c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 435–446, 2022. https://doi.org/10.1007/978-3-030-96299-9_42

436

V. Dobrynin et al.

The competence-based approach is more complex than just a knowledge test: it is structured in a multi-component form. This approach is aimed at “overcoming the main drawback of the existing system of professional training: the gap between the theoretical and practical aspects of the professional activity formed in the course of training” [22]. NPEC work group defines competence as “combination of skills, abilities and knowledge needed to perform a specific task” [11]. This paper aims to introduce a new methodology for assessing digital competencies in the ICT field, and in particular in the Cybersecurity-related topics. In the next Section the motivation of our work will be discussed, while in Sect. 3 is detailed the Competencies Model proposed. In Sect. 4 is proposed a pathway for implementation of our model based on Data Warehousing systems, and the last two Sections are devoted to the discussion of results, the conclusions and the planned future work.

2

Motivation: The Digitalization in 21th Century and the Need for an Assessment Methodology

Since the middle of the last century, there has been growing influence of digital technologies on each spheres of society. The development of the digital economy is one of the priority areas for most countries. The creation of our own digital, high-tech products within the country is a goal of the development of the digital economy which counties try to reach. The requirements for specialists are changing and increasing, as a consequence of these transformation: digital competences are becoming relevant. The COVID-19 pandemic is further evidence of the urgent need to develop digital competences. Many business processes, including educational ones, have been transformed and become digital. Researchers predict that even after all temporary restrictive measures are lifted, many of the processes that became digital during the pandemic will remain online [1,12,20]. While being able to work with digital products is an important ability, it is equally important to be able to work safely with modern technology. The ability to safely handle digital products is reflected in digital security competence. The need for this competence is growing as it is required not only for IT specialists, but also for many others. Most of the work positions are related to the use of computers and access to the Internet at the moment, which directly requires the possession of digital competences, including security competences. In addition, mobile devices with Internet access have become an integral part of the lives of modern citizens. According to statista.com, the number of smartphone users from 2016 to 2021 grew from 3.668 billion to 6.378 billion and continues to grow. The number of smartphone users will reach 7.516 billion by 2026, according to the forecasts of the same site [13]. Moreover, there is evidence that the number of unique mobile internet users stood at 4.28, In 2020 billion. These data confirm the growth in the use of digital devices and the Internet for personal purposes, which expands the circle of citizens who need to develop the competences discussed in the article.

A New Structured Model for ICT Competencies Assessment

437

The rapid growth of cybercrime has been confirmed by many studies [5]. CybersecurityVentures.com experts predicted that cybercrime will cost the global economy $ 6.1 trillion a year by 2021 and reach $ 10.5 trillion in annual losses by 2025 [15] The recent Covid-19 pandemic and the consequently mandatory lockdown (and home-working activities) of business and public sector employees made the situation worse. Interpol reports an increasing rate in Cybercrime activities and new emerging threats related to the pandemic [9]. One solution to this problem can be the development of security competences for the conscious use of digital devices and the Internet. As a conclusion, the development of digital competences in general and security competences in particular is more of a necessity in the modern world than a superfluity. The development of competences is a complex system: the development of educational material, a methodology for checking the formation of competences, etc. In the following chapters, a methodology for assessing the formation of competences and its implementation is proposed.

3

The Methodology for Assessing the Level of Competence

For the full-fledged development of the competence-based approach, it is necessary to create a transparent, understandable system for assessing competences. In this part we present MMACK, a new Meta-model for assessing ICT competencies and knowledge which allows to assess the level of competence formation among students, as well as track a path of their competences development. The source of the research is the result of the authors’ many years of work, tested in the educational process in various disciplines at the Dubna State University. The level of competence is one of the principles of organizing the technology of acquiring meta-knowledge by students. Meta-knowledge is the basis for creative independent acquisition of new knowledge and competencies. This approach improves the organizational, scientific and methodological activities of the teaching staff, as well as the structure and content of the educational and methodological complex of disciplines. Confirmation is the topics of defended master’s theses, published articles. In the proposed model the competence level (CL) is defined as a knowledge of the subject area. We believe, the competence level can be measured as a function (F) of three variables: UNderstanding (UN), Have The Skill (HSK) and ABility (AB). CL = F (U N, HSK, AB)

(1)

In the Formula 1 the three variables are defined as follows. It is to be noted that the HSK and AB are both Skills, which are automated components of a person’s conscious action, which are generated in the process of its implementation. UN - the student’s ability to recognize the meaning (content) of a text or speech. This ability is based on vocabulary of words, concepts, definitions,

438

V. Dobrynin et al.

terms, abbreviations and etc. Generally, a teacher relies on the previous acquired basic knowledge or makes an analogy with the processes already studied when explained a new topic or giving a lecture in a discipline. Additionally, a student has a feeling or adequate perception of the topic, and he grasps what the teacher is explaining. HSK - the ability of a student to solve typical problems of the discipline in an automatic mode (without the involvement of consciousness, on reflexes). Solving problems without involving reflexes can be achieved by solving typical tasks or through training. Important indicators of skill are - the time spent on solving the problem and the quality of the solution (there are mistakes or not). AB - the student’s ability to compare, compare - identify similarities and differences - highlight, argue, reason. The method of performing an action mastered by the subject, provided by the totality of acquired knowledge and skills; the ability to perform an action according to certain rules, and the action has not yet reached automation. It is formed through exercises and creates the ability to perform an action not only in familiar, but also in changed conditions. AB has two interpretations, in contrast of UN: be able to use certain methods that allow you to make a decision; highlight the meaning and essence of what they are talking about or about what they write. This concept reflects the fact that the student not only understood the material, but also analyzed it: comparison, juxtaposition, highlighting, reasoning. Consequently, one chain of thoughts, generate another chain of thoughts - a chain of cause and a chain of effect in reasoning (Table 1). Table 1. Samples of tests for assessing competence variables. Type of test Samples UN test

Students need to find its meaning for the corresponding term

HSK test

Students have to determine what essence is reflected in the selected text and how this essence is highlighted

AB test

To give students different interpretation of a concept and they have to note the similarities and differences between interpretations of the concept

The levels of UN, HSK and AB are determined by a system of testing rules (ST). ST includes: tests (TST), testing rules (W), rules for assessing the test result. As a result of testing, the levels of UN, HSK and AB are determined. It is assumed that the tests to be carried out multiple times. We assume that tests are conducted once time a semester, however, the frequency of tests may vary depending on the education (training, teaching) system. The Fig. 1a shows the change of CL depending on time (the figure depicts the CL of 4 students who were measured three times in some intervals). Moreover, this system allows to track the change in each indicator of CL (UN, HSK, AB) over time. Figure 1b shows the CL model, where the x, y, z axes

A New Structured Model for ICT Competencies Assessment

439

Fig. 1. a) Time-dependent student measurement graph; b) Three-dimensional representation of changes in student knowledge. (Color figure online)

correspond to the level UN, HSK, AB. Consequently, the cube reflects a set of levels UN, HSK, AB. The scale of levels UN, HSK, AB consists of 3 levels: low, average and high. The black cube reflects the medium level UN, HSK, AB. The scale of levels UN, HSK, AB consists of 3 levels, Therefore, the cube we are considering consists of 27 cubes - spaces where a point with a student’s level of knowledge can be (Fig. 2). From the central cube, cubes can be laid left-right, up-down along three axes (UN, HSK, AB). Red, yellow and green are the colors of the cubes that correspond to the low, average and high CL levels.

Fig. 2. Displaying levels of UN, HSK, AB. (Color figure online)

A sample of three-dimensional representation of students’ knowledge over time is presented in the Fig. 3. The y-axis is the names of the students and the z-axis is the time axis. This system allows to make a slice for a particular student and see the dynamics of changes in the level of each indicator. Dynamics of changes in CL in time for one student is presented in the Table 2. General recommendations for studying, recommendations of additional materials, and also refine the education program can be made based on the results analyze of the CL for a specific group of students (Fig. 3). Data from Table 2

440

V. Dobrynin et al.

Fig. 3. Three-dimensional representation of students’ knowledge over time. Table 2. Slice of a cube. Dynamics of changes in SL in time for one student. Time UN

HSK

AB

T1

30

40

32

T2

37

47

30

T3

34

53

36

...

...

...

...

can be used for a deeper analysis for each student: to determine the dynamics of particular student development in a certain subject area and, on the basis of it, make recommendations for further studying.

4

Toward Competence Model Implementation

In order to plan an on-field test of our new Competence Model, we choose to explore the feasibility of implementation using Data Warehousing (DWH) software. A Data Warehousing may be defined as a data-driven decision support system that supports the decision-making process in a strategic sense and, in addition, operational decision-making [14]. The early implementations of decision support systems based on data extracted from operational systems dates back to late 80s, and in the 1991 Bill Inmon, generally considered the father of data warehousing, publishes the book Building the Data Warehouse [8], in which DWH systems have been described in terms of architecture and data modelling. Since the early 90s, big Companies are using DWH systems to support their strategic decision-making processes. In the last years, also government agencies make use of DWH systems for the same purpose. In the education sector DWH systems are used by some Universities as a decision support system, and an overview of the scientific papers about DWH systems in the education sector

A New Structured Model for ICT Competencies Assessment

441

may be found in [16] and [10]. A remarkable paper dealing with the use of DWH system in Higher Education is [17], and a paper that deals on enhancing digital competences for Higher Education is [21]. In DWH systems, the information is delivered to users using Online Analytical Processing (OLAP), which is a technique that enables multi-dimensional and multi-level analysis on a large volume of data, providing aggregated data visualizations with different perspectives [19]. The OLAP processing is based on multidimensional analysis of data in the form of “Facts”; a fact is a measure (or a set of measures) which represents a fact about the managed entity or system. In Fig. 4 is shown an example of the hypercube representation of the sales of a firm: The measurement of fact are the units sold and the amount of the revenue, and the dimensions of hypercube are Time, Products and Branch related to a particular Sales fact.

Fig. 4. The hypercube structure of Facts.

It is clear that the OLAP hypercube structure fits in perfectly with the representation of our competences model as shown in Fig. 3, so it’s worth to explore the feasibility of the implementation using Data Warehouse software. In the Table 3 are listed the main advantages of this choice and the issues to be faced. Those issues require that the implementation phase of our competences model must be preceded by a careful analysis in the design phase. In this paper, we deal with a preliminary design framework.

442

V. Dobrynin et al. Table 3. Advantages and issues of using DWH in implementation.

5

Advantages

Issues

– If DWH software is already in use, the cost and the effort of this implementation may be very low – If DWH software is not yet used, the introduction of this king of software may be used by University to implement a number of features to support decision-making processes (e.g. scientific production analysis, cash flow monitoring, ...) – The data may be automatically extracted from existing Student Career University systems – Ease of competences model extension (e.g. adding new attributes) – Opportunity to perform competence analysis in relationship with other student’s attributes (curriculum, learning cycle, ...)

– The use of commercial (i.e.: non open-source) DWH software may be very costly – The introduction of DWH could lead to a heavy organisational effort for the University

Architecture of the Proposed System

For the implementation, a simple two-layers architecture is chosen [7]. The raw data, originating from the Student Career DB, are subjected to a filtering phase in a Data Staging area using ETL (Extract, Transform and Load) Tools and loaded to the Data Warehouse, and then the users can process the data of own interest via Data Marts (Fig. 5). The results of the Competences Test have to be loaded on the Student Career DB of the University, and the DWH database is fed with data via ETL Tools. Going further in the design phase, prior to design the logical model (the STAR schema), we decide to use a conceptual model to provide a higher level of abstraction in describing the warehousing process and the structure of the information needed in all its aspects. The model proposed by Golfarelli et al. [6], namely Dimensional Fact Model (DFM), was chosen due to its simplicity and ease of implementation. The DFM diagram for Competences tests is drawn in Fig. 6. The Fact is Competences Test and the measures are UN, HSK and AB, the three variables of the function F (Eq. 1). There are three dimensions of interest in this preliminary design: Time, Student and Course. Time is hierarchically structured in two alternative ways, Day-Month-Year and Day-Semester-Year, in order to enable queries both based on academic time or in natural time. Also the dimension Student is structured in a simple hierarchy: we use only Student Career as an example, but many different structures may be designed.

A New Structured Model for ICT Competencies Assessment

443

Fig. 5. Basic architecture of DWH system.

Fig. 6. DFM conceptual diagram of Competences test.

To design the logical model, we use a CASE tool to produce the STAR schema [18] which defines the tables, the attributes, and the relationships, according with the data structure represented using the DFM model previously designed. The STAR schema has been automatically obtained using this tool, exploiting its ability to forward-engineer the conceptual model and generate the schema for the target database system you need to work with. We use for this work the experimental CASE tool BIModeler (Business Intelligence Modeler) [3].

444

V. Dobrynin et al.

Fig. 7. The STAR schema of Competences Test.

In Fig. 7 is drawn the STAR schema, in which the table beginning with “F ” is the fact and the tables beginning with “D ” are the three dimensions (Course, Student and Time).

6

Discussion

The presented research topic is part of a set of problems associated with training specialists with a high level of motivation for independent development and improvement of such qualities as insight, empathy (in a broad sense) and empathy as a tool of cognition, reflexivity, intuition, criticality, rationality in the field of ICT. These concepts used are fundamental for the study of various complexes of competencies, for which it is necessary to create measurement scales, measurement methods, assessment methods, comparison methods, using modern technologies (fuzzy sets, soft computing, fuzzy logics, artificial intelligence and modeling methods). From this point of view, one can consider the dynamics of the development of the presented topic, both from a theoretical point of view and from a practical point of view. From the point of view of futurology, the topic raised is a small (almost insignificant, but important) part of the digitalization of the world education system, in which the main problem from the present time (the time of informatization and robotization and intellectualization) is what a person is (as a bioenergy information system) of the future, living and creating in the digital world.

A New Structured Model for ICT Competencies Assessment

7

445

Conclusions and Future Work

In this paper has been presented a new competence meta-model MMACK, the future works for competence model will to develop tests which allow to estimate competence variables (UN, HSK, AB). The tests will be focused on measurement the level of Cybersecurity-related competences. In this research work has also been discussed a possible implementation of the competence model using Data Warehousing software, and have been presented a basic architecture and the conceptual and logical design. The future work for implementation of the system will be the choice of the DWH software to be used; in this way a proof-of concept of a working system will led us to conduct the first practical experiments.

References 1. Barragan-Sanchez, R., Corujo-Velez, M.C., Palacios-Rodriguez, A., RomanGravan, P.: Teaching digital competence and eco-responsible use of technologies: Development and validation of a scale. Sustainability 12(18), 7721 (2020). https:// doi.org/10.3390/su12187721, https://www.mdpi.com/2071-1050/12/18/7721 2. Burkov, A.V.: Implementation of bologna declaration: Problems in Russian higher education. Vestnik Samara State Univ. Econ. 9(47), 9–12 (2008) 3. Cazzella, S.: Business intelligence modeler, alpha version - b115 (1998). http:// www.bimodeler.com/ 4. Commission, E.U.: The bologna process and the european higher education area (2020). https://ec.europa.eu/education/policies/higher-education/bolognaprocess-and-european-higher-education-area en 5. Conteh, N., Royer, M.: The rise in cybercrime and the dynamics of exploiting the human vulnerability factor. Int. J. Comput. 20(1), 1–12 (2016). https://ijcjournal. org/index.php/InternationalJournalOfComputer/article/view/518 6. Golfarelli, M., Maio, D., Rizzi, S.: The dimensional fact model: A conceptual model for data warehouses. Int. J. Coop. Inf. Syst. 07(02n03), 215–247 (1998). https:// doi.org/10.1142/S0218843098000118, https://doi.org/10.1142/S0218843098000118 7. Golfarelli, M., Rizzi, S., Turricchia, E.: Modern software engineering methodologies meet data warehouse design: 4wd. In: Cuzzocrea, A., Dayal, U. (eds.) Data Warehousing and Knowledge Discovery, pp. 66–79. Springer, Heidelberg (2011) 8. Immon, W.H.: Building the Data Warehouse. Wiley, New York (1993) 9. Interpol: Covid19: Cybercrime analysis report - august 2020 (2021). https://www. interpol.int/content/download/15526/file/COVID19CybercrimeAnalysisReportAugust2020.pdf 10. Jayashree, G., Priya, C.: Comprehensive guide to implementation of data warehouse in education. In: Peng, S.L., Son, L.H., Suseendran, G., Balaganesh, D. (eds.) Intelligent Computing and Innovation on Data Science, pp. 1–8. Springer Singapore, Singapore (2020) 11. Jones, E.A., Voorhees, R.A.: Defining and Assessing Learning: Exploring Competency-Based Initiatives. Report of the National Postsecondary Education Cooperative Working Group on Competency-Based Initiatives in Postsecondary Education. U.S. Department of Education, National Center for Education Statistics (2002)

446

V. Dobrynin et al.

12. Lasi´c-Lazi´c, J., Milkovi´c, M., Rosanda ˇzigo, I.: Digital competences as core competences for lifelong learning, pp. 5911–5915 (2020). https://doi.org/10.21125/ edulearn.2020.1537 13. Ceci, L.: Mobile internet usage worldwide - statistics & facts. https://www.statista. com/topics/779/mobile-internet/ 14. Linstedt, D., Olschimke, M.: Building a Scalable Data Warehouse with Data Vault 2.0. Morgan Kaufmann, Burlington (2015) 15. Morgan, S.: Cybercrime to cost the world $ 10.5 trillion annually by 2025. https:// cybersecurityventures.com/cybercrime-damages-6-trillion-by-2021/ 16. Moscoso-Zea, O., Paredes-Gualtor, J., Luj´ aN-Mora, S.: A holistic view of data warehousing in education. IEEE Access 6, 64659–64673 (2018). https://doi.org/ 10.1109/ACCESS.2018.2876753 17. Namnual, T., Nilsook, P., Wannapiroon, P.: System architecture of data warehousing with ontologies to enhance digital entrepreneurs’ competencies for higher education. Int. J. Inf. Educ. Technol. 9(6), 414–418 (2019) 18. Ponniah, P.: Data Warehousing Fundamentals for IT Professionals, 2nd edn. Wiley, New York (2010). https://doi.org/10.18178/ijiet.2019.9.6.1237 19. Queiroz-Sousa, P.O., Salgado, A.C.: A review on OLAP technologies applied to information networks. ACM Trans. Knowl. Discov. Data 14(1), 1–25 (2019). https://doi.org/10.1145/3370912 20. S´ a, M.J., Serpa, S.: Covid-19 and the promotion of digital competences in education. Univ. J. Educ. Res. 8(10), 4520–4528 (2020) 21. Santoso, L.W.: Yulia: Data warehouse with big data technology for higher education. Procedia Comput. Sci. 124, 93–99 (2017) 22. Nyushchenko, V.G.: On the readiness of the teaching staff to implement the competence based approach in teaching [in Russian], pp. 264–266 (2012). https://doi.org/10.26439/ciis2021.5581, https://revistas.ulima.edu.pe/index.php/ CIIS/article/view/5581 23. Voorhees, R.A.: Competency-based learning models: A necessary future. New Dir. Inst. Res. 2001(110), 5–13 (2001)

Automated Evaluation Tools for Web and Mobile Accessibility: A Systematic Literature Review João Dias1

, Diana Carvalho1,2(B) , Hugo Paredes1,2 , Paulo Martins1,2 Tânia Rocha1,2 , and João Barroso1,2

,

1 UTAD - University of Trás-os-Montes e Alto Douro, Quinta de Prados, 5001-801 Vila Real,

Portugal {dianac,hparedes,pmartins,trocha,jbarroso}@utad.pt 2 INESC TEC, Rua Dr. Roberto Frias, 4200-465 Porto, Portugal

Abstract. This research aims at investigating which web accessibility and usability tools, with the focus on the ones that warrant automation, are available to assess the quality of interfaces for people with disabilities and/or special needs, enabling them to navigate and interact with web and mobile apps. Our search strategy identified 72 scientific articles of the most rated conferences and scientific journals, from which 16 were considered for the systematic literature review (SLR). We found that, despite the existence of various tools either for web or mobile apps, they are not completely effective, covering less than 40% of all the problems encountered. Also, no tool was found capable of adapting the application interfaces according to the type of disabilities that users may present. Therefore, a new tool could be a welcome advancement to provide full accessible and usable experiences. Keywords: Multimedia interfaces · Web accessibility and usability · Mobile accessibility · Digital inclusion · Automatic evaluation tool · Disability

1 Introduction The Information and Communication Technologies (ICT) have become a commodity that everyone needs, as it is built upon the most important commodity of the next millennium: information [16], accessed especially through most typical search engines or via mobile devices. It has great potential for making significant improvements in the lives of disabled people, as they can compensate physical or functional limitations by developing content that allows them to perceive, understand, navigate and interact with web and mobile apps. As we move towards a highly connected world, it is critical that the web be accessible and usable by anyone, regardless of individual disabilities [16]. Users are a heterogeneous and multicultural public, with different abilities and disabilities (visual, hearing, cognitive and motor impairments) [13], and they should be able to search information and interact with the web regardless of the device [19]. As Tim Berners-Lee said, “The power of the web is in its universality. Access by everyone regardless of disability is an essential aspect” [22]. According with World Health Organization (WHO), Disability © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 447–456, 2022. https://doi.org/10.1007/978-3-030-96299-9_43

448

J. Dias et al.

is an umbrella term for impairments, activity limitations and participation restrictions [23]. Disabilities can affect people in different ways, even when one person has the same type of disability as another person. Disability covers impairments, activity limitations, and participation restrictions. Impairment be a problem in body function or structure, inactivity limitation is a struggle faced by a person in executing a task or action, while a participation restriction is a problem encountered by a person involving life situations [5]. Most websites and mobile apps available are developed under the tacit assumption that the final user will not have any kind of disability. Thus, accessibility aspects are not contemplating in their development process [17] and many users encounter interaction barriers and cannot access much of the content in web or mobile apps [17]. Sometimes due to these problems, people are unable to deal with certain bureaucracies online, having to move to institutions, causing enormous inconvenience. Another big problem associated with accessibility is loss of the potential economic value of people with disability [3]. In order to make people use the information, there are a set of recommendations and strategies to expand universal access of use of the ICT. Web should have three fundamental characteristics: be accessible, usable, and inclusive. The European Union (EU) Directive uses the four principles of the WCAG 2.1, requiring that public sector organizations across the EU take steps to make sure their websites are “Perceivable, Operable, Understandable and Robust” [18]. This paper aims to review the literature about automated evaluation tools for web and mobile accessibility and usability. We analyze their strengths and weaknesses, what can be done to improve evaluation and usability of the applications and identify the tools related with accessibility assessment. To best achieve this aim, it was decided that a SLR would be developed, according to the guidelines established by Kitchenham [12]. The results of such SLR will provide both a valuable research for developing new applications keeping in mind the issues for accessibility and usability as well as the development of applications that allow revolutionizing the man-machine iteration, making accessible and usable the applications that already exist and do not follow the WCAG 2.1 criteria. This document is structured in six sections: the method is described in Sect. 2; the data extraction results are shown in Sect. 3; the discussion of research questions is represented in Sect. 4; we conclude with Sect. 5 and future work in Sect. 6.

2 Method SLR is a type of literature review and an approach to conducting a rigorous study with a goal of collecting data and found evidence in order to draw conclusions about research investigation, giving responses to our research questions. To carry out the task of produce an SLR on automated evaluation tools for web and mobile accessibility and usability, we used the methodology suggested by Kitchenham. The author [14] proposed the guideline for systematic reviews in software engineering, created as adaptation of several existing guidelines, mainly from medicine. In computer science most relevant sources are online and the research process concentrates in scientific databases [8, 14]. The 3 steps required in doing an SLR are as

Automated Evaluation Tools for Web

449

follows [12]. The first step is to identify the need for the review, ask research questions, define the review protocol and validate it. The second step is to make a systematic review, identifying relevant researches, selecting primary studies, evaluating the quality of them, extracting the relevant data and synthesizing the extracted data. Lastly, it is necessary to report the systematic review, write and validate the review report. 2.1 Identifying the Need for the Systematic Review One of our goals with this study is to evaluate the need of a new platform or tool that allows to have better web and mobile apps, so that people with disabilities can use it. To do that, we analyze multiple works of research through a systematic process in the field of automated evaluation tools for web accessibility and usability. It is also necessary to know if there are tools that make pages accessible and usable for the user automatically and adapted to their disability type. 2.2 Research Questions As the first step in our SLR process, we formalized the goals of our review into the following set of research questions: RQ1 - Which considerations can be drawn about automated accessibility evaluation tools in web and mobile apps? RQ2 - Which technological solutions can be implemented to minimize the problems of accessibility and usability on web and mobile apps? RQ3 - What are the existing tools on the market? 2.3 Exclusion Criteria We did not consider the following types of work for this SLR: papers that were not written in English; papers that only have the title and/or abstract written in English; articles added to the searched databases before 2010; multiple versions of the same article; unreviewed reports and technical papers; theses; covers (journal or conference proceedings covers); posters; discussions; low relevant websites and PowerPoint presentations. 2.4 Search Process We designed an SLR protocol to guide the search process to find articles that would answer the research questions. Databases of scientific papers were queried. The decision as to what databases to choose was made in three steps. First, an SLR (such as reference [1]) was read to see what databases they used to automatically retrieve relevant papers. At this point, several keywords were extracted. Many of them are variations of the same terms, as shown in Table 1. From the terms listed, we defined as our core the following: Web Accessibility; Mobile Accessibility; WCAG 2.1; Automatic Evaluation Tool; Disability. The period of SLR covers the last 11 years, from 2010 until 2021, the time period covered by the original expert literature review in order to analyze the current state of the art. We performed our database search using logical commands “AND” and “OR” for a chain of keywords, looking to have as much relevant information as possible to answer our RQs.

450

J. Dias et al. Table 1. Search terms synonyms.

Sets

Synonyms keywords

Set 1

Web accessibility, mobile accessibility, accessibility, accessibility metrics

Set 2

Usability, usability guidelines, universal usability

Set 3

Digital inclusion, social inclusion, inclusion

Set 4

Automatic assessment tool, automatic evaluation, evaluation tools, accessibility evaluation

Set 5

Assistive, assisted technology, assistive technology

Set 6

Disability, disabled, impaired, impairment, rehabilitation

We conducted the search through query strings applied into digital libraries to find as many of the known empirical studies as possible. Three digital libraries were considered: ACM Digital Library, IEEE Xplore and SCOPUS. The automated search to our SLR was performed on November 13th, 2021 and the query that returned the best result was: “web accessibility” OR “web usability” AND “automatic assessment tool” OR “automatic evaluation tool” AND “disabled people” OR “disability”. The combination of keywords resulted in a set of 11314 papers based from the past 11 years. Then, refining search terms we obtained a set of 1426 journal articles, 7740 conference papers and 510 magazines. Consequently, in phase 2, we filtered by title (T), by abstract (A) and finally by introduction (I) to obtain 289 potentially relevant studies. In phase 3, according to benefits and quality (BF) attributes expected, we selected 101 papers and, finally, of those we selected the most relevant 16 articles to present a variety of automatic evaluation tools for web and mobile accessibility and usability. The results obtained can be found on Table 2. Table 2. Summary of our search results Database name

Results

Jour./Conf./Mag.

T

A

I

BF

Relevant papers

ACM

5111

499/4169/179

457

243

194

46

7

IEEE Xplore

2391

286/2024/36

557

241

79

45

3

Scopus

3812

641/1547/295

140

84

16

10

6

11314

1426/7740/510

1154

568

289

101

16

Total

2.5 Data Extraction In this step, data are collected from each of the studies included in the review. This is done using the data collection form, which was designed during the development of the review protocol. Information records were saved in an Excel sheet, containing the following fields: 1) article ID; 2) title; 3) keywords; 4) abstract; 5) publication date; 6) authors;

Automated Evaluation Tools for Web

451

7) conclusion; 8) publication name; 9) publication type; 10) number of citations; 11) number of references; 12) quartile and 12) H-index. Later, the Excel sheet was exported into a Sql database and a PHP application was developed, to facilitate the analysis of each article inserted in the database. After defining which articles were relevant to the literature review, we performed an initial analysis on the titles and abstracts; and then a more complete one focused on the introduction and conclusions section of the works. Finally, we evaluated the relevance of the studies to answer our RQs.

3 Data Extraction Results In the literature, it is possible to find studies related to automated evaluation tools for web and mobile accessibility in order to understand the problems of many web and mobile apps and their compliance with the existing accessibility guidelines. We found ten studies that have similarities with ours. Below we analyze the pros and cons of these tools and draw ideas for the implementation of our platform. In article [6], Google Chrome plugins help to test accessibility criteria during the development phase of applications. Each plugin presents different results regarding the success criteria. The analysis showed that individual tools have poor coverage of the WCAG 2.1 success criteria. Thus, human analysis is always recommended. The disadvantage is that the plugins are limited to Google Chrome. In article [24] MAC results showed that 67% of accessibility issues was identified by the tool. As future work, the authors mention the scope should be expanded to cover the issues related to people with learning and cognitive disabilities using technologies, such as image recognition and natural language processing. In this way people with other types of disabilities will remain excluded. In article [17], WAL tool introduces accessibility features at runtime in legacy Web systems via dynamic source code manipulation. It also works as a plugin that puts an accessibility bar on websites. It already covers some types of disabilities, such as visual impairment, dyslexia, and reading difficulties. Does not yet include user authentication. It’s a good tool, although works only like a plugin. In article [10], Automatica11y tool is able to fix source code issues that can be detected by the HTML Code Sniffer. The number of errors on the website is reduced by using this tool and is faster than manual refactoring methods. It only works as an HTL code parser. In article [2], SiMor is presented as a valuable validation tool that gathers and centralizes numerous functionalities and characteristics that are not fully contemplated and integrated in anyone validators. However, to date it only follows the WCAG 2.0 guidelines. In article [27], with Active-Prediction, artificial intelligence is inserted into web accessibility evaluation and proposes a semi-supervised machine learning method to obtain all the accessibility results by predicting. It’s a novel method to obtain the accessibility results AND we can use into our platform in future. In article [4], Social4all allows a set of accessibility problems to be solved without modifying the original page code. The proposed platform can analyze websites and

452

J. Dias et al.

detect many accessibility problems automatically; after this, a guided assistant is used to offer adequate solutions to each detected problem. In article [11], the authors identified AChecker as the best open-source tool currently available to check web accessibility. Moreover, only the WCAG 2.0 guidelines are supported currently. In article [26], the authors presented URLSamp, a novel web page sampling method based on URL clustering. It’s very limited to our scope. In article [9], the authors, after comparing the results of six accessibility assessment tools applied to a set of web pages, concluded that it is not recommended employing only the tools and leave out human judgment.

4 Discussion of Research Questions 4.1 RQ1 - Which Considerations Can Be Drawn About Automated Accessibility Evaluation Tools in Web and Mobile Applications? Automated accessibility evaluation tools are applications that analyze web page code verifying specific sets of guidelines [25]. Its advanced search functionality enables evaluation tools to be looked for according to various criteria sets of guidelines, language, type of tool, technology, provided assistance, scope and license type. It can also be classified by type: API (Application Programming Interface, authoring tool plugin, browser plugin, command line tool, desktop application, mobile app and online tools) [7]. Usability guidelines are supposed to help web designers to design usable websites. But unfortunately, studies carried out that applying these guidelines by designers is difficult, essentially because of the way of structuring or formulating them [20]. Many tools present in the studies analyzed are of this type, i.e. they are limited to analyzing the source code and detecting accessibility errors based on the WCAG guidelines, contributing little or nothing to making it instantly accessible. It would be very useful to have good tools to evaluate the accessibility and usability of applications, initially web-based, and more recently for mobile devices. Although there are numerous assessment tools to date, the literature does not seem to provide an answer to the problems related to automatic inspection of the WCAG 2.1 guidelines for the purpose of assessing the accessibility and usability of web and mobile apps in the most proper way. The analysis of effectiveness of the 10 state of the art accessibility evaluation tools presented in the previous section, in terms of coverage, completeness and correctness corroborates that employing tools alone and leaving out human judgment is indeed not recommended. As stated by the WAI “We cannot check all accessibility aspects automatically. Human judgment is required. Sometimes evaluation tools can produce false or misleading results. “Web accessibility evaluation tools cannot determine accessibility, they can only assist in doing so” [21]. 4.2 RQ2 - Which Technological Solutions Can Be Implemented to Minimize the Problems of Accessibility and Usability on Web and Mobile Apps? Regarding the 10 selected studies related to automatic evaluation tools for web and mobile accessibility and usability, we can see that evaluation tools already exist in quite

Automated Evaluation Tools for Web Table 3. Summary of automated evaluation tools for web accessibility and their features (adapted from [7]).

Table 4. Summary of automated evaluation tools for mobile accessibility and their features (adapted from [15]).

Tool

Type

Features

Tool

OS

Features

A-Checker

Online

AE, MF, PM, ST

Android Lint

Android

AC, ANT

Accessibility Check

Online

AE, RG

Expresso

Android

CC, TTS, S, AC, ANT

Eval Access Online Funconal Accessibility Online Eval- uator

AE, RG Roboletric

Android

CC, TTS, S, AC, ANT

MATE

Android

CC, TTS, S, AC, ANT

ForApp

Android

V, ANT

Accessibility Scanner

Android

CC, TTS, S, ANT

PUMA

Android

TTS, S, ANT

IBM AbilityLab

Android

CC, ANT

Hera

Online

HiSoware® Cynthia Online Says™ Portal

AE, RG, ST AE, MF, PM, CV, RG, ST, MS AE, RG

Sortsite

Online/Offline AE, RA, ST, TG

TAW

Online

AE, RG

Torquemada

Online

AE, RG

Total validator

Online

AE, RG

WAVE

Online

AE, PM

TAW Standalone

Applicaon

AE, MF, PM, CV, LP, RA, MS, FL

Web Accessibility inApplicaon spector

AE, LP, RG

WAVE Firefox toolbar Extension

AE, PM, LP, RA, RP

Mozilla/Firefox AccessiExtension bility

AE, LP, RA, RP, RG

Foxability

AE, RA, RP, RG, FL

Extension

453

Accessibility Inspector IOS

C, ANT

Mobile web AccesIOS sibility Checker

CC, V, K, ANT

EarlGrey

IOS

CC, TTS, S, AC, ANT

KIF

IOS

CC, TTS, S, AC, ANT

AccScope

Windows PhoneANT

Inspect

Windows PhoneC, ANT

AccChecker

Windows PhoneAC, C, ANT

AC - Actionable elements; AE - Automatic Evaluation; ANT - Alternatives for non-text; C - Consistency; CC - Color Contrast; CV - Annotated code view; FL - Flexibility; K - Keyboard; LP - Local Pages; MF - Manual filling; PM - Page presentation modification; RA - Restricted-access pages; RP Rendered-page evaluation; S - Spacing; ST Support for training; TTS - Touch target size; V – Visible

considerable number, each with its strengths and weaknesses. To help designers and developers, it would be good to have a tool that evaluate applications (during and after design) and alert them to errors. Besides that, there was no mention of any tool capable of converting web or mobile app transparently for the user, into accessible and usable applications, according to an adaptability system based on the user’s profile. This means that, depending on the type of profile registered—visual, hearing, cognitive and motor impairments—and after logging into the system, the platform will adapt the web page or application to be used, to the type of disability of the user. In other words, the platform consists of a website and app where the page to visit or the app to emulate is inserted, and it will take care of making them accessible and usable by users, even if there are

454

J. Dias et al.

deficiencies in source code. This is where the greatest contribution of this research work may lie. 4.3 RQ3 - What Are Similar Existing Tools on the Market? We briefly reference each testing tool found during investigation. For each tool we identify which accessibility guidelines it provides support. Table 3 presents free automated evaluation tools for web accessibility found during the research. The accessibility of mobile apps can also be tested by automated tools. Table 4 presents automated evaluation tools for mobile apps found during the research.

5 Conclusion This paper reports the results of SRL on automated evaluation tools for web and mobile accessibility and usability. Given the high importance of the topic, researchers contribute to thousands of articles published over the electronic database. Hence, to refine these bulky amounts of articles, SLR is clearly the best approach. One of the study’s limitations is the restriction of database search from 3 databases. Although we believe, the results do provide reasonable insights into the state of the art of automated evaluation tools for web and mobile accessibility and usability. To date, the literature seems to report some solutions to the problems of automatically inspecting the guidelines with the aim of evaluating websites and mobile apps. There are many evaluation tools on the market capable of evaluating websites. Although, there was no mention of any tool capable of converting web or mobile applications transparently for the user, into accessible and usable applications, according to an adaptability system based on the user’s profile. This means, depending on the type of user profile, and after logging into the system, the platform will adapt the web page or application to be used to the type of disability of the user. This is where the greatest contribution of this research work may lie. Acknowledgments. The research leading to these results has received funding from the European Union’s Horizon 2020 - The EU Framework Programme for Research and Innovation 2014–2020, under grant agreement 101006468.

References 1. Almendra, N., Moquillaza, A., Paz, A.: Web Accessibility Evaluation Methods: A Systematic Review (2019). https://doi.org/10.1007/978-3-030-23535-2_17 2. Banchoff, C., Harari, I., Rajoy, G., Defalco, M.: SiMor: An intensive web accessibility analyzer based on rules. In: Proceedings of the 2016 42nd Latin American Computing Conference, CLEI 2016, pp. 1–6 (2017). https://doi.org/10.1109/CLEI.2016.7833407 3. Baptista, A., Martins, J., Gonçalves, R., Branco, F., Rocha, T.: Web accessibility challenges and perspectives: A systematic literature review. In: 2016 11th Iberian Conference on Information Systems and Technologies (CISTI), pp. 1–6 (2016). https://doi.org/10.1109/CISTI. 2016.7521619

Automated Evaluation Tools for Web

455

4. Crespo, R.G., Espada, J.P., Burgos, D.: Social4all: Definition of specific adaptations in Web applications to improve accessibility. Comput. Stand. Interf. 48(2016), 1–9 (2016). https:// doi.org/10.1016/j.csi.2016.04.001 5. Disabled World: Disabilities: Definition, Types and Models of Disability - Disabled World, pp. 1–3 (2019). https://doi.org/DW#170-17.172.98-5a 6. Frazão, T., Duarte, C.: Comparing accessibility evaluation plug-ins. In: Proceedings of the 17th International Web for All Conference (W4A ’20), pp. 1–11. Association for Computing Machinery, New York, NY, USA (2020). https://doi.org/10.1145/3371300.3383346 7. Fuertes, J.L., González, R., Gutiérrez, E., Martínez, L.: Hera-FFX: A Firefox add-on for semi-automatic web accessibility evaluation. W4A 2009 - International Cross-Disciplinary Conference on Web Accessibility, pp. 26–35 (2009) 8. Genero, M., Cruz-Lemus, J., Piattini, M.: Métodos de investigación en ingeniería del software (2015) 9. Harper, S., Yesilada, Y., Vigo, M.: W4A Camp Report: “2012 Edition”. In: Proceedings of the 10th International Cross-Disciplinary Conference on Web Accessibility (W4A ’13). Association for Computing Machinery, New York, NY, USA (2013). https://doi.org/10.1145/ 2461121.2461145 10. Ikhsan, I.N., Candra, M.Z.C.: Automatically: An automated refactoring method and tool for improving web accessibility. In: Proceedings of 2018 5th International Conference on Data and Software Engineering, ICoDSE 2018, pp. 1–6 (2018). https://doi.org/10.1109/ICODSE. 2018.8705894 11. El Kabani, I., Zantout, R., Hamandi, L., Mansi, S.: Improving web accessibility. Int. J. Appl. Eng. Res. 11(12), 7836–7845 (2016). https://www.scopus.com/inward/record.uri?eid=2-s2. 0-84989199345&partnerID=40&md5=03a7f4af17508a4f7b923a0bdf2dbf65 12. Kitchenham, B.: Procedures for Performing Systematic Reviews, Version 1.0. Empir. Softw. Eng. 33(2004), 1–26 (2004) 13. Luján-Mora, S., Masri, F.: Evaluation of web accessibility: A combined method. Information Systems Research and Exploring Social Artifacts: Approaches and Methodologies. April 2018, pp. 314–331 (2012). https://doi.org/10.4018/978-1-4666-2491-7.ch016 14. Quiñones, D., Rusu, C.: How to develop usability heuristics: A systematic literature review. Comput. Stand. Interf. 53(2017), 89–122 (2017). https://doi.org/10.1016/j.csi.2017.03.009 15. Silva, C., Eler, M.M., Fraser, G.: A survey on the tool support for the automatic evaluation of mobile accessibility. In: ACM International Conference Proceeding Series (DSAI 2018), pp. 286–293. Association for Computing Machinery, New York, NY, USA(2018). https://doi. org/10.1145/3218585.3218673 16. Sohaib, O., Kang, K.: E-Commerce web accessibility for people with disabilities. In: Gołuchowski, J., Pa´nkowska, M., Linger, H., Barry, C., Lang, M., Schneider, C. (eds.) Complexity in Information Systems Development. LNISO, vol. 22, pp. 87–100. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-52593-8_6 17. Teotonio, W., Gonzalez, P., Maia, P., Muniz, P.: WAL: A tool for diagnosing accessibility issues and evolving legacy web systems at runtime. In: 2019 19th International Working Conference on Source Code Analysis and Manipulation (SCAM), IEEE, pp. 175–180 (2019). https://doi. org/10.1109/SCAM.2019.00028 18. The Chang School and Digital Education Strategies: The Evolution of Web Accessibility – Professional Web Accessibility Auditing Made Easy (2019). Retrieved from 22 June 2020. https://pressbooks.library.ryerson.ca/pwaa/chapter/the-evolution-of-web-accessibility/ 19. Ullmann, D., Jones, H., Williams, F., Williams, R.C.: Information and communications technologies for the inclusion and empowerment of persons with disabilities in Latin America and

456

20.

21.

22.

23. 24. 25. 26.

27.

J. Dias et al. the Caribbean. In Information and communications technologies for the inclusion and empowerment of persons with disabilities in Latin America and the Caribbean, United Nations publication (2018). Retrieved from https://repositorio.cepal.org/bitstream/handle/11362/43744/ 4/S1800975_en.pdf Vanderdonckt, J., Beirekdar, A., Noirhomme-Fraiture, M.: Automated evaluation of web usability and accessibility by guideline review. In: Koch, N., Fraternali, P., Wirsing, M. (eds.) ICWE 2004. LNCS, vol. 3140, pp. 17–30. Springer, Heidelberg (2004). https://doi.org/10. 1007/978-3-540-27834-4_4 WAI: Selecting {Web} {Accessibility} {Evaluation} {Tools}. Web Accessibility Initiative (WAI) (2017). Retrieved from 29 January 2020 https://www.w3.org/WAI/test-evaluate/tools/ selecting/ Williams, J.R.: Developing Performance Support for Computer Systems: A Strategy for Maximizing Usability and Learnability, 1st edn. CRC Press (2004). https://doi.org/10.1201/978 0203228531 World Health Organization: Disabilities (2020). Retrieved from 13 November 2021. http:// www.emro.who.int/health-topics/disabilities/index.html Yan, S., Ramachandran, P.G.: The current status of accessibility in mobile apps. ACM Trans. Access. Comput. 12(1), 1–31 (2019). https://doi.org/10.1145/3300176 Yesilada, Y., Harper, S. (eds.): Web Accessibility. HIS, Springer, London (2019). https://doi. org/10.1007/978-1-4471-7440-0 Zhang, M.N., Wang, C., Bu, J.J., Yu, Z., Zhou, Y., Chen, C.: A sampling method based on URL clustering for fast web accessibility evaluation. Front. Inf. Technol. Electron. Eng. 16(6), 449–456 (2015). https://doi.org/10.1631/FITEE.1400377 Zhang, M., Wang, C., Bu, J., Li, L., Yu, Z.: An optimal sampling method for web accessibility quantitative metric and its online extension. Internet Res. 27(5), 1190–1208 (2017). https:// doi.org/10.1108/IntR-07-2016-0205

My Buddy: A 3D Game for Children Based on Voice Commands Diana Carvalho1,2(B)

, Tânia Rocha1,2

, and João Barroso1,2

1 UTAD - University of Trás-os-Montes e Alto Douro, Quinta de Prados, 5001-801 Vila Real,

Portugal {dianac,trocha,jbarroso}@utad.pt 2 INESC TEC, Rua Dr. Roberto Frias, 4200-465 Porto, Portugal

Abstract. Mobile devices, as smartphones and tablets, have presented an exponential growth, being part of our everyday life, particularly considering children [1]. Their daily habits are undoubtedly influenced by technology and the applications they use can affect socialization and learning processes [2]. Specifically, games are the most popular type of applications and have the potential to change attitudes and behaviours. Emphasizing the importance of this area, we decided to create a serious game that stimulates the children’ responsibility for taking care of pets while they play, called “My Buddy”. In this paper, we present the development and assessment process of a 3D serious game, where the user is asked to interact with a pet and nurture it. The interface was developed based on the universal design philosophy, presenting itself attractive to children without disabilities, but also accessible to children with visual or motor disabilities. As such, we present a multimodal interface based on touch and speech commands. The game was tested in terms of usability, with a heuristic evaluation, and the results obtained highlight the potential of such interfaces. Keywords: 3D technology · Computer graphics · Human-computer interface · Universal design · Serious game · Speech interaction modality · Children education

1 Introduction Nowadays, many schools already resort to serious games with the aim of helping students to learn. Indeed, these are games that have another purpose besides entertainment and, as such, are used to promote learning or encourage behaviour changes, having widely proved themselves in various areas of expertise, namely education, healthcare, marketing and other businesses and industries. The power of serious games is that they are entertaining, engaging and immersive, combining learning strategies and game elements, as challenges and rewards, to teach specific matters. Naturally, teachers are aware of the massive potential of using serious games in classrooms: students end up feeling more motivated given that they are having fun, often without realizing that they are learning as well [3–6]. Indeed, serious games have a © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 457–466, 2022. https://doi.org/10.1007/978-3-030-96299-9_44

458

D. Carvalho et al.

positive effect in increasing students’ motivation, proving to be a valuable tool to assist in teaching [7]. Furthermore, these games are also an important resource to help children improve their cognitive and social levels [8–10]. Just by using them, children develop basic skills and competences [11] in different fields: (1) speech, as giving instructions, discussing and sharing opinions, giving directions and answering questions; (2) mathematics, by calculating the game scores; (3) reading, as they need to understand the game’s dialogue displayed on the screen; (4) and even sociability, when children play together or talk about their favorite games. Therefore, in addition to being a resource for educational issues, serious games are also used in schools to deal with problems, whether at the level of communication [12], spatial skills [13], problem solving exercises [14] or mathematical queries [15]. Likewise, there are also serious games specially designed for children with disabilities. However, their interfaces are not as accessible when compared to other technologies [16–18]. Despite the general interest in the development of adapted games, there is still a large number of people who are not able to play them due to having some sort of disability [19–21]. Undeniably, the accessibility of a product is crucial during the development phase of a tool. However, it is also seen as an obstacle, since it is necessary to integrate new interaction paradigms (e.g., “text-to-speech”, “speech-to-text”, voice processing models), produce hardware, or even adapt devices/controllers [22] to this end. The fact is that games must be accessible to as many people as possible, regardless of their abilities and/or skills, and these circumstances turn out to be discouraging when developing more accessible and inclusive games. While accessibility is a more common condition when it comes to serious games, in entertainment games this is still not a concern for most game designers [23]. In this context, when developing a game for children, it is necessary to present an accessible and aesthetically appealing design for all user profiles and not just one. Thus, we have to consider both the game mechanics and the graphical elements that captivate various player profiles, aiming for greater adhesion [24]. The game’s graphic elements must follow their mechanics without compromising the features and also its optimization. These elements cover not only the main aspects of any game but also secondary components, such as menus and buttons that must also be contextualized in order to avoid confusion and problems during interaction. Therefore, considering the mentioned strengths of serious games, our intention is to create an accessible mobile game, named “My Buddy”, that can be played by children with visual or motor disabilities, using voice commands as the interaction modality for input and output of information. Despite the target audience’ specificity, we intend to approach a universal design philosophy and, thus, create a game that is also appealing for children without disabilities. After presenting a brief outline of the related work in the field and identifying two popular mobile games for children, we believe it is important to also clarify our foundations for the game created and its features. Furthermore, we describe the methodology used for our evaluation and discuss its results. By reporting our observations and pointing out our conclusions, we believe that we can provide insights on how children may approach, perceive and interact with the serious game.

My Buddy: A 3D Game for Children Based on Voice Commands

459

2 Related Work In this section, we briefly lay out scientific background related to serious games and their importance in education. Additionally, in order to clarify our game design and our choices during the development process, we also showcase and analyse two different mobile games that are in accordance with our goals. 2.1 The Importance of Serious Games for Education The “gameplay” is the core of all games and, to summarize, it can be described as the decisions and actions a player has to take to overcome different challenges [31]. Naturally, the interactive aspect of the game is what makes it fun. Contrarily to oneway learning experiences (e.g. books or videos), in games the players can genuinely interact with the matters at hand, being allowed, and even encouraged, to experiment different outcomes of their actions. Indeed, the main goal of serious games is to rely on the “serious” scope of learning and not to simply entertain. Children learn intuitively by interacting and experimenting with the objects around them and this is the reason why serious games are so promising when it comes to education [32]. Indeed, serious games have the aptitude to successfully enhance people’s skills and proficiency, offering direct and continuous feedback and making them feel motivated [12]. This approach has shown to be effective, having a positive impact on students [13], and can be implemented to instruct large sectors of the population, even with different motivations and social backgrounds [14]. Consequently, these games are proving to be important assets in the learning process regarding different scopes: education for elders [30], simulations [15], training [16], health care [17], therapy [18], among others. This is an ongoing research area that shows a lot of potential, since serious games are incorporating features of our daily lives and presenting more realism than ever before [19]. In fact, they may play an important role regarding children with visual and motor disabilities due to their ability to engage users and allow them to explore the interface at their own pace. 2.2 Mobile Games for Children Having performed a brief exploration with the focus on available mobile games for children, two games were identified due to their similarity to our game’s main goal. They are mobile solutions, one in 3D and the other in 2D, and allow children to play and take care of a pet. However, these games only provide touch as the interaction modality, which means that people with motor or visual disabilities have difficulty playing them. The only voice feature that both provide is the repetition of what the user says with a high frequency. Nevertheless, these are games with a lot of popularity on the market. Next, we analyse these two games and their features, underlining their strengths and weaknesses. Firstly, Talking Tom [25] is a 3D mobile game for Android and iOS. The base version is free, but it displays advertising elements that can be removed by buying a premium version. The main character of this game is a talking cat that repeats everything the players says, also encouraging them to take care of the animal by performing daily activities, e.g.

460

D. Carvalho et al.

eating, going to the bathroom and sleeping. The main character of this game demonstrates happiness, hunger or sleepiness, and these afflictions are only resolved depending on the activities the players perform. Another extra feature are the so-called mini games, like “bubble shooter” or “slingshot”, among others, designed to test skill, reflexes and puzzle solving ability. Talking Tom also offers rewards as the player advances through the different levels (it offers between 9 to 999 different levels), unlocking new items and giving away coins, and provides the possibility to buy (using the digital coins of the game) new assets to embellish the set and the character, such as new combinations of fur, clothes or furniture. On the other hand, Pou [26] is a 2D mobile game, also created for Android and iOS. As the previous game, Pou reveals a pet and the player needs to address its basic needs and take care of the animal when it is sick. Likewise, the pet repeats what the player says and the game provides mini games. Moreover, it has several difficulty levels with the possibility for customization. On the contrary, the player can buy in-game coins with real money by browsing the online store for the game’s accessories. The analysis of the previous applications demonstrates several aspects to be taken into account when developing games for children. The first point relates to the difference between the applications in terms of design: the first, “Talking Tom”, was developed in 3D and the other, “Pou”, was developed in 2D. Curiously, although there are big differences in terms of design, the number of downloads for both games is practically the same, making us assume that even though the “Pou” game is visually less attractive, children also choose to play with it because it provides similar features as “Talking Tom”. Regarding the graphical elements, there are criteria that must be considered: the quality-optimization balance, as this directly affects the game storage size; and resources comsuption, since a game with high quality elements (large amount of polygons and complex textures) can consume a lot of resources on a mobile device or not even work on devices with less capacity. Moreover, considering a children’s game it is also necessary to take into account the particularities of the target audience: in most mobile games aimed at younger users, the static graphical elements (background themes, environments, menus) are generally very colourful and with elements allusive to the theme of the game [24]. On the other hand, regarding dynamic graphical elements, such as the characters, non-player characters (NPCs) and secondary components (e.g. menu animations), the design is simpler and cleaner, simplifying the movement and not requiring as much gameplay’s coordination as compared to other more complex games that target older users. Lastly, we emphasize that these games do not provide voice as either input nor output modality to help people with motor and visual difficulties to play. Whenever the players desire, they can exit the game by touching the corresponding button or simply say the word “exit” aloud. 2.3 Game Design The serious game designed for this study was developed with specific features, described next. When the player launches the game, the main screen displays its name and the primary menu with three options: game mode (nurture the pet, maze and obstacles), instructions and the exit button. By choosing the game mode for nurturing the pet, a

My Buddy: A 3D Game for Children Based on Voice Commands

461

new screen is presented (Fig. 1) where the user has the option to feed the pet and put it to sleep. These actions can be accomplished by using the corresponding buttons or by resorting to commands given by speech. Here, the player can instruct the game to undertake different actions, like “eat” and “sleep”, and the system will provide audio feedback.

Fig. 1. Game screen of My Buddy in “nurture the pet” mode

Secondly, if the player selects the maze option in the main menu, the system will present three available levels for the player to choose from. Figure 2 presents one of the mazes available to the player. As soon as the player finishes the maze, a congratulatory message is displayed, asking the player if he wants to continue to the next level or return to the main menu. Lastly, the game’s obstacles mode also offers three levels of difficulty. In this game mode, if the player touches any obstacle, he loses the game, but if they reach the end of the route, they win. When it finishes, whether they lose or win, a message is displayed, asking if the user wants to continue to the next level or return to the main menu (Fig. 3). During the maze and obstacle modes, the player can use speech commands to orient the animal, making it run, jump, turn right or left.

Fig. 2. Game screen of My Buddy in “maze” Fig. 3. Game screen of My Buddy in mode “obstacles” mode

462

D. Carvalho et al.

On the other hand, the player is also given the chance to understand the instructions, where it is demonstrated, with the help of images and voice, how the game works.

3 Game Evaluation We resorted to the use of heuristics in order to carry out the study. A heuristic evaluation is done by analysing an interface and getting an opinion about what is, or is not, correct about it. According to several paradigms, there are a number of rules that must be followed in order to accomplish the tests [28]. 3.1 Methodology For the game’s interface evaluation process, Nielson’s ten usability heuristics [29] will be used as a basis, divided into the following steps: evaluation preparation, short and objective sessions, consolidation of results, prioritization of problems encountered and, finally, the completion of a conclusive report on the interface in question. The ten heuristics that will be used in evaluating the application are: [H1] Visibility of system status: the design should always keep users informed about what is going on, through appropriate feedback within a reasonable amount of time. [H2] Match between system and the real world: the design should speak the users’ language and use words, phrases, and concepts familiar to the user, rather than internal jargon. [H3] User control and freedom: users often perform actions by mistake and need a clearly marked "emergency exit" to leave the unwanted action without having to go through an extended process. [H4] Consistency and standards: users should not have to wonder whether different words, situations, or actions mean the same thing. [H5] Error prevention: the design should prevent problems from occurring in the first place. Either eliminate error-prone conditions, or check for them and present users with a confirmation option before they commit to the action. [H6] Recognition rather than recall: minimize the user’s memory load by making elements, actions, and options visible, as the user should not have to remember information from one part of the interface to another. [H7] Flexibility and efficiency of use: flexible processes can be carried out in different ways, so that people can pick whichever method works for them. [H8] Aesthetic and minimalist design: interfaces should not contain information which is irrelevant or rarely needed. [H9] Help users recognize, diagnose, and recover from errors: error messages should be expressed in plain language (no error codes), precisely indicate the problem, and constructively suggest a solution. [H10] Help and documentation: help and documentation content should be easy to search and be focused on the user’s task, helping users to understand how to complete their tasks.

My Buddy: A 3D Game for Children Based on Voice Commands

463

3.2 Procedures Before proceeding with the tests, it is necessary to determine: (1) what is the general meaning that the user gives to the system/application he is going to analyse; (2) what is the profile of the users who will use the application; (3) what do they intend to do with the product; and (4) in which situations/scenarios will the application be used. It is important to note that these tests must be done by at least three reviewers and each one will provide their own evaluation that will later be grouped into one. 3.3 Results The evaluation process was as follows: (1) Assess the compliance of the game with the principles of the selected heuristics, recording which rules were violated and where; (2) Assess the seriousness of the problems found; (3) Generate a report with the findings and comments.

Fig. 4. Percentage of issues encountered for each heuristic

Figure 4 represents, for each heuristic, the proportion of the problems found, which are represented by the percentage of dark grey: the higher percentage of dark grey, the more problems were found. Indeed, it was possible to verify that 50% of the evaluated heuristics contained usability problems. More frequent problems were found in [H3], because on the game’s development phase emergency exits are scarce, that is, the way in which the player can exit the game and save the progress is limited. While in the mini game of obstacles and nurture the pet the exit option is not important, as it goes against the philosophy of the game, in the maze mini game this option can be necessary and practical. Problems were encountered in the heuristics [H5], [H6], [H8] and [H9], even if minor: they include the lack of prevention and error recovery by the user. Indeed, most of the feedback that the user has when he wrongfully performs an action is when he has already made the mistake, and not before he has done it. Also, there are still some features that require memorization, but in the mini games’ context, this situation is not necessarily adverse. Some issues were encountered regarding

464

D. Carvalho et al.

the aesthetic design of the menus and buttons. Undoubtedly, even though this is not the general focus of the game’s construction itself, it is important to have a coherent graphic line in order to avoid confusing the users. Table 1 outlines the overall issues reported. Based on the heuristic tests, we found that most of the problems found had to do with the fact that the application did not have an early and sophisticated error control. In this regard, changes will be made with the goal of achieving a system that prevents errors effectively and without affecting the user’s gameplay. In addition, a reconstruction of the game’s graphical interface will have to be performed. With all the usability conditions in mind, we believe that after fixing the reported issues, the game will be ready for user testing. Table 1. Description of the issues encountered for each heuristic Usability heuristics

Description of the issues

[H3] User control and freedom

Some situations do not provide “exits” for the user

[H5] Error prevention

Lack of events that signal possible errors, mainly during the “obstacles” mode

[H6] Recognition rather than recall

In the “obstacles” mode, there is a need for memorization

[H8] Aesthetic and minimalist design

The position of some buttons is irregular

[H9] Help users recognize, diagnose, and recover from errors

There is no error control in the “maze” mode when the player hits a wall

4 Conclusions In a general context, the accessibility of the game meets the basic requirements, especially considering the defined target audience. As far as usability is concerned, it is necessary to continue to develop a more robust and responsive error control. When it comes to gameplay, the mini games are intuitive, however it is necessary to carry out tests with users to detect and implement possible changes. As future work, it is our goal to add new animations to the game’s character, from playing, barking and strolling, as well as correcting the existing ones. Indeed, the current animations for sleeping, eating and changing the direction of the character’s progress could be improved. New features are also in store in the “nurture the pet” mode, from picking up objects and teaching tricks; as well as new levels of difficulty for mazes and trying to change the environment of each maze at each level, thus providing a greater diversity to the user; and new obstacles in the obstacle mini game. New mini games are also a possibility to consider as they provide new user experiences and differentiate gameplay. Moreover, it is of the utmost importance to initiate the testing phase with real users for feedback on the usability of the interface. In short, with the aforementioned changes

My Buddy: A 3D Game for Children Based on Voice Commands

465

and additions to the game, we expect to present a game that can provide entertainment to a greater number of possible users, with different accessibility profiles. Admittedly, another focus of our future work will be to ensure that the game is accessible to various user profiles. Acknowledgments. This work is a result of the project INOV@UTAD, POCI-01–0247-FEDER049337, financed by FEEI and supported by FEDER, through the Competitiveness and Internationalization Operational Program. Furthermore, we thank all people who directly or indirectly helped in this study, particularly to Mariana Montenegro, Rui Carvalho e João Teixeira.

References 1. Palazzi, C.E., Maggiorini, D.: From playgrounds to smartphones: Mobile evolution of a kids game. In: Consumer Communications and Networking Conference (CCNC), 2011 IEEE, pp. 182–186. IEEE (2011) 2. Hromek, R., Roffey, S.: Promoting social and emotional learning with games: "It’s fun and we learn things". Simul. Gaming (2009). https://doi.org/10.1177/1046878109333793 3. Gee, J.P.: What Video Games Have to Teach Us About Learning and Literacy. Palgrave Macmillan, New York (2003) 4. Jenkins, H.: Game Theory: How should We Teach Kids Newtonian Physics? Simple, Play Computer Games (2002) 5. Kebritchi, M., Hirumi, A.: Examining the pedagogical foundations of modern educational computer games. Comput. Educ. 51(4), 1729–1743 (2008) 6. Squire, K.: Changing the game: What happens when video games enter the classroom? Innov. J. Online Educ. 1(6) (2005) 7. Rosas, R., et al.: Beyond Nintendo: Design and assessment of educational video games for first and second grade students. Comput. Educ. 40(1), 71–94 (2003) 8. Csikszentmihaliy, M.: Flow: The Psychology of Optimal Experience. Harper & Press, New York (1990) 9. Provost, J.A.: Work, Play and Type: Achieving Balance in Your Life. Consulting Psychologist Press, Palo Alto, CA (1990) 10. RogoÔ¨Ä.B.: Aprendices del pensamiento. El desarrollo cognitivo en el contexto social. Barcelona (1993) 11. Griffiths, M.: The educational benefits of videogames. Educ. Health 20(3), 47–51 (2002) 12. Horn, E., Jones, H.A., Hamlett, C.: An investigation of the feasibility of a video game system for developing scanning and selection skills. J. Assoc. People Severe Handicaps 16, 108–115 (1991) 13. Masendorf, F.: Training of learning disabled children’s spatial abilities by computer games. Zeitschrift fur Padagogische Psychologie 7, 209–213 (1993) 14. Hollingsworth, M., Woodward, J.: Integrated learning: Explicit strategies and their role in problem solving instruction for students with learning disabilities. Except. Child. 59, 444–445 (1993) 15. Okolo, C.: The effect of computer-assisted instruction format and initial attitude on the arithmetic facts proficiency and continuing motivation of students with learning disabilities. Exceptionality 3, 195–211 (1992) 16. Bierre, K., Chetwynd, J., Ellis, B., Hinn, D.M., Ludi, S., Westin, T.: Game not over: Accessibility issues in video games. In: 11th International Conference on Human-Computer Interaction (HCII’05). Lawrence Erlbaum Associates, Inc (2005)

466

D. Carvalho et al.

17. Westin, T., Bierre, K., Gramenos, D., Hinn, M.: Advances in game accessibility from 2005 to 2010. In: Stephanidis, C. (ed.) UAHCI 2011. LNCS, vol. 6766, pp. 400–409. Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-642-21663-3_43 18. Yuan, B., Folmer, E., Harris, F.C.: Game accessibility: A survey. Univ. Access Inf. Soc. 10, 81–100 (2011) 19. Atkinson, M.T., Gucukoglu, S., Machin, C.H.C., Lawrence, A.E.: Making the mainstream accessible: redeÔ¨Åning the game. In: Sandbox ‘06: Proceedings of the 2006 ACM SIGGRAPH Symposium on Videogames, pp. 21–28. ACM, New York (2006) 20. Bierre, K., Ellis, B., Hinn, M., Ludi, S., Westin, T.: Whitepaper: Game Not Over: Accessibility Issues in Video Games (2005). http://www.igda.org/accessibility/hcii2005_gac.pdf 21. Grammenos, D.: Game over: learning by dying. In: CHI ‘08: Proceeding of the 26th Annual SIGCHI Conference on Human Factors in Computing Systems, pp. 1443–1452. ACM, New York (2008) 22. Bierre, K., et al.: Accessibility in Games: Motivations and Approaches (2004) 23. Grammenos, D., Savidis, A., Georgalis, Y., Stephanidis, C.: Access invaders: Developing a universally accessible action game. In: International Conference on Computers for Handicapped Persons, pp. 388–395. Springer, Berlin, Heidelberg (2006). https://doi.org/10.1007/ 11788713_58 24. Kultima, A.: Casual game design values. In: Proceedings of the 13th International Mindtrek Conference: Everyday Life in the Ubiquitous Era, pp. 58–65. ACM (2009) 25. Outfit7:https://play.google.com/store/apps/details?id=com.outfit7.mytalkingtomfree&hl= pt_PT (2017) 26. Zakeh: https://play.google.com/store/apps/details?id=me.pou.app&hl=pt_PT (2016) 27. Referência para o github—adicionar 28. Smith, S.L., Mosier, J.N.: Guidelines for Designing User Interface Software. Mitre Corporation, Bedford, MA (1986) 29. Nielsen, J., Molich, R.: Heuristic evaluation of user interfaces. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, pp. 249–256. ACM (1990) 30. Carvalho, D., Bessa, M., Peres, E., Magalhaes, L., Guedes, C., Oliveira, L.: Developing a multi-touch serious game to fight the digital divide: The Portuguese ATM: A pilot case study. In: Proceedings of the 7th Iberian Conference on Information Systems and Technologies (CISTI) (2012) 31. Katie, S., Eric, Z.: Rules of Play: Game Design Fundamentals, p. 3. The MIT Press, Cambridge, Massachusetts. ISBN 978–0–262–24045–1 (2004) 32. Craig, L., Lennart, N., Charlotte, S.: Dissecting play – investigating the cognitive and emotional motivations and affects of computer gameplay. In: Proceedings of CGAMES 08. University of Wolverhampton, Wolverhampton, UK. ISBN 978–0–9549016–6–0 (2008)

Demography of Machine Learning Education Within the K12 Kehinde Aruleba1(B) , Oluwaseun Alexander Dada2 , Ibomoiye Domor Mienye3 , and George Obaido4 1

2

Department of Information Technology, Walter Sisulu University, Mthatha, South Africa Institute for Molecular Medicine Finland (FIMM), University of Helsinki, Biomedicum 2U, 00290 Helsinki, Finland [email protected] 3 Department of Electrical and Electronic Engineering Science, University of Johannesburg, Johannesburg, South Africa [email protected] 4 Department of Computer Science and Engineering, University of California, San Diego, USA [email protected] Abstract. The rapid increase of artificial intelligence and machine learning tools and technologies around us has led to a rise in the daily interaction between humans and these technologies. Children nowadays are very likely to interact with these tools in different contexts, such as at home, recreation centres, and schools. While some children are already exposed to and working with these technologies, others are still far behind in the digital world. In this paper, we use the 2020 Stack Overflow Developer Surveys dataset to examine the demography of K12 students who are already using machine learning tools at school or their workplace. Over 55% of the respondents are younger than 24 years. The finding shows that there is still a significant gender gap in the IT field, with only 2% of 138 respondents identified as female. Also, with only four African countries represented in the dataset, Africa is still behind regarding machine learning in K12. Keywords: K12 education survey

1

· Artificial intelligence · Stack overflow

Introduction

Often referred to as the skill of the century, machine learning (ML) and artificial intelligence (AI) are now used to create intelligent systems and machines that perform tasks that would generally require intelligence. It has become part of our daily lives that it now feels quite normal when a device plays a game, recognises the voice, and detects faces. The importance and benefits of AI have made a growing number of companies use technology to enhance their decisionmaking, production, and marketing and evaluate their business models. These c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 467–474, 2022. https://doi.org/10.1007/978-3-030-96299-9_45

468

K. Aruleba et al.

benefits are not limited to the IT and computer science sectors alone. Other sectors directly related to our daily lives, such as automobiles (electric and selfdriving cars), government, banking and finance, and healthcare [6,9,16], now use AI algorithms to perform several activities. Beyond the broad applications of AI, ethics is another equally important aspect of AI-enabled systems. Despite the wide use of AI-enabled devices and technologies, there are potentials to spread inequality and misuse of these technologies. AI systems deployed in transportation, finance, and healthcare could be life-threatening to society if safety were not considered when implementing them. For example, deploying a facial recognition model, autonomous vehicle on real roads, deepfake videos, and AI-based grading systems are still controversial and debatable questions with much ongoing research. Hence, AI education is essential to inform engineers and scientists on the harm these systems can cause if they are not properly developed. People now need to study AI alongside society to understand how it works fully, and the responsible use because being digital literate is no longer sufficient in the AI era. Getting AI education and exposure to these technologies at a young age can also assist young engineers and scientists develop fair, transparent, and accountable systems. The increase in the automation of our daily activities has made AI education an essential enabler to future opportunities where success depends on creativity, intellect, and the right skills. Hence, to make the AI future workforce more inclusive and diverse, it is essential to have proper AI education. This paper focuses on K-12 students exposed to and using AI/ML algorithms in the classroom or workplace. Existing literature on ML in K12 has always focused on pedagogy and how ML is taught at the K12 level. To the best of the authors’ knowledge, no work in the literature has addressed the demography of K12 students who are already using ML in the classroom and at the workplace. Most of the AI education available today is designed for a more general audience which can only be understood by adults in a learning context due to the level of complexity and abstract in many AI models. However, in K-12 learning contexts, they require a unique style such as scaffolding learning and emphasis on student engagement. Therefore, there is a clear need to examine how AI curriculum and learning tools are designed to incorporate the unique need of K-12 and, most significantly, students from disadvantaged backgrounds. Understanding how ML models the world is a type of data literacy that can empower young learners to understand, explore, and question the data-driven systems they encounter daily (such as voice detection and face recognition) and ML systems [22]. ML is also an essential aspect of computational thinking and participation agenda at the K12 level [7]. The rest of this paper is organised as follow Sect. 2 discusses the background and related work. Section 3 presents the methodology of this study. Section 4 shows the paper discussion and findings, and Sect. 5 shows the conclusion of the paper.

Demography of Machine Learning Education Within the K12

2

469

Background and Related Work

The rapid increase of AI and ML tools and technologies around us has led to a rise in the daily interaction between humans and these technologies. Children nowadays are very likely to interact with these tools (YouTube, Siri, and Alexa) in different contexts such as at home, recreation centres, and schools. Notwithstanding this increase, most children still do not know when interacting with AI or how AI works. This becomes even worse for children from disadvantaged backgrounds whose parents do not even understand basic technology, with little or no internet access and limited education. All these have led to the growing AI literacy gap. Thus, education and AI researchers have called for better learning experiences that allow AI literacy to empower all learners to be critical users of these tools and technologies regardless of their demographic background. In addition, having AI in the curriculum of K-12 education worldwide will offer young students the opportunity to understand how AI can be used and see themselves as future developers of AI. Recently, different initiatives and research on introducing AI/ML to young learners have emerged [4,14]. Most of these initiatives are implemented in the Global North and focus on learning environments that enable learners to develop fundamental AI applications with block-based programming [5], teaching the principles of object recognition to high-school students, workshops, and training on how to use interactive data visualisation tools [2], and structured series courses that teach basic AI concepts and utilising existing AI tools [15]. A recent review in Marques et al. [8], discusses 30 educational initiatives aimed at AI/ML basics and neural networks in K12 education. The most popular of these initiatives and tools are IBM Watson-based (ML for kids), AIenhanced children’s robotics kits (e.g., Calypso for Cozmo6, AI-in-a-box7), Google’s teachable machine, and the Wolfram Alpha-based (ML for middle schoolers) [17]. For example, the AI4K12 project [18] for teaching AI from kindergarten, primary and high school is gaining much attention. Without understanding some ML principles, most services and apps kids are exposed to appear to them like magic. Video streaming, video recommendation even when they have never seen the video before, home assistants acting based on voice command and face recognition used to unlock phones are all complex for them to understand. Having a clear picture and understanding that all these technologies are smartly developed, which is in no way intelligent as humans are, is essential to describe these technologies. Gaining a comprehensive understanding and the required skills related to AI can be essential for children to succeed. Governments and agencies are slowly beginning to acknowledge this. In 2017, the House of Lords in the United Kingdom created a committee on AI. By 2018, the committee published a report that proposed that irrespective of the rapid development of AI, it is certain that AI will drive and impact future generations. Hence, the education system must address the needs of future generations through the following objectives: first, make children aware and be ready for a future with AI and prepare them for a possibly unpredictable job market. Second, assist young learners in

470

K. Aruleba et al.

understanding how the everyday technologies they interact with function; this can inspire a new generation of AI researchers and software engineers. Lastly, train future professionals to develop ethical and safe AI systems across different fields such as medicine, finance, and automobile. The prospects for young people are rapidly changing, and those with AI-related skills have the chance of having a competitive edge. A report examining the comparison of different Asian countries’ readiness and competitiveness in AI showed that the next generation of workforces must be equipped with relevant skills and experiences. These skills will enable them to manage and work with intelligent systems, critical thinking and problem solving, and interpret complex data. Many countries are currently including computational thinking in their K-12 curricula. However, little effort is being made to build better data literacy, and understanding of related concepts of the algorithmic process [10].

3

Methodology

The current study examines ML at K12. In particular, the study investigates the countries where ML is being taught at K12. Also, the demographics of the K12 students were discussed in this paper. To address this, the publicly available Stack Overflow Developer Surveys (SODS) 20201 dataset was used. This survey is the largest survey of diverse representation of programmers worldwide. The SODS 2020 survey dataset contains 65,111 data of respondents of different educational qualifications. However, this paper focuses on only the 138 respondents who chose secondary school (e.g., American high school, German Realschule, or Gymnasium) or primary/elementary school when asked for their education levels. Few variables of interest were extracted from the many variables in the dataset. Our data analysis was done using Python programming libraries such as pandas, collections, NumPy, and matplotlib. The analysis revolved around the different information present in the data set.

4

Results and Discussion

This study focuses on K12 students that have taken the SODS survey and are currently using ML algorithms either at their workplace or school. Over 55% of the respondents are younger than 24 years old, 19.6% are between 25 and 34 years, and the others are above 35 years. According to [12], the United States has historically been the leader and most contributor in AI-related outputs. This is because of the significant capital funding for AI and the ever-growing tech industry, aside from being the home to thousands of AI startups and companies. Figure 1 shows the location of each of the respondents. This was seen in the SODS dataset. The United States has the highest number of respondents (N = 20; 14.5%). This could be the considerable number of established AI 1

https://insights.stackoverflow.com/survey/2020#overview.

Demography of Machine Learning Education Within the K12

471

resources and models available for children in the country. India (N = 17; 12.3%), United Kingdom (N = 9; 6.5%), and Germany (N = 8; 5.7%) were the next to the United States, respectively.

Fig. 1. Country of K12 respondents

Many African countries such as Nigeria, South Africa, and Kenya are now developing their AI strategies. However, there are still significant challenges (such as lack of funding, infrastructures, and highly skilled labour), making Africa lag behind in the AI race. Addressing these issues has been the focus of many researchers in recent times [1,3]. From the 47 countries shown in Fig. 1, only 4 African countries with K12 students are exposed to ML based on the data presented in SODS. Nigeria leads other African nations with just three respondents, which is only 2.2% of the sample size. Then Kenya (N = 2; 1.4%), South Africa (N = 2; 1.4%) and Uganda (N = 1; 0.7%). This result indicates an immediate need to improve and strengthen the education sector with the continent. This will help to meet the new digital opportunities. As discussed in Sect. 2, exposing

472

K. Aruleba et al.

and investing in students at a young age (Primary school level), creating different AI initiatives, and mentorship programs will assist in building the next AI practitioner from the continent.

Fig. 2. Respondents population per gender

Figure 2 shows the respondent’s gender. From Fig. 2, 87% of the respondents are male, 2% are female, and 11% preferred not to say their gender. This is an indication of the gender gap in the AI field. According to the report in [20], there is a considerable gender gap in AI, but unsurprising given the scarcity of women in STEM and computer science fields. Furthermore, as few as 13.5% of the machine learning practitioners are female, while 18% of software developers and 21% of computer programmers identify as women [20]. A significant reason for this low presence of women in the STEM and computer field could be linked to gender inequality and the lack of early exposure to technology. Gender inequality is out of the scope of this paper, so we focus on education. Zhou and Xu [21] highlights that female instructor in post-secondary teaching have less experience and confidence in using computers for teaching. They often learn how

Demography of Machine Learning Education Within the K12

473

to use computers and other technology devices from others, while male instructors often learn from their own experience. In the survey in [21], more females than males stated their worries about lack of training opportunities and unstable software and hardware as a major challenge in computer use. These findings can be related to Spotts et al. [13], which explained that women are less confident of their experiences and skills in using computers than men. Women need adequate training and technology transfer because women’s adoption of engineering technologies depends on their skills and knowledge in operating, managing, and repairing specific machines [11]. Proper training and inclusion will increase the adoption of ICT by women at workplaces.

5

Conclusion

This study analysed the SODS data for the year 2020. The 2020 SODS contains 65,111 data of different IT professionals. However, for this study, we focused on K12 students’ data, i.e., the data analysed are only for respondents who have indicated that they use machine learning systems and are secondary or primary school students. The findings discussed in Sect. 4 showed that the United States of America has the highest number of young people who are already using machine learning models in their daily activities. Also, the gender gap in the IT field was seen in the dataset as only 2% of the respondents identified as female.

References 1. Abebe, R., et al.: Narratives and counternarratives on data sharing in Africa. In: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pp. 329–341 (2021) 2. Bishop, F., Zagermann, J., Pfeil, U., Sanderson, G., Reiterer, H., Hinrichs, U.: Construct-a-vis: Exploring the free-form visualization processes of children. IEEE Trans. Vis. Comput. Graph. 26(1), 451–460 (2019) 3. Cisse, M.: Look to Africa to advance artificial intelligence. Nature 562(7728), 461– 462 (2018) 4. Hitron, T., Wald, I., Erel, H., Zuckerman, O.: Introducing children to machine learning concepts through hands-on experience. In: Proceedings of the 17th ACM Conference on Interaction Design and Children, pp. 563–568 (2018) 5. Kong, S.C., et al.: Curriculum activities to foster primary school students’ computational practices in block-based programming environments. In: Conference Proceedings of International Conference on Computational Thinking Education, pp. 84–89 (2017) 6. LaRosa, E., Danks, D.: Impacts on trust of healthcare AI. In: Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, pp. 210–215 (2018) 7. Mariescu-Istodor, R., Jormanainen, I.: Machine learning for high school students. In: Proceedings of the 19th Koli Calling International Conference on Computing Education Research, pp. 1–9 (2019) 8. Marques, L.S., Gresse von Wangenheim, C., Hauck, J.C.: Teaching machine learning in school: A systematic mapping of the state of the art. Inf. Educ. 19(2), 283–321 (2020)

474

K. Aruleba et al.

9. Panesar, A.: Machine Learning and AI for Healthcare. Springer, New York (2019) 10. Pangrazio, L., Selwyn, N.: ‘personal data literacies’: A critical literacies approach to enhancing understandings of personal digital data. New Media Soc. 21(2), 419– 437 (2019) 11. Rola-Rubzen, M.F., Paris, T., Hawkins, J., Sapkota, B.: Improving gender participation in agricultural technology adoption in Asia: From rhetoric to practical action. Appl. Econ. Perspect. Policy 42(1), 113–125 (2020) 12. Savage, N.: The race to the top among the world’s leaders in artificial intelligence. Nature 588(7837), S102–S104 (2020) 13. Spotts, T.H., Bowman, M.A., Mertz, C.: Gender and use of instructional technologies: A study of university faculty. High. Educ. 34(4), 421–436 (1997) 14. Srikant, S., Aggarwal, V.: Introducing data science to school kids. In: Proceedings of the 2017 ACM SIGCSE Technical Symposium on Computer Science Education, pp. 561–566 (2017) 15. Steinbauer, G., Kandlhofer, M., Chklovski, T., Heintz, F., Koenig, S.: A differentiated discussion about AI education k-12. KI-K¨ unstliche Intelligenz, pp. 1–7 (2021) 16. Sun, T.Q., Medaglia, R.: Mapping the challenges of artificial intelligence in the public sector: Evidence from public healthcare. Gov. Inf. Q. 36(2), 368–383 (2019) 17. Tedre, M., et al.: Teaching machine learning in k-12 computing education: Potential and pitfalls. arXiv preprint arXiv:2106.11034 (2021) 18. Touretzky, D., Martin, F., Seehorn, D., Breazeal, C., Posner, T.: Special session: AI for k-12 guidelines initiative. In: Proceedings of the 50th ACM Technical Symposium on Computer Science Education, pp. 492–493 (2019) 19. Vartiainen, H., Tedre, M., Valtonen, T.: Learning machine learning with very young children: Who is teaching whom? Int. J. Child-Comput. Interact. 25, 100182 (2020) 20. VentureBeat: How more women in AI could change the world. https:// venturebeat.com/2018/04/15/how-more-women-in-ai-could-change-the-world/ (2018). Accessed 25 Sept 2021 21. Zhou, G., Xu, J.: Adoption of educational technology: How does gender matter? Int. J. Teach. Learn. High. Educ. 19(2), 140–153 (2007) 22. Zimmermann-Niefield, A., Turner, M., Murphy, B., Kane, S.K., Shapiro, R.B.: Youth learning machine learning through building models of athletic moves. In: Proceedings of the 18th ACM International Conference on Interaction Design and Children, pp. 121–132 (2019)

Educational Workflow Model for Effective and Quality Management of E-Learning Systems Design and Development: A Conceptual Framework Kingsley Okoye1,2(B) 1 Writing Lab, Institute for Future of Education, Office of the Vice President for Research

and Technology Transfer, Tecnologico de Monterrey, 64849 Monterrey, NL, Mexico [email protected] 2 School of Architecture Computing and Engineering, College of Arts Technologies and Innovation, University of East London, London, UK

Abstract. The use of learning management systems (LMS) in handling learning activities, e-contents, or curriculum that underlies the educational processes has shown to be effective, particularly in recent times. However, in settings where the resultant systems are not effectively built or deployed, this can result in delivery of low-quality educational services, or the developed systems not being able to integrate, support, and provide the expected outcome by and for the stakeholders. To this effect, this paper introduces a conceptual framework for learning called eLMS workflow model, that aims to improve educational learning systems and the identified challenges. It provides an automated administrative system for learning that shows to meet the needs of the stakeholders (students, staffs, universities, auxiliary services) by ensuring that the educational information and e-contents are easily located, effectively processed, and maintained whilst ensuring appropriate storage and use of the systems. Technically, the system (e-LMS) was designed to include enabling components such as automatic notification on request for learning activities, queries, archiving of request history, and tracking of users’ actions, that allows the users and stakeholders to effectively integrate different sets of information and services (curricula, projects, administrative, and operational activities). Thus, it supports infrastructural design, time, and cost to use of information technologies for educational/learning purposes. Keywords: IT in education · e-Learning · Process modelling · LMS · Educational data · Educational innovation · Higher education

1 Introduction The application/use of learning management systems (LMS) in Education have unwaveringly matured over the years. Pedagogically, information and communication technologies (ICT), allied to the LMS systems, are often developed and brought into the various organizations with the goal of helping to foster and solve the business needs [1, © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 475–490, 2022. https://doi.org/10.1007/978-3-030-96299-9_46

476

K. Okoye

2]. For instance, in context of education domain, LMS are used to support the various learning activities and curriculum design for the learners [1, 3, 4]. Thus, learning information or the e-contents can easily be accessed, effectively processed, and maintained whilst ensuring appropriate storage and use of the systems. However, in settings where the resultant systems are not effectively built or deployed (e.g., to help improve the efficacy of the educational services and operations, or learning processes for the teachers and students), this can result in delivery of low quality service or learning outcomes that do not meet the expected needs or performance [5–8]. In an effort to address the identified challenges, González et al. [9] have previously argued that the stakeholders (e.g. educational institutions) and learning service or tool providers (e.g. software developers, e-content designers) have to look out for not just a well-designed learning system (fitfor-purpose) that meets the operations or business needs, but also a system or educational model that aims to address the technical challenges that are or may be encountered during the different phases of its implementation [3, 10–16]. Typically, an automated learning management system should support interoperable machine-to-machine and human-tomachine interaction, system performance versus cost of implementing the solutions, and making the operations and/or functional aspects profitable [5, 17]. Gaining effectiveness in design and development of e-learning systems or platforms must involve a series of activities or process modelling steps, that are idiosyncratically carried out to ensure concrete implementation and management of the said processes or systems per se. The main focus must be to anticipate and predict as much as possible, the problems or technical challenges that may impact the progression of the (LMS) system design and development process. For this purpose, this study proposed the e-LMS framework for e-learning systems design that integrates the process of quality planning, control, and implementation of the sets of activities (workflows) or development phases, aimed to ensure that e-learning projects or systems are designed and completed successfully, by boring in mind the estimated project time and objectives (e.g., achieving the expected outcome by the project sponsors or educators irrespective of the risk or constraints that may inadvertently emerge or encountered during this process) [18]. Moreover, in order to achieve the technical and/or pedagogical objectives of any LMS system; it is expected that planning and procurement of the resources must be specific and realistic, in addition to being measurable and achievable. Thus, such kinds of (educational) projects often allied to the SMART methodology or goals (Specific, Measurable, Achievable, Relevant, and Time-bound) [19, 20] have shown to be efficient and fit for purposes for which it is intended or serves, particularly as it concerns attaining quality in the educational system and learning outcomes [6, 17]. Kerzner [21] note that there has been a revolution in the quality of e-learning systems. The author [21] observed that improvements have occurred not only in aspects of quality and usability of the developed systems such as LMS [5, 9], but also in the many operations, maintenance, or sustenance of the resultant systems and models [6]. Specifically, Kerzner [21] defined quality management as “the process of ensuring that the end results of the developed systems, e.g. LMS, will meet the expectation of the users”. Moreover, the works of [5, 6, 8, 22, 23] sees “quality” as the ability of any information system to fulfill its purpose, for example, by considering the nature, size, data, intricacy, and users of the systems. In another depiction, Okoye et al. [24, 25] sees quality

Educational Workflow Model for Effective and Quality Management

477

as the ability of the collection of tasks involved in the different stages of the systems’ design and implementation to fulfill requirements by the intended users (educational institutions, learners) by taking into account the users’ profiles, goals or expectations (e.g., learning behaviours, styles, accessibility, and usability factors, etc.) [24, 25]. By definition, quality management of e-learning systems must allow for SMART control of the several procedures and/or plans that are set out in the learning project management and deployment. Along these lines, this study utilized the Plan-Do-Check-Act (PDCA) approach [26] to propose a workflow model for e-learning project management (e-LMS) implemented by identifying at first, the quality objectives of the e-learning systems, and then breaks those components down into manageable units (workflows) to provide an effective and controlled flow of activities or tasks that are needed to achieve the educational objectives. The PDCA cycle have shown to provide a higher level of quality assurance or procedures for measuring, for instance, whether the e-LMS workflow model is adequate or relevant, towards supporting ample implementation, interoperability, and usability of the developed system. Thus, the e-LMS model, which was grounded on the PDCA method, acts as a key performance evaluation/indicator used in measuring the learning process workflow model and outcome, by critically analyzing the impact of the resultant system against the predefined SMART metrics.

2 Background Information Since what most users of e-learning systems or platforms do is to view, read, and interact with the contents that are embodied in the said educational systems; there has been growth in demand for quality and easy-to-use systems that meets the purpose and goals of the intended users (educational institutions, instructors, learners) [11, 22, 24]. While Fresen and Hendriks [27] note that the design and provision of e-learning systems ought to focus on the quality of support, or yet still, a clear understanding of the different procedures being subsidized (defined) by the development or management team [28, 29]. It is expected that LMS system’ developers must make use of the available skills and resources to figure out plans/metrics towards achieving the expected quality of the resultant systems [5, 6, 17]. Besides, quality is becoming a significant issue or discourse in e-Learning systems design and development [5, 30]. There is need for Educators or LMS developers to consider the standards (procedures) that are critical towards ensuring quality of the e-learning systems, for instance, by setting the “accessibility” and “usability” of the resultant tools/platforms as one of the quality objectives [5, 6, 25, 30]. They also need to make clear the context of the plans/design by defining SMART metrics [19, 20] needed to measure the quality of the systems’ development and implementation in real-time. Hirata [30] notes that there is a vast number of guidelines and principles in connection with quality of the information/educational systems, and that those guidelines are not only useful for setting objectives, evaluation, or improvement of the e-learning systems/projects, but also in determining suitable requirements and actions plans to achieving the set-out goals [31]. According to Hirata [30], it is indispensable to explicate or make clear the quality metrics or objectives towards achieving a maximum understanding of the (LMS) implementation

478

K. Okoye

process, as well as, how to use those standards for effective management of the systems in general. Quality management of e-learning systems as specified by International Standard Organisation (ISO) in part of ISO/IEC19796–1 [32], stated that when developing elearning systems or projects, that Quality Management Plan (QMP) or Control Measures (CM) must be put in place as requirements/part of the project. Moreover, Quality Assurance (QA) [6] undertakings are also to be carried out not only at specific stages or points of the project lifecycle, for instance, to ensure the Quality and Control Process (QCP), but also at every stage of the development and implementation process. This is where the PDCA method [26], as described in this paper, is effective and paramount in ensuring that the resultant learning systems and development lifecycle are quality-based. Indeed, the PDCA method is useful for (quality) assurance or management, not only by meeting the project objectives, but also in guaranteeing that the underlying workflows are effectively implemented. Having considered the quality objectives or metrics in line with the ISO standards or guidelines, e.g. e-LMS framework described in this study, the LMS system is represented to cover the users, system’ owners, and the development processes in general. Thus, a fit-for-purpose system are eventually implemented [25, 33]. Likewise, on the other hand, Thom et al. [34] defined logical procedures (relations) between process activities, such as the educational process in this paper, as workflow activity patterns (WAPS) [35, 36]. WAPS are common structures involving the interaction between individual entities (process instances) and the control-flow constructs used to model the semantics of the activities being performed (e.g., LMS) [35, 36]. Technically, WAPS (workflow management systems) assume that a process can be divided into small, unitary actions called activities or tasks [37] (see: Figs. 1, 2a, and 2b). To perform or execute a given process, one must perform the sets (or perhaps a subset) of the activities that comprises it [12, 24]. Henceforth, an activity is an action that represents as a semantic unit at some level, which can be thought of as a function that modifies the state of the process in question (e.g., the learning process) in terms of the semantics of the patterns (or tasks), and can be designed, implemented, or computed manually, semi-automatically, or automatically [12].

3 Method Description and Theoretical Framework The development and application of methods that enable (good) quality management of e-learning systems/projects dates back to the first half of the twentieth century [38]. Ever since then, there have been several research and pedagogical discourses on what methods are best or adequate for ensuring quality management of e-learning projects and design [8, 11, 22, 26, 39, 40]. For some of the existing methods/approaches in the current literature, an in-depth comparison of their strengths and weakness have also been reviewed/debated [6, 22, 30, 32, 41–45]. The second milestone was during the early twenty-first century, when studies like that of Clegg et al. [46] put forward ethics, tools, and methods, envisioned useful towards ensuring the quality of e-learning management systems. According to the authors [46], learning management systems or applications should not be seen as alternatives to total

Educational Workflow Model for Effective and Quality Management

479

process management (e.g., the learning process in itself) but rather as a collection of individual activities and concepts that supports the overall quality objectives of the system. While, Salah et al. [47] noted that electronic learning management systems have gained more attention over the decades, as it tends to provide basic means for controlling and improving the quality of a large number of (e-learning) tools or platforms per se. Adina-Petru¸ta et al. [48] have also looked into the way Six Sigma [39, 43] is applied in higher education, and on integrating Six Sigma as one of the quality management strategies; namely the model ISO 9000, for the development and continuous improvement of universities [48]. On the other hand, Geraldi et al. [49] stated that uncertainty and dynamics of e-learning systems appear to challenge the ideology of quality, especially as it concerns recurring operations and called for project-tailored solutions (fit-for-purpose) [50]. Along those lines, the method of this paper aimed not just to adopt such type of quality ideology in the design of the proposed e-LMS workflow model, but also builds upon emerging propositions or methods that allow for effective application of the method (e-LMS workflow model) in real-time settings. Thus, the PDCA-based method which is described in detail in the next section of this paper. 3.1 Design Approach This study adopts the Plan-Do-Check-Act (PDCA) approach [26, 51] to propose an eLMS method that shows to be useful for effective management and implementation of e-learning systems. The PDCA was first introduced by Deming (1986) in 1950 [51], who as of then referred to the method as the “Deming wheel” [52]. His approach, allied to the “theory of management”, was based on long-term commitment to new learning and new philosophy that are purportedly required for transforming management styles or projects [53]. Ever since then, many tools and methods have been proposed that adopts the ideology, for instance, as utilized in this study to model the e-LMS framework. Likewise, studies like Kirsch et al. [54] notes that for an effective design and evaluation of projects, e.g., the e-learning management systems, that the developers must follow a primary set(s) of principles or cycle, which consists of planning, implementation, control, and iterative feedback. They [54] found the PDCA cycle as a state-of-the-art approach that is quite often used in almost every process modelling or analysis project. According to them [54], the PDCA is particularly used to ensure quality control (quality assurance) procedures or measures that are suitable for modelling of systems, such as the e-LMS, as well as the quality of the outcomes. Although, Geraldi et al. [49] noted that for effective implementation of the PDCA approach, that the stakeholders (project managers, developers) may also experience difficulties in choosing an appropriate framework that ensures the quality and evaluation of the project. To this end, the work done in this paper proposes the e-LMS workflow model to illustrate a more effective way towards the design and implementation of e-learning systems, by bearing in mind the quality of the underlying administrative processes, and the educational/learning purposes or pedagogical intentions. Technically, PDCA-based methods must focus on classifying the design and development process, e.g. the e-LMS workflow model in this paper, into four different phases [55]:

480

• • • •

K. Okoye

Plan – planning of the quality management procedures Do – implementation of the procedures Check – conformance checking of the quality metrics or criteria Act – deployment of the method or system.

Plan (Phase 1): is the first step to the e-LMS workflow model described in this study context. The planning at this stage is used to determine the systems requirements, which includes: • • • • • •

What the objectives of the educational system or e-learning project is? Why the objective was set? Where the objective and plan will be used? When the objective and plan will be used? Who will be responsible for what? and How the objectives will be measured?

This phase (Plan) helps the system’ (LMS) developers/team to identify the problems, and potential bottlenecks. Breaks those down into manageable and smaller components, and then, outline the metrics or measures that will be used to solve the problems. Do (Phase 2): is used to control and monitor the project as it progresses. At this stage, both the managers and the implementation team are to ascertain (determine) whether the requirements meets or are capable of solving the objectives as set out in the plan (phase 1). The “Do” phase allows the project team to evaluate whether the quality control mechanisms (or prototype) put in place will in real-time make a development impact or not, without running the risk of affecting the progress of the project and/or delivery time. Check (Phase 3): covers the systems’ quality assurance by measuring and determining whether the performance or execution criteria being set out in Phases 1 and 2 is appropriate and relevant to accomplishing the system’ objectives. The “Check” is carried out by evaluating the identified plans (phase 1) against the procured requirements (phase 2) on a regular basis, to ensure that the LMS system per se, will meet the expected goal. Thus, this phase (Check) acts as a key performance evaluation indicator used in measuring the outcome of the “Do” and “Plan” phase by critically analyzing their impact against some pre-defined quality control metrics. Act (Phase 4): ensures or guarantees the entire project development and implementation process (and take corrective actions if need be). Practically, the “Act” transforms the slated steps in Phases 1 to 3 into an integrated and/or executed process. In summary, the Plan-Do-Check-Act (PDCA) helps in ensuring that quality and consistency are maintained, and adhered to at all times, throughout the entire process of developing and implementation of the e-LMS system. In fact, the approach (PDCA) is used to effectively initiate and provide answers to rapidly evolving e-Learning design problems or fixing questions, such as; What is this? Why do it? When do I use it? Who does it? How do we do it? etc. [56]. Fundamentally, the above questions can be explicitly answered and implemented to address e-Learning systems development issues, as follows:

Educational Workflow Model for Effective and Quality Management

• • • • • • •

481

what are the quality objectives of the e-learning system or project? how does it meet the need and requirements by the stakeholders/sponsors/users? who are the intended users? and what are their objectives in using the system? who will be responsible for ensuring that those quality objectives are met? when do they implement the quality plans or procedures? what is the conceptual model or framework of the e-learning application? and how will the management team measure whether those quality criteria are met?

This study aimed to address the identified issues or learning design problems by proposing the e-LMS framework, a workflow model that can be used to build an effective or quality e-learning systems [5, 8]. Exclusively, the method (e-LMS) can help e-learning project managers/developers to determine whether unabridged quality objectives of the resultant systems are met. The e-LMS framework puts into consideration the (quality) requirement by the users when drawing the quality management plan and objectives. Quality objectives and planning of any e-learning system must put into consideration what the users, stakeholders or sponsors expect at the end of the project, and must be designed and implemented to meet the requirements by the said users/stakeholders through a quality adherence method, such as the PDCA approach. In turn, this ensures that quality is being maintained and adhered to at all times throughout the entire process of the e-learning system design, development, and deployment. 3.2 The e-LMS Workflow Model In Fig. 1, the study introduces the architecture of the e-LMS design framework, and how it is used to enable an effective workflow model for ample implementation of e-learning process or curriculum development. It shows the main components and different stages (tasks) for its implementation (Fig. 1), the resultant workflows (Fig. 2a and 2b), and its application towards an automated administrative learning system and management for the educators or institutions. This was done specifically to show how the different components of the e-LMS model integrates, and is capable of ensuring the quality of e-learning system design and development. Figure 1 explains the proposed framework with a description of the different tasks and subtasks that constitutes the process. Figures 2a and 2b represents the e-LMS workflow model for implementation of the framework. As shown in the flowchart (Fig. 1), the process of designing the “system prototype” involves a number of “quality assurance (QA)” checks that are supposedly required to be passed or completed, before proceeding to next steps of the project, such as the “procurement of resources and development of the system”. Subsequently, the functional components of the system are tested before deployment and handover, as shown in Fig. 1.

482

K. Okoye

Fig. 1. Architecture of the e-LMS framework

Accordingly, the following figures (Fig. 2a and 2b) represents a flow diagram of the different components (workflow breakdown) that constitutes the e-LMS system as defined in Fig. 1. Some of the key components or tasks desegregated in order to realize an effective and quality use of the LMS system includes: • Identification of the main tools and instructional equipment (hardware, software) including computer systems that rather adapt to low power supplies.

Educational Workflow Model for Effective and Quality Management

483

• Installation of power backups (e.g., inverters, UPS) to ensure continuous and steady power supply. • Cabling of wires using standard installation guidelines and tools. • Install hardware components making sure they are properly mounted to reduce the risk of damage. • Design of database and/or queries using database management applications that run on any operating system (OS) to ensure ease of use and stability of the system. • Create/use softwares that can run on any platform or over a network. • Provide security measures to cater for or check potentiality of wrong and/or misplacement of data.

Fig. 2a. Implementation of the e-LMS workflow model and main components.

484

K. Okoye

• Provide documentation, user guide and training of staffs to ensure appropriate use and implementation of the system, and then • Deploy an automated administrative and management system for learning that reduces the operational cost, time to and/or process development, and management or elearning platforms to a high degree. As shown in Fig. 2a and 2b, a number of quality “Checks” are implemented in the process after each “Plan” or “Do” tasks before proceeding to the next stages of the system development or “Act” based on the PDCA approach. For instance, in Fig. 2a, we can see from “identifying the system requirement and resources”, to the “system design and modeling”, that quality checks was implemented before the infrastructural (e.g., hardware and software) implementation phases. Whereas, in Fig. 2b, the “systems’ integration and components testing” were done before the final “deployment of the system”.

Fig. 2b. Implementation of the e-LMS workflow model and main components.

4 Discussion Evaluation of the functional components of e-learning systems, such as the e-LMS framework, particularly as it concerns application of the sets of activities or workflows in real-time, is critical in determining the quality of the developed systems. The quality

Educational Workflow Model for Effective and Quality Management

485

of performance of such system is determined by the rate at which the different components can effectively satisfy the business needs. While Nath and Singh [57] note that existing methods/studies do not stipulate specific (or necessary) criteria for evaluating the quality of the systems. Clearly, the quality performance (e.g., the e-LMS workflow model) and its real-time application is believed to be achieved through the SMART lens or metrics, as well as the manageable and reassuring phases of the PDCA. Thus, the individual tasks breakdown (workflows) provides an effective structure (increased agility) for design and development of the systems, and the fact that it is able to meet the objectives/requirements by the users/stakeholders. PDCA approach provides developers with consistency (effective management procedure), and the capacity to swiftly respond to not just changes that may neededly be required during the system design and development process, but as well as, is able to satisfy or meet the stakeholders’ needs. Which according to Nath and Singh [57], is the biggest outcome or benefits of using such method (PDCA) in carrying IT projects or systems development. Furthermore, other factors the study considered essential in the evaluation of the eLMS system is interoperability. Whereas the system is designed to support the integration of different (functional) pedagogical or technological components. For instance, data can be shared or exchanged over the network, and is capable of running on any platform. It enables data mining or information retrieval technologies to be used. Data entry or information retrieval are relayed to the central database (see: Figs. 1 and 2), e.g., new data entries are concretely correlated with the historic or existing data, and can be relayed out again (retrieved) via the interfaces and/or queries modules. Indeed, a major consideration in architecting the model (e-LMS) was to show the benefit of integrating an automated administrative and management system for learning, whilst keeping changes to the existing system or environment, as few and efficient as possible. Another important factor the study considered in its design and evaluation, is the security aspects of the system. The work showed the need to ensure that the exchange of data and information across the platform is being authenticated and authorized. This was done in order to maintain a good level of data/information integrity and confidentiality. For instance, encryption of data at the requestor’s side and decryption of information at the providers’ side. Pimenidis and Georgiadis [58] noted that providing a securitydriven application or system is an essential criterion for achieving trust and managing the loyalty of intended users of any system. In summary, amongst the many benefits of implementing the e-LMS workflow model (Figs. 2a and 2b) includes: an automated administrative system for learning that enable the students, staffs, or service managers to effectively create information and didactical tools/technologies in support of the curricula, educational process, and the business operations in the wider spectrum. In turn, is capable of commercially making the institutions per se to be competitive amongst the others. Also, the e-LMS system supports an automatic notification on requests, status, data archiving or history, and tracking of user actions [10, 11, 24]. Thus, allows for an effective database structure or record management. Moreover, information and data security are guaranteed due to the aforenoted features. The main implications or benefit of implementing the e-LMS system in practice, is that it reduces operational expense and time that is usually associated with the batch

486

K. Okoye

processing learning systems, by providing an effective and automated system for administration and learning for the learners in the different institutional settings or education at large [59].

5 Conclusion The evolvement and integration of learning management systems and technology into the educational systems, have enabled many institutions to continue to gain efficiency and effectiveness in its pedagogical and administrative processes. The major benefits of the e-LMS framework and workflow model described in this study; is its pedagogical capacity to support an automated system for learning for the higher institutions, and its administrative capacity to minimize the time and cost of managing the educational processes on the whole. This paper identified some of the key functional components of e-learning systems, and puts forward a method which can help educators to integrate quality in its process or curricula design. The study illustrated and discussed the practical application and implication of the e-LMS framework in terms of the quality of performance, interoperability, and security. It can be said that method (e-LMS) is an idiosyncratically and technological response for the educational institutions in achieving effectiveness in the demand or need for an automated/technology-enhanced model for learning management, e-content development, and curricula. It is noteworthy to mention that, although the work has identified and introduced a workflow model for quality control or implementation of the e-learning systems, its application particularly with regards to the different contexts or educational settings may vary, and could be limited to some extent. This is owing to the fact that the work does not consider the impact of the method in terms of the stakeholders’ varying settings, experiences, or transformational adoption of such systems. Future works could look into those aspects by taking into account the institutions’ perceptions or user-centric experiences, in order to provide a more widespread authentication of the work already done in this study. Acknowledgment. The author would like to acknowledge the financial support of Writing Lab, Institute for Future of Education, Tecnologico de Monterrey, in the publication of this work.

References 1. Chiu, M.-S.: Exploring models for increasing the effects of school information and communication technology use on learning outcomes through outside-school use and socioeconomic status mediation: the Ecological Techno-Process. Educ. Tech. Res. Dev. 68(1), 413–436 (2019). https://doi.org/10.1007/s11423-019-09707-x 2. Piedade, M.B., Santos, M.Y.: Business intelligence in higher education: enhancing the teaching-learning process with a SRM system. In: 5th Iberian Conference on Information Systems and Technologies, pp. 1–5 (2010) 3. Okoye, K., Nganji, J.T., Hosseini, S.: Learning analytics: the role of information technology for educational process innovation. In: Advances in Intelligent Systems and Computing (AISC). Proceedings of IBICA-WICT 2019, vol. 1180, pp. 272–284, Springer (2021). https:// doi.org/10.1007/978-3-030-49339-4_28

Educational Workflow Model for Effective and Quality Management

487

4. Urmi, S.S., Taher, K.A.: Integrating ICT in teaching and learning at university level focusing education 4.0. In: 2021 International Conference on Information and Communication Technology for Sustainable Development, ICICT4SD 2021, pp. 300–304 (2021) 5. Medne, A., Lapina, I., Zeps, A.: Sustainability of a university’s quality system: adaptation of the EFQM excellence model. Int. J. Qual. Serv. Sci. 12(1), 29–43 (2020) 6. Mourad, M.: Quality assurance as a driver of information management strategy: Stakeholders’ perspectives in higher education. J. Enterp. Inf. Manage. 30(5), 779–794 (2017) 7. Gamede, V.W.: Cultural implications for learners’ effectiveness as governors of schools in rural South Africa. SA J. Educ. 40(3) (2020). https://doi.org/10.15700/saje.v40n3a1655 8. Supriadi, O., Mutrofin, M.: Management capability in a structural modelling of the quality of economics and accounting education in Indonesia. SA J. Educ. 40(1) (2020). https://doi.org/ 10.15700/saje.v40n1a1658 9. González, R.M., Martin, M.V., Arteaga, J.M., Álvarez Rodríguez, F.J., Ochoa Ortíz Zezzatti, C.A.: Web service-security specification based on usability criteria and pattern approach. J. Comput. 4(8), 705–712 (2009). https://doi.org/10.4304/JCP.4.8.705-712 10. Schultz, M.: Enriching process models for business process compliance checking in ERP environments. In: vom Brocke, J., Hekkala, R., Ram, S., Rossi, M. (eds.) DESRIST 2013. LNCS, vol. 7939, pp. 120–135. Springer, Heidelberg (2013). https://doi.org/10.1007/978-3642-38827-9_9 11. Alles, M., Brennan, G., Kogan, A., Vasarhelyi, M.A.: Continuous monitoring of business process controls: a pilot implementation of a continuous auditing system at Siemens. Int. J. Account. Inf. Syst. 7(2), 137–161 (2006). https://doi.org/10.1016/J.ACCINF.2005.10.004 12. Okoye, K., Islam, S., Naeem, U., Sharif, M.S., Sharif MhD, S.: Semantic-based process mining technique for annotation and modelling of domain processes. Int. J. Innov. Comput. Inf. Control 16(3), 899–921 (2020) 13. Firmansyah, E., Herdiana, D., Yuniarto, D.: Examining readiness of e-Learning implementation using information system readiness impact model. In: 2020 8th International Conference on Cyber and IT Service Management (CITSM) (2020) 14. Widger, L., Denny, M., Benke, M., Pajnkihar, M., Bruen, C., Madden, C.: The strategic implementation and augmentation of Technology Enhanced Learning (TEL) In Third Level Education: a critical lens. In: 10th International Technology, Education and Development Conference. (INTED2016), vol. 1, pp. 1289–1297 (2016) 15. Herodotou, C., Rienties, B., Boroowa, A., Zdrahal, Z., Hlosta, M.: A large-scale implementation of predictive learning analytics in higher education: the teachers’ role and perspective. Educ. Technol. Res. Dev. 67(5), 1273–1306 (2019). https://doi.org/10.1007/s11423-019-096 85-0 16. Bezanilla, M.J., Olalla, A.M.G., Castro, J.P., Ruiz, M.P.: A model for the evaluation of competence-based learning implementation in higher education institutions: criteria and indicators. Tuning J. High. Educ. 6(2), 127–174 (2019). https://doi.org/10.18543/tjhe-6(2)-201 9pp127-174 17. Pawlowski, J.M.: The quality adaptation model: adaptation and adoption of the quality standard ISO/IEC 19796-1 for learning, education, and training. Educ. Technol. Soc. 10(2), 3–16 (2007) 18. Miller, B.: The Purpose of Project Management and Setting Objectives. 21 Ways to Excel at Project Management (2018). https://www.projectsmart.co.uk/purpose-of-project-manage ment-and-setting-objectives.php. Accessed on 04 Oct 2021 19. CFI: SMART Goal - Definition, Guide, and Importance of Goal Setting, Cooporate Finance Institute (2021). https://corporatefinanceinstitute.com/resources/knowledge/other/ smart-goal/

488

K. Okoye

20. Bu, L., Chen, C.H., Ng, K.K.H., Zheng, P., Dong, G., Liu, H.: A user-centric design approach for smart product-service systems using virtual reality: a case study. J. Clean. Prod. 280, 124413 (2021) 21. Kerzner, H.: Project Management: a Systems Approach to Planning, Scheduling, and Controlling, vol. 1877. Wiley, New York (2017) 22. Mansouri, S., Eftekhar, F., Heidarnia, S.: The application of quality management in e-learning, by QFD technique and based on customers’ needs (A case study in an Iranian University). In: 3rd International Conference on eLearning and eTeaching, ICeLeT 2012, pp. 45–52 (2012). https://doi.org/10.1109/ICELET.2012.6333364 23. Deepwell, F.: Embedding quality in e-learning implementation through evaluation, educational technology & society. Educ. Technol. Soc. 10(2), 34–43 (2007) 24. Okoye, K., Tawil, A.-R.H., Naeem, U., Bashroush, R., Lamine, E.: A semantic rule-based approach supported by process mining for personalised adaptive learning. Procedia Comput. Sci. 37, 203–210 (2014). https://doi.org/10.1016/j.procs.2014.08.031 25. Okoye, K., Jahankhani, H., Tawil, A.-R.H.: Accessibility of dynamic web applications with emphasis on visually impaired users. J. Eng. 2014(9), 531–537 (2014) 26. Riis, J.O., Achenbach, M., Israelsen, P., Hansen, P.K., Johansen, J., Deuse, J.: Dealing with complex and ill-structured problems: results of a Plan-Do-Check-Act experiment in a business engineering semester. Eur. J. Eng. Educ. 42(4), 396–412 (2016). https://doi.org/10.1080/030 43797.2016.1189881 27. Fresen, J.W., Hendrikz, J.: Designing to promote access, quality, and student support in an advanced certificate programme for rural teachers in south Africa. Int. Rev. Res. Open Distance Learn. 10(4) (2009). https://doi.org/10.19173/irrodl.v10i4.631 28. Boyd, L.G., Fresen, J.W.: Quality promotion and capacity development could they come to the aid of weary South African academics? (2004). http://hdl.handle.net/2263/7509. Accessed on Oct 04 2021 29. Fresen, J.W.: Quality assurance practice in online (web-supported) learning in higher education: an exploratory study. Thesis on Quality Assurance in Higher Education (2005) 30. Hirata, K.: Information model for quality management methods in e-Learning. In: Sixth International Conference on Advanced Learning Technologies, ICALT 2006, vol. 2006, pp. 1147–1148 (2006) 31. UNESCO: Global Education Coalition (2021). https://en.unesco.org/covid19/educationres ponse/globalcoalition. Accessed on 17 Aug 2021 32. ISO: Information technology-Learning, education and training-quality management, assurance and metrics - Part 1: general approach technologies (2005) https://webstore.iec.ch/pre view/info_isoiec19796-1%7Bed1.0%7Den.pdf. Accessed on 04 Oct 2021 33. IMS: IMS Global Learning Consortium Announces Agreement with ISO/IEC on Emerging Metadata Standards for Learning Accessibility and Personalization | IMS Global Learning Consortium (2008). http://www.imsglobal.org/pressreleases/pr080306.html. Accessed on 04 Oct 2021 34. Thom, L.H., Reichert, M., Iochpe, C.: Activity patterns in process-aware information systems: basic concepts and empirical evidence. Int. J. Bus. Proc. Integr. Manage. 4(2), 93 (2009) 35. van der Aalst, W.M.P.: Process Mining: Data Science in Action. Springer, Berlin (2016). https://doi.org/10.1007/978-3-662-49851-4_1 36. van der Aalst, W.M.P.: Process Mining: Discovery, Conformance and Enhancement of Business Processes. Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-642-19345-3 37. Ferreira, D.R., Thom, L.H.: A semantic approach to the discovery of workflow activity patterns in event logs. Int. J. Bus. Proc. Integr. Manage. 6(1), 4–17 (2012) 38. Shewhart, W.A.: Statistical method from the viewpoint of quality control. Nature 146(3692), 150 (1940). https://doi.org/10.1038/146150e0

Educational Workflow Model for Effective and Quality Management

489

39. Nonthaleerak, P., Hendry, L.C.: Six sigma: literature review and key future research areas. Int. J. Six Sigma Compet. Adv. 2(2), 105–161 (2006) 40. Bhuiyan, N., Baghel, A.: An overview of continuous improvement: from the past to the present. Manag. Decis. 43(5), 761–771 (2005). https://doi.org/10.1108/00251740510597761 41. Zashchitina, E.K., Pavlov, P.V.: Higher education today: mass or individual approach. In: 2019 IEEE International Conference Quality Management, Transport and Information Security, Information Technologies IT and QM and IS 2019, pp. 653–656 (2019). https://doi.org/10. 1109/ITQMIS.2019.8928325 42. Dahlgaard, J.J., Dahlgaard-Park, S.M.: Lean production, six sigma quality, TQM and company culture. TQM Mag. 18(3), 263–281 (2006). https://doi.org/10.1108/09544780610659998 43. Andersson, R., Eriksson, H., Torstensson, H.: Similarities and differences between TQM, six sigma and lean. TQM Mag. 18(3), 282–296 (2006). https://doi.org/10.1108/095447806106 60004 44. Romero, C., Ventura, S.: Educational data mining and learning analytics: an updated survey. WIREs Data Min. Knowl. Discov. 10(3) (2020). https://doi.org/10.1002/widm.1355 45. Okoye, K., Nganji, J.T., Hosseini, S.: Learning analytics for educational innovation: a systematic mapping study of early indicators and success factors. Int. J. Comput. Inf. Syst. Ind. Manage. Appl. 12, 138–154 (2020) 46. Clegg, B., Rees, C., Titchen, M.: A study into the effectiveness of quality management training: a focus on tools and critical success factors. TQM J. 22(2), 188–208 (2010) 47. Salah, S., Carretero, J.A., Rahim, A.: The integration of quality management and continuous improvement methodologies with management systems. Int. J. Prod. Qual. Mang. 6(3), 269– 288 (2010) 48. Adina-Petru¸ta, P., Roxana, S.: Integrating six sigma with quality management systems for the development and continuous improvement of higher education institutions. Procedia. Soc. Behav. Sci. 143, 643–648 (2014). https://doi.org/10.1016/J.SBSPRO.2014.07.456 49. Geraldi, J.G., Kutsch, E., Turner, N.: Towards a conceptualisation of quality in information technology projects. Int. J. Project Manage. 29(5), 557–567 (2011) 50. Coetzee, D.: The effect of the ideology of new managerial professionalism on the South African education system. South Afric. J. Educ. 39(4) (2020). https://doi.org/10.4314/saje. v39i4 51. Deming, W.E.: Elementary principles of the statistical control of quality. JUSE (1950). https://www.worldcat.org/title/elementary-principles-of-the-statistical-control-of-quality-aseries-of-lectures/oclc/2518026. Accessed on 05 Oct 2021 52. Moen, R., Norman, C.: Evolution of the PDCA Cycle. In: 7th ANQ Congress, Tokyo 2009, 17 Sept 2009, pp. 1–11. https://www.anforq.org/activities/congresses/index.html 53. Deming, W.E.: Out of Crisis - Deming’s 14 Points for Management (2000). https://deming. org/explore/fourteen-points/. Accessed on 05 Oct 2021 54. Kirsch, M., Vogg, I., Hosten, N., Fleßa, S.: Quality management in a radiological practice: experiences with a certification for DIN EN ISO 9001:2000. Eur J. Radiol 75(1), e1–e8 (2010) 55. Chen, I.J., Coccari, R.I., Paetsch, K.A., Paulraj, A.: Quality managers and the successful management of quality: an insight. Qual. Manage. J. 7(2), 40–54 (2018). https://doi.org/10. 1080/10686967.2000.11918891 56. Bellamy, S.D.: PDCA problem solving guide: a simplified guide to Team Approacch To Problem Solving - Group Total Quality Manager - 22 August 2000 (5TH Revision - Health Version) (2000). https://www.scribd.com/presentation/492287862/Team-Approacch-to-Pro blem-Solving-pdca 57. Nath, A., Singh, R.: Evaluating the performance and quality of web services in electronic marketplaces. E-Services J. 7(1), 43–59 (2010). https://doi.org/10.2979/ESJ.2010.7.1.43

490

K. Okoye

58. Pimenidis, E., Georgiadis, C.K.: Web services for rural areas—security challenges in development and use. Comput. Electron. Agric. 70(2010), 348–354 (2010) 59. UNESCO: National learning platforms and tools (2021). https://en.unesco.org/covid19/edu cationresponse/nationalresponses. Accessed on 17 Aug 2021

Designing Green Routing and Scheduling for Home Health Care Hossein Shokri Garjan1 , Alireza Abbaszadeh Molaei1 , Fariba Goodarzian2 , and Ajith Abraham2,3(B) 1 Department of Industrial Engineering, Babol Noshirvani University of Technology, Babol, Iran

[email protected]

2 Machine Intelligence Research Labs (MIR Labs), Scientific Network for Innovation

and Research Excellence, Auburn, WA 98071, USA [email protected], [email protected] 3 Center for Artificial Intelligence, Innopolis University, Innopolis, Russia

Abstract. Nowadays, Home health care (HHC) procurement has become a hot topic of research in recent years due to the importance of HHC services for the care of the elderly. With the growth of the percentage of elderly people in different cases, we are witnessing concerns about providing health services to these people in the community. With getting older, the demand for Home Health Care increases. HHC includes a wide range of medical, paramedical and social services that can be provided at home and can be an alternative to receiving these services in a location other than the hospital. Also, due to the possibility of conflict in different countries in the future, with the spread of diseases such as Covid-19 and turning all the facilities and medical and health potential of countries to these epidemics, the need for medical services and home care for the elderly and sick people increases. In this research, a green routing problem is designed for the Home Health care network for the elderly. The network is structured in such a way that the medical service provider with services teams provides services to a group of patients located in a geographical area. The problem is presented as a multi-period mixed integer mathematical model. The purpose of the model is to maximize profits under carbon dioxide emission limits. In this model, an attempt has been made to address the environmental aspects as well. Finally, the mathematical model is solved in GAMS software with numerical examples and its results and performance are presented. Keywords: Home Health Care (HHC) · Green routing problem · Scheduling problem · COVID-19 epidemic

1 Introduction Different types of disasters, such as the spread of infectious diseases, can threaten human lives along with natural disasters. History has shown that pandemics can cause extraordinary suffering and death [1, 2]. COVID-19 is a highly transmissible and pathogenic viral infectious disease that caused a global epidemic and resulted in the loss of many lives worldwide [3]. As of July 22, 2020, more than 15 million cases of COVID-19 have © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 491–504, 2022. https://doi.org/10.1007/978-3-030-96299-9_47

492

H. S. Garjan et al.

been reported worldwide, with nearly 618,000 deaths [4]. Hospital readiness is crucial in organizing national and local responses to infectious disease epidemics. However, it soon became clear that hospitals were under pressure [5]. Even a conservative estimate shows that the health needs of the coronavirus epidemic far exceed the capacity of USA’s hospitals [6]. In older communities where an increasing number of people with illness and disability live longer, monitoring the quality of home health care delivery is important for sustainable health and social care [7]. Health systems and payers are increasingly recognizing a subset of functionally deficient adults who have difficulty accessing the health care system. To optimize their care, home adults need coordinated medical care at home [8, 9]. Like other older communities, the Czech Republic needs to improve the quality of home care. Currently, about 7% of the elderly covered by health care use home care [10]. Home Medical Care (HBMC) refers to clinical practices that provide physician or nurse-led interdisciplinary care for adults living at home, with functional impairments, and serious illnesses that make it difficult to access traditional primary care. Offers. This includes primary home care and home palliative care [11]. Home health care, i.e. meeting and nursing patients in their homes, is a growing part of the medical services business [2, 12]. The purpose of home care services is to help people with functional limitations to live in their own homes. Home care is considered as a possible strategy to reduce hospital use among the elderly [13]. Home care leads to a significant improvement in patients’ quality of life, as they continue to live in a familiar environment at home. On the other hand, this leads to significant cost savings in the health care system, as it avoids the costs of hospitalization [14]. A wide range of skills among home care nurses can prevent unnecessary hospitalization and emergency care [15]. In home services, to serve a range of customers spread across a wide geographical area, a service team starts at the origin (General Service Provider) and visits each dedicated customer before returning to the origin [16]. The rest of the article is described below. In the next section, a summary of the research background of the subject is collected. Section 3 describes the definition of the problem and its assumptions. Also, in Sect. 4, the formulation of the mathematical model of the problem is described. In Sect. 5, the solution method and the results of implementing the mathematical model in the software are presented. Finally, Sect. 6 summarizes the present study and ideas for future research.

2 Literature Review In recent years, optimization of home health care routing has attracted a lot of attention instead of home health care planning. This is because logistics costs of home care are a major concern in organizing home health care [17]. Kergoshen et al. presented a mathematical model of correct linear programming to solve the problem of vehiclespecific routing and the collection of laboratory samples of patients in the province of Quebec. Also, to solve the mathematical model, two methods of forbidden search and neighborhood search have been used [18]. Goodarzian et al. examined the multi-objective multi-product and Multi-period drug supply chain network along with the production of purchase, distribution, inventory order and holding-allocation-routing problem under

Designing Green Routing and Scheduling for Home Health Care

493

uncertainty. Their model type is MINLP and they have used meta-heuristic methods to solve the model [19]. Nahli et al. presented a multi-objective approach based on linear integer programming for human resource allocation and home health care routing. Numerical results obtained from solving the model show that using the proposed method, a good balance can be established between the preferences of patients and caregivers and travel costs [20]. In Breckers et al. A dual-objective model of routing and scheduling home health care is presented. The objectives of the model include minimizing operating costs while increasing the level of service provided to customers by taking into account their preferences [21]. Goodarzian et al. designed a network of green drug supply chains in the state of uncertainty. Their mathematical model type is fuzzy dual-purpose MILP. The purpose of this network is to consider the environmental impacts of establishing pharmacies and hospitals and to reduce the greenhouse effects and control of pollutants. Several metaheuristic algorithms and two new hybrid algorithms have been used to solve the model. Experimental data were obtained through simulation [22, 23]. Ferfita et al. conducted a study in the field of vehicle routing by considering the time window. The purpose of this study is to optimize the appointment of home nurses and the order of visits [24]. Dasserl et al. provided an integrated mixed-time programming formulation for the multifunctional home health care problem. The objectives of the proposed model are to minimize the total working time of caregivers, optimize the quality of services and minimize the difference in working time between nurses and nursing assistants. A memetic algorithm has also been proposed to solve the mathematical model of multi-objective optimization [25]. Shiri et al. used a mixed integer programming model for home health care network design is presented. In this research, the interactive method and robust optimization approach for the problem of multi-purpose home network under uncertainty have been used. The computational results showed the competencies of the proposed powerful model in an uncertain environment and the superiority of the Nimbus method as a new interactive method [26]. Fathollahi-Fard et al. presented a mixed integer linear programming (MILP) model for optimizing multi-stage and multi-period health care. This article is able to solve the problem of optimizing home health care with total cost and patient satisfaction as complex goals and constraints such as patients performing pharmacy tasks, scheduling and routing constraints, delivery time, balancing travel time travel constraints using a two-objective optimization approach [17]. Jean and Van addressed the issue of vehicle routing and appointment scheduling by assigning the team to home care services. To solve it, an innovative algorithm based on Tabu search (TS) was used [27]. Demirblick et al. addressed the issue of home nurse planning. The goal of their model is to maximize the average number of daily appointments for a nurse. Their main new assumption was that patients arrive dynamically in each time period [28]. Manukuska et al. provided a model for day-to-day planning of health care services. The objectives of the model include optimizing the total distance traveled by all caregivers, general service delays starting beyond time windows, and maximum overall operation delays [29]. Environmental pollution is one of the biggest challenges of the present century [30]. Over the past few decades, great strides have been made to incorporate ethical and environmental responsibilities into the core culture of today’s business

494

H. S. Garjan et al.

world. With more emphasis on such responsibilities, an increasing number of companies have opted for “green” (environmentally friendly) designs as competitive strategic weapons [31]. Considering environmental considerations and considering the amount of greenhouse gas emissions from nearly a decade ago, in the literature on individual supply chain issues had begun [2, 32, 33]. Goodarzian et al. developed an optimization approach for the drug supply chain network by considering delivery time and usability with multi-purpose transportation. Innovative method was used to solve the model and two proposed meta-heuristic algorithms improved social engineering optimization (ISEO) and a hybrid of firefly and simulated annealing algorithm (HFFA-SA) were developed for validation [34]. Luo et al. introduced a mixed-integer programming model, addresses the issue of green routing and scheduling in home health care (HHC) with limited simultaneous visits and carbon emissions. The aim of this study is to design a rational logistics route while reducing the impact of the home health care network on the environment. The solution method includes an innovative approach based on two precise methods using Grubi solution method and dynamic programming method (DM) [35]. Increased trade, logistics and transportation have led to increased emissions of harmful gases into the Earth’s atmosphere and pollution [36]. Vehicle Routing Problem (VRP) is a combined optimization and programming problem that finds the optimal set of routes that the fleet of vehicles must pass to reach a specific set of customers [37]. The Green Vehicle Routing (GVRP) problem is a recent alternative to standard traditional vehicle routing models [34, 38]. Transportation plays an important role in the sustainable management of logistics. Traditionally, the main purpose of solving the problem of vehicle routing was to minimize the cost of distance traveled. Today, economic benefits cannot be the only driver of achieving sustainability, and environmental issues must also be considered [39]. In their own research, Michael et al. examined a vehicle routing issue with economic and environmental aspects simultaneously in order to achieve a sustainable vehicle routing decision. Also, firefly algorithm has been used to solve the problem of vehicle routing and TOPSIS technique has been used to integrate economic and environmental factors [39]. Foroutan et al. In this paper, have examined the issue of routing and planning of green vehicles with a heterogeneous transportation vehicles, including reverse logistics in the form of collecting returned goods with early and late costs. Their goal is to minimize operating and environmental costs and create a balance between them [40]. Goodarzian et al. Examined a multi-objective, multi-level, multi-product, and multiperiod issue for a sustainable medical supply chain network, as well as the distribution of drugs related to Covid-19 patients and drug production and delivery periods due to the perishability of some drugs [41]. Gupta, P., et al. presented an Iot-based health monitoring system for emergency medical services that flexibly demonstrates the collection, integration, and operation of Iot data that is capable of supporting ICU services.The proposed result of this project is to provide appropriate and efficient medical services to patients by connecting and collecting information through health status monitors [42]. Pasquadibisceglie and et al., proposed a new solution for evaluating critical indicators using a non-invasive monitoring. The proposed system is a smart device consisting of a mirror to a camera that

Designing Green Routing and Scheduling for Home Health Care

495

captures a person’s face and uses photo plethysmography to instantly estimate heart rate and heart rate saturation, and can be easily used at home [43]. In the event of epidemics such as Covid-19 that all the capacity of the health care system is used to deal with it. Also, for reasons such as the decline in the quality of response to the elderly in epidemic conditions, the dangerous presence of the elderly in centers such as hospitals due to the spread of the disease and the refusal of the elderly to attend these centers, the existence of home medical care centers are more needed. Based on previous studies in the field of home medical care, this article attempts to provide a new model for the issue of green routing of home medical services for the elderly. The objectives of the proposed model include increasing profits and reducing costs for the home health care network. This model also deals with the environmental aspects and the effects of the home medical care transportation network on the environment.

3 Problem Definition The green routing of elderly Home Health Care is a multi-period Mixed integer linear programming and defined on a graph G = (N0 , A) consisting of nodes and arcs that N0 = N ∪ {0} represents a set of nodes containing a set of points N = {1, 2, ...., n} as a set of patients in Different geographical areas and {0} as a center providing medical services with a set of different medical service teams. We also show the set of arcs with A and thus A = {(i, j) : i, j ∈ N0 , i = j} indicate the relationship between different points of the model network. The T = {1, ......, τ } set represents time periods. It is assumed that the medical service center has only three services: 1) Laboratory affairs, injections and bandage. 2) Home Visit Doctors and check-up. 3) Performs physiotherapy and massage therapy. And depending on the needs of patients and in order to respond to it, one of the teams that has expertise related to these three categories of services is sent. Also, according to the type of service, the number of service team members and the equipment they need, K-type heterogeneous vehicles are used for transportation. Each service team starts its journey from the origin node and returns to the origin node after visiting the patients. According to the explanations, Fig. 1 shows a schematic view of the flow of Home Health Care services for the elderly. Which includes a medical care center and several patients in different geographical locations that these patients receive by medical care teams in different periods of service. The proposed model is a green routing problem that in addition to meeting the needs of patients pursues environmentally friendly goals. The purpose of the proposed mathematical model is to maximize profits under carbon dioxide emission limits, which are obtained by subtracting costs from revenue from service delivery. Model network costs include: cost of travel between points, cost of service team, cost of carbon dioxide emissions by vehicles, and cost of non-response to customers. The total service time of medical service teams to patients is limited and fixed. It is also assumed that all costs are clear and precise in the planning horizon. Another

496

H. S. Garjan et al.

Medical Services Cen Elderly home patients

Medical care team Medical care vehicle

Fig. 1. Schematic view of the green routing problem model of home medical services

assumption is that, in order to minimize the number of customers who are not served. For each unit of customers who suffered from non-response, a penalty was considered and we try to be able to reduce the costs of non-response to customers as much as possible by considering this section.

4 Mathematical Formulation 4.1 Indices, Parameters and Decision Variables In this section, the indices used to introduce the indices, parameters and variables of the model are defined. Indices: i, j o i, j s k t

Total points index, i, j ∈ N0 Medical Center Index, 0 ∈ N0 Patient Index, i, j ∈ N Medical Service Index, s ∈ S Vehicle index, k ∈ K Period index, t ∈ T Parameters

disij coijk α1k ϕ∗ cfs cvk picis mcois θ 1ijk

The distance between points i and j Cost of travel between i and j points by vehicle k Carbon dioxide emissions for the vehicle k Estimated average cost of carbon emissions Fixed cost of hiring medical team s Fixed cost of using the vehicle k The cost of a fine for each patient i who has not received medical care s Income from each patient i who has received medical care s Time required to cross arc(i,j) by vehicle k

Designing Green Routing and Scheduling for Home Health Care

θ 2is ψ

497

Time required to provide team service s to the patient i Maximum service time

Decision variables: Zijskt The binary variable takes the value (1) if the medical team j crosses the arc (i,j) with the vehicle k in period t. Otherwise (0). Vist The binary variable takes the value (1) if the medical team s is assigned to the patient i in period t. Otherwise (0). 4.2 Mathematical Model According to the definition of Indices and the definition of the problem in the previous sections, the problem of responding to home care services for the elderly during the Covid-19 epidemic is modelled as a mixed integer linear programming as follows, taking into account environmental considerations. Max TP = Q1 − (Q2 + Q3 + Q4 + Q5 + Q6 )

Q1 =



mcois . Vist

(1)

(2)

i∈N s∈S t∈T

Q2 =



cfs . Vist

(3)

picis .(1 − Vist )

(4)

i∈N s∈S t∈T

Q3 =

 i∈N s∈S t∈T

Q4 =

  

disij . coijk . Zijskt

(5)

i∈N0 j∈N0 s∈S k∈K t∈T

Q5 =

  

cvk . Zijskt

(6)

ϕ ∗ . α1k . disij . Zijskt

(7)

i∈N0 j∈N0 s∈S k∈K t∈T

Q6 =

   i∈N0 j∈N0 s∈S k∈K t∈T

Subject to:

498

H. S. Garjan et al.



Z0jskt ≤ 1 , ∀ s ∈ S, k ∈ K , t ∈ T

(8)

Zi0skt ≤ 1 , ∀ s ∈ S, k ∈ K , t ∈ T

(9)

j∈N

 i∈N



Zijskt =

i∈N0



Zjiskt , ∀ i ∈ N , s ∈ S, k ∈ K , t ∈ T

(10)

i∈N0



Vist = 1

, ∀i∈N

(11)

s∈S t∈T



Zijskt = Vist , ∀ i ∈ N , s ∈ S, t ∈ T

(12)

j∈N0 k∈K

 i∈N

θ 2is . Vist +



θ 1ijk . Zijskt ≤ ψ , ∀ k ∈ K , s ∈ S , t ∈ T

(13)

Zijskt , Vist ∈ {0, 1} , ∀ i ∈ N0 , j ∈ N0 , s ∈ S, k ∈ K , t ∈ T

(14)

i∈N j∈N

Equation (1) represents the objective function of the problem that maximize the profit from patient care for the medical center and is obtained by deducting costs from revenues. Equation (2) shows the income from medical services for the elderly for the medical service center. Equation (3) shows the costs of using medical groups. Equation (4) shows the costs of fines for non-response to patients. Equation (5) calculates the cost of transportation or in other words the cost of travel between nodes. Equation (6) calculates the costs of using different vehicles to refer patients to medical care teams. Equation (7) shows the costs of carbon emissions by the transportation system. Equations (8) and (9) indicate that the origin and destination of all vehicles is the origin node. In this way, medical care teams start their travel from the medical service center and return to the medical service center after treating the patients. Equation (10) ensures the continuity of the route for medical service teams in such a way that if we enter a patient’s node in a period, we must leave it in the same period with the same vehicle. Equation (11) indicates the limitation that only one medical care team should be assigned to each patient per period. Equation (12) ensures that when it is greater than 1, then the medical team should leave the node of the patient or medical service center, in fact showing the correlation of allocation and routing decisions. Equation (13) indicates that the total time of medical services for patients in the nodes as well as the time for medical teams to cross the routes by vehicles should not be more than time. Equation (14) also defines the decision variables of the mathematical model.

Designing Green Routing and Scheduling for Home Health Care

499

5 Solution Method and Numerical Results Organizations’ attention to environmental issues in supply chain-related decisions has led to environmental decisions as well as environmentally friendly decisions. In this paper, an mixed integer linear programming model is proposed to maximize profits while reducing carbon dioxide emission costs. This model seeks to provide home medical care for the elderly during and after Covid-19. In order to evaluate the performance of the proposed definitive model, five sample problems with small to large dimensions have been considered. The proposed mathematical model is then solved using GAMS software and Ciplex solver on a system with an Intel (R) Core (TM) i5-4200U CPU @ 1.60 GHz processor and 8 GB of RAM. The results of solving the model as well as the different dimensions of the sample problems are shown in Table 1. A uniform distribution is used to generate other parameters of the proposed model. Table 1. The results obtained from solving the model with GAMS software Sample

Dimensions of the problem

Objective value (total profit)

Cost of CO2 emissions

Solving time (seconds)

o

i

s

k

t

1

1

3

3

2

2

361

128

00.262

2

1

10

3

3

5

1287

402

00.397

3

1

20

3

5

8

3856

498

00.746

4

1

35

3

8

15

6386

668

02.872

5

1

60

3

10

20

12065

849

10.079

The issue of providing medical services to the elderly at home seeks to meet the medical needs of the elderly. According to the 3 types of medical services defined for the medical service center, how to assign medical service teams to the elderly in different time periods for problem example (1) is shown in Table 2. Table 2. How medical teams were assigned to patients in different periods Patient

Medical team 1

1

Period 2

3

*

1

2

*

2

*

*

3

*

*

500

H. S. Garjan et al.

In this section, the effect of changing the distance between nodes on supply chain performance in terms of total network profit, network costs and emission costs are investigated. For this purpose, in Fig. 2, the amount of changes in the profit target function, the cost of greenhouse gas emissions and the current costs of the home health care network for different percentages of increase in the distance between nodes parameter are shown in separate diagrams. The results show that the profit objective function shows different behavior in terms of greenhouse gas emission costs and current costs of the home health care network by changing different amounts of distance between nodes. In this way, increasing the distance, increases the emission costs of the greenhouse gases and the running costs of the system and on the other hand, reduces the profits of the centers that provide home health care services. To the extent that a 150% increase in distances brings negative benefits and losses for the service provider center. Therefore, Home Health Care service providers are only able to provide home care services within a certain range. To have a balance between costs and benefits at the same time, and this geographical area of service is determined by parameters such as distance, shipping cost, carbon emission cost, and time required to travel between nodes. 600

Value

400 200 0 Objective function -200

0%

25%

Cost of CO2 emission Current system costs 50%

75%

100%

150%

Fig. 2. Comparison of changes in the objective function and current system costs and the cost of greenhouse gas emissions relative to changes in distance between nodes

This study presents a proposed green routing model for the problem of home Health services for the elderly. Another result of solving this model is to determine the optimal routes for teams to move towards patients and provide services to them and return to the medical service center. Figure 3 shows the optimal routes and how to allocate machines to medical care teams for example (1). (The indice t indicates the time period. The indice s indicates the type of medical service. And the indice k indicates the type of vehicle). In this paper, a proposed model of green routing of Home Health services for the elderly during Covid-19 was presented. Then the proposed mathematical model in different dimensions was solved using GAMS software and the results were expressed. Based on the obtained results, it can be confirmed that the proposed model can be useful and practical to decision makers in designing a green network to meet the medical needs of the elderly during the Covid-19 epidemic and beyond. The proposed model also seeks to reduce the costs of emitting carbon dioxide from its heterogeneous vehicle fleet.

Designing Green Routing and Scheduling for Home Health Care

501

3

1 t=1 s=2 k=2

t=2 s=3 k=1

2

Fig. 3. Schematic view of the optimal paths for the sample problem (1)

6 Conclusions Hospitals and medical centers are at the forefront of epidemic diseases. The experience of countries at the time of the outbreak of infectious diseases showed that these centers and the health care system are easily under pressure and the level of quality of medical services and care for ordinary patients is reduced. These centers also become high-risk places for patients. This situation is clearly visible now that the world is involved in the Covid-19 pandemic. These experiences show that today and even in the future, Home Health Care and home medicine can be an auxiliary arm for the health care system in the event of an epidemic. To receive medical and care services in safer and healthier environments for patients and high-risk groups such as the elderly. Also, as time goes on and the demand for home care services increases, planning to achieve an efficient program to meet the needs of patients becomes more difficult and complex and new challenges are faced every day. Therefore, in this article, we have tried to provide an integer planning model for the green routing problem of home health care services. In this network, medical service centers are served by teams of patients who are scattered in a specific geographical area and return to the medical service center. The purpose of the model is to maximize the benefits of providing services to patients for the medical center. Also, in this model, attention has been paid to the environmental aspects and reduction of carbon dioxide emissions caused by the network transportation system. The proposed model was implemented for simulated data in different dimensions in GAMS software. Problem solving in different dimensions and the results obtained from it confirm that the proposed model can help decision makers in designing and planning a home medical care network for the elderly. As a basis for future research, the proposed model of Home Health Care can be implemented in a case study and the model can be evaluated for better performance. Some parameters of the problem can be considered indefinitely. Also, solving the model in larger dimensions with meta-heuristic algorithms is another issue that can be considered. In future work, the model can be planned several times with goals such as increasing the level of service to patients and addressing the sustainability aspects of the model. Patients’ opinions, such as receiving service from a specific caregiver, can also be incorporated into the future model to be more similar to real-world situations. Acknowledgement. This research has been financially supported by The Analytical Center for the Government of the Russian Federation (Agreement No. 70-2021-00143 dd. 01.11.2021, IGK 000000D730321P5Q0002).

502

H. S. Garjan et al.

References 1. Dasaklis, T.K., Pappis, C.P., Rachaniotis, N.P.: Epidemics control and logistics operations: a review. Int. J. Product. Econ. 139(2), 393–410 (2012) 2. Goodarzian, F., Abraham, A., Fathollahi-Fard, A.M.: A biobjective home health care logistics considering the working time and route balancing: a self-adaptive social engineering optimizer. J. Comput. Des. Eng. 8(1), 452–474 (2021). https://doi.org/10.1093/jcde/qwaa089 3. Shereen, M.A., et al.: COVID-19 infection: origin, transmission, and characteristics of human coronaviruses. J. Adv. Res. 24, 91 (2020) 4. Colaneri, M., et al.: Honors and Costs of Being Heroes: Risk Factors and Epidemiology of COVID-19 Infection among Health Care Workers of a Severely Hit COVID-19-Referral Hospital in Italy (2020) 5. Farsalinos, K., et al.: Improved strategies to counter the COVID-19 pandemic: Lockdowns vs primary and community healthcare. Toxicol. Rep. 8, 1–9 (2021) 6. Emanuel, E., Persad, G., Upshur, R.: Allocation of scarce medical resources in the time of Covid-19. N. Engl. J. Med. 10, 1–11 (2020). Published online ahead of print 2020 7. Hoff, T.: Medical home implementation: a sensemaking taxonomy of hard and soft best practices. Milbank Q. 91(4), 771–810 (2013) 8. Ornstein, K.A., et al.: Epidemiology of the homebound population in the United States. JAMA Intern. Med. 175(7), 1180–1186 (2015) 9. Ghasemi, P., Goodarzian, F., Gunasekaran, A., Abraham, A.: A bi-level mathematical model for logistic management considering the evolutionary game with environmental feedbacks. Int. J. Logist. Manage. (2021). https://doi.org/10.1108/IJLM-04-2021-0199. (ahead-of-print) 10. Yamada, Y., Kisvetrová, H., Topinková, E.: Evaluation of a pilot study to introduce outcome based home care in the Czech Republic. Kontakt 16(3), e149–e154 (2014) 11. Ritchie, C.S., Leff, B.: Population health and tailored medical care in the home: the roles of home-based primary care and home-based palliative care. J. Pain Symptom Manage. 55(3), 1041–1046 (2018) 12. Bertels, S., Fahle, T.: A hybrid setup for a hybrid scenario: combining heuristics for the home health care problem. Comput. Oper. Res. 33(10), 2866–2890 (2006) 13. Landi, F., et al.: A new model of integrated home care for the elderly: impact on hospital use. J. Clin. Epidemiol. 54(9), 968–970 (2001) 14. Carello, G., Lanzarone, E.: A cardinality-constrained robust model for the assignment problem in home care services. Eur. J. Oper. Res. 236(2), 748–762 (2014) 15. Harteloh, P.P.: The meaning of quality in health care: a conceptual analysis. Health Care Anal. 11(3), 259–267 (2003) 16. Xiang, Z., Chu, C., Chen, H.: The study of a dynamic dial-a-ride problem under timedependent and stochastic environments. Eur. J. Oper. Res. 185(2), 534–551 (2008) 17. Fathollahi-Fard, A.M., Ahmadi, A., Goodarzian, F., Cheikhrouhou, N.: A bi-objective home healthcare routing and scheduling problem considering patients’ satisfaction in a fuzzy environment. Appl. Soft Comput. 93, 106385 (2020). https://doi.org/10.1016/j.asoc.2020. 106385 18. Kergosien, Y., Ruiz, A., Soriano, P.: A routing problem for medical test sample collection in home health care services. In: Matta, A., Li, J., Sahin, E., Lanzarone, E., Fowler, J. (eds.) Proceedings of the International Conference on Health Care Systems Engineering, pp. 29– 46. Springer International Publishing, Cham (2014). https://doi.org/10.1007/978-3-319-018 48-5_3 19. Goodarzian, F., Hosseini-Nasab, H., Muñuzuri, J., Fakhrzad, M.-B.: A multi-objective pharmaceutical supply chain network based on a robust fuzzy model: a comparison of meta-heuristics. Appl. Soft Comput. 92, 106331 (2020). https://doi.org/10.1016/j.asoc.2020. 106331

Designing Green Routing and Scheduling for Home Health Care

503

20. En-nahli, L., Allaoui, H., Nouaouri, I.: A multi-objective modelling to human resource assignment and routing problem for home health care services. IFAC-PapersOnLine 48(3), 698–703 (2015) 21. Braekers, K., et al.: A bi-objective home care scheduling problem: analyzing the trade-off between costs and client inconvenience. Eur. J. Oper. Res. 248(2), 428–443 (2016) 22. Goodarzian, F., et al.: A new bi-objective green medicine supply chain network design under fuzzy environment: hybrid metaheuristic algorithms. Comput. Indust. Eng. 160, 107535 (2021) 23. Goodarzian, F., Bahrami, F., Shishebori, D.: A new location-allocation-problem for mobile telecommunication rigs model under crises and natural disasters: a real case study. J. Ambient. Intell. Humaniz. Comput. 1−19 (2021) https://doi.org/10.1007/s12652-021-03461-w 24. Frifita, S., Masmoudi, M., Euchi, J.: General variable neighborhood search for home healthcare routing and scheduling problem with time windows and synchronized visits. Electron. Notes. Disc. Math. 58, 63–70 (2017) 25. Decerle, J., et al.: A memetic algorithm for multi-objective optimization of the home health care problem. Swarm Evol. Comput. 44, 712–727 (2019) 26. Shiri, M., Ahmadizar, F., Mahmoudzadeh, H.: A three-phase methodology for home healthcare routing and scheduling under uncertainty. Comput. Indust. Eng. 158, 107416 (2021) 27. Zhan, Y., Wan, G.: Vehicle routing and appointment scheduling with team assignment for home services. Comput. Oper. Res. 100, 1–11 (2018) 28. Demirbilek, M., Branke, J., Strauss, A.: Dynamically accepting and scheduling patients for home healthcare. Health Care Manag. Sci. 22(1), 140–155 (2018). https://doi.org/10.1007/ s10729-017-9428-0 29. Mankowska, D.S., Meisel, F., Bierwirth, C.: The home health care routing and scheduling problem with interdependent services. Health Care Manag. Sci. 17(1), 15–30 (2013). https:// doi.org/10.1007/s10729-013-9243-1 30. Balasubramanian, S., Shukla, V.: Green supply chain management: an empirical investigation on the construction sector. In: Supply Chain Management: An International 31. Min, H., Kim, I.: Green supply chain research: past, present, and future. Logist. Res. 4(1–2), 39–47 (2012). https://doi.org/10.1007/s12159-012-0071-3 32. Benjaafar, S., Li, Y., Daskin, M.: Carbon footprint and the management of supply chains: insights from simple models. IEEE Trans. Autom. Sci. Eng. 10(1), 99–116 (2012) 33. Goodarzian, F., Kumar, V., Abraham, A.: Hybrid meta-heuristic algorithms for a supply chain network considering different carbon emission regulations using big data characteristics. Soft. Comput. 25(11), 7527–7557 (2021). https://doi.org/10.1007/s00500-021-05711-7 34. Goodarzian, F., Kumar, V., Ghasemi, P.: A set of efficient heuristics and meta-heuristics to solve a multi-objective pharmaceutical supply chain network. Comput. Indust. Eng. 158, 107389 (2021) 35. Luo, H., Dridi, M., Grunder, O.: A green routing and scheduling problem in home health care. IFAC-PapersOnLine 53(2), 11119–11124 (2020) 36. Gajanand, M., Narendran, T.: Green route planning to reduce the environmental impact of distribution. Int. J. Log. Res. Appl. 16(5), 410–432 (2013) 37. Dantzig, G., Ramser, J.H.: The truck dispatching problem. Manage. Sci. 6, 80–91 (1959) 38. Hosseini-Nasab, H., Lotfalian, P.: Green routing for trucking systems with classification of path types. J. Clean. Prod. 146, 228–233 (2017) 39. Micale, R., Marannano, G., Giallanza, A., Miglietta, P.P., Agnusdei, G.P., La Scalia, G.: Sustainable vehicle routing based on firefly algorithm and TOPSIS methodology. Sustain. Futures 1, 100001 (2019). https://doi.org/10.1016/j.sftr.2019.100001 40. Foroutan, R.A., Rezaeian, J., Mahdavi, I.: Green vehicle routing and scheduling problem with heterogeneous fleet including reverse logistics in the form of collecting returned goods. Appl. Soft. Comput. 94, 106462 (2020)

504

H. S. Garjan et al.

41. Goodarzian, F., Taleizadeh, A.A., Ghasemi, P., Abraham, A.: An integrated sustainable medical supply chain network during COVID-19. Eng. App. Artif. Intell. 100, 104188 (2021). https://doi.org/10.1016/j.engappai.2021.104188 42. Gupta, P., et al.: IoT based smart healthcare kit. In: 2016 International Conference on Computational Techniques in Information and Communication Technologies (ICCTICT). IEEE (2016) 43. Pasquadibisceglie, V., Zaza, G., Castellano, G.: A personal healthcare system for contact-less estimation of cardiovascular parameters. In: 2018 AEIT International Annual Conference. IEEE (2018)

The Bi-level Assembly Flow-Shop Scheduling Problem with Batching and Delivery with Capacity Constraint Hossein Shokri Garjan1 , Alireza Abbaszadeh Molaei1 , Nazanin Fozooni2 , and Ajith Abraham3,4(B) 1 Department of Industrial Engineering, Babol Noshirvani University of Technology, Babol, Iran

{Hossein.shokri,alirezaabbaszadeh}@nit.ac.ir

2 Department of Industrial Engineering, Ferdowsi University, Mashhad, Iran 3 Machine Intelligence Research Labs (MIR Labs), Scientific Network for Innovation

and Research Excellence, Auburn, WA 98071, USA [email protected] 4 Center for Artificial Intelligence, Innopolis University, Innopolis, Russia

Abstract. In most manufacturing and assembly systems, a number of operations are performed on each job. Most of these operations are performed in the same order on all tasks, ie the works flow in the same direction. In such an environment, known as flow shop, the machines are arranged in series. In this the bi-Level assembly flow-shop scheduling problem with Capacity Constrains batching and delivery system is presented. Here, m is a single machine that do different parts of the job, and in the second part, number of machines have the duty of assembly. In this paper, a mixed nonlinear integer math model is formulated. The objective of this model is include minimizing the cost of delays, delivery, and categorization. For solving the model in small dimensions, the Branch and Bound method are used in GAMS and finally numerical examples and Analysis are done. Keywords: Assembly flow shop · Scheduling · Batched delivery system · Mathematical model

1 Introduction The first study of two-stage production scheduling was reviewed [1]. With the advent of modern technologies in production and industry, many efforts have been made to coordinate production and distribution in supply chain management [2, 3]. The production planning process, which helps to find the optimal use of available resources to meet market demand, has always been one of the most important activities in manufacturing and service companies. Meanwhile, the issue of scheduling and sequencing operations as one of the final stages of production planning, plays a significant role in achieving its goals. In the manufacturing industry, the issue of scheduling, depending on the production environment and the constraints and functions of the target, can be a simple or very complex issue. The issue of flow shop has attracted the attention of many researchers due © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 505–516, 2022. https://doi.org/10.1007/978-3-030-96299-9_48

506

H. S. Garjan et al.

to its widespread use in industry and economics. Flow Shop is one of the optimization issues in computer science and operations research, which is also called line layout. This refers to a process or shop that runs on a flow when all jobs have the same processing path. It is used for frequent productions that include constant job flows and a relatively large production volume. Flow control in these jobs is easy in this case, but it is difficult to respond to product diversity. The main goal is to find the best plan and arrangement for those jobs in order to minimize longevity. Nowadays, various industries are moving towards lowering their price. In this regard, in manufacturing companies, production planning and distribution is an important issue. However, in supply chain planning, including classification and delivery, the cost of classification and the capacity of the categories in question have not been considered. The structure of the article is as follows: in Sect. 2 the research background is presented and then in Sect. 3 the problem is stated and the problem is formulated by presenting a programming model of mixed nonlinear integer and finally, in Sect. 4, a case study and performance evaluation are reviewed.

2 Literature Review A scheduling process is a process in which resources are allocated to perform a set of tasks over a period of time, and generally two sets of objective functions can be used in scheduling problem. Rahman et al. [4] design a mathematical model and study the production and distribution issues with start-up time, which depends on the order for several customers. To solve these problems, three meta-heuristic algorithms were used. The goal is to minimize the total cost of the flow shop. A case study in industry has been used to validate the problem. Jabbari et al. [5] developed a mathematical model for examining the timing of a custom production system that includes a shop flow production line with a parallel assembly line that produces different products in two stages. The goal is to increase the manufacturing time of goods using efficient planning. Because the NP-hard model uses genetic algorithm (GA) and particle swarm optimization methods to solve the model. Zhang et al. [6] examined preventive maintenance activities in a two-step flow shop schedule using a MILP model. Heuristic method was used to solve the model [7–12]. Yang et al. [12] investigated a new distributed assembly flow scheduling issue with flexible assembly and batch delivery. The goal is to minimize the total cost of delivery and delays. Seven algorithms including variable neighborhood descent (VND) algorithm, and two Iterated Greedy (IG) have been used to solve the model. The results showed that the proposed batch allocation could significantly improve the solution. Basir et al. [13] studied integrated production and distribution planning on two stages of flow shop assembly with batch delivery system. Their mathematical model is MILP. The goal is to plan jobs in two-phase flow shop and deliver completed products in an appropriate number with a minimum number of weighted latency jobs and total grouped delivery costs. An improved two-level genetic algorithm has been used to solve the model. Ceylan et al. [14] investigated a coordinated planning problem for a multi-stage supply chain network. The MILP problem has two purposes. The goal is to minimize the total cost of the supply chain. The structure of this supply chain is multi-product

The Bi-level Assembly Flow-Shop Scheduling Problem

507

and multi-period. To integrate the issue of flexible flow shop planning, a supply chain network has been integrated. The results showed that the sum of weights is the best MODM method to solve the problem. Cheng and Kahlbacher [15] raised issues related to batch delivery planning [16, 17]. Today, the main industries are moving towards Just in Time (JIT) production and expanding the world market, manufacturing companies are moving more and more towards competition in order to reduce their costs. Therefore, integrated production and distribution planning is very important for manufacturing companies [18]. The problem of two-stage assembly is widely used in industry. Mazdeh and Rostami [19] proposed a mixed linear algorithm and a single branch algorithm for dual-flow scheduling problems to minimize latency and delivery costs. The problem of flow shop production scheduling is an issue of np-hard optimization in the production scheduling literature [20]. Belabid et al. [21] designed a mathematical model for the flow shop problem that has a start-up time independent of the sequence. The mathematical model is of the MILP type. The goal is to minimize the maximum completion time. Several innovative and meta-innovative methods have been used to solve the model. Taxido et al. [22] examined the issue of permutation flow shop scheduling problem using a combination of firefly algorithm and variable neighborhood search algorithm. The flow shop scheduling issue is an NP-hard issue. In order to find high quality solutions in a reasonable computational time, heuristic and meta-heuristic algorithms have been used to solve the problem. In order to test the effectiveness and efficiency of the proposed method, we used a set of benchmark samples with different sizes of articles. Wang et al. [23] proposed a mathematical model to address the issue of flow shop schedule delivery for multiple customers. In this case, first the products are produced as alternatives and then they are delivered to multi customers. The purpose of this article is to minimize the total cost of delay and batch delivery. To solve the model, innovative methods of genetic algorithm and variable neighborhood search and a new meta-heuristic method have been used. Al-Bihadili et al. [24] examined a multi-objective flow shop optimization model with the arrival of new work and machine failure. The purpose of this is to increase the working time according to the stability and strength. To solve the model, the Biased Randomized algorithm was used to repeat the greed [25]. Ochi et al. [26] studied the flow shop assembly schedule, which aims to minimize make span. To solve the model, iterated greedy-based method was used and to improve the quality of the solution, four search methods were used. Then, comparative results between the performances of algorithms were performed. Pessoa et al. [27] investigated a flow shop planning issue with delivery dates and cumulative returns. Their goal is to maximize profits. Computational experiments were performed by considering the methods of solving the problem and the local search method. In this paper, the issue of flow shop assembly schedule has been studied, and the delivery and classification system has also been considered in this issue. This is a NPhard model. The purpose of the model is to minimize the cost of delays, delivery and classification. To solve the MILP model, the branch and boundary method has been used.

508

H. S. Garjan et al.

3 Problem Description In this paper we have N jobs belonging to H customer and nj denote the total number of jobs related to the customer by j, so the sum of the total number of jobs related to customers shows the total jobs N. In order to complete each task, each M part must be completed, these components are completely unique, and in the first stage of the process, there is an independent M machine that is responsible for performing the various components of each task. When the components of each work are completed by the machines in the first stage, in order to assemble the parts, the work must be assigned to one of the assembly machines that is free. After completing the work, it is necessary to decide on the classification and submission of work. After completion, each task can be sent to the customer, or if the classification is done, it must wait for the completion of all tasks related to that category, and after completing all the tasks in that category, then it will be sent to the customer. Due to this, the maximum number of categories can be considered equal to the total number of works, and according to the explanations given, if the category is sent to the customer, the time of sending the category is equal to the time when the last task related to that category is completed. The cost that we accept for the delivery of each category is proportional to the size of the category, but the size of the categories is not unlimited, so that if the total size of the work related to a customer exceeds the size of the category We have to increase the number of categories, which will cost us more categories. The goal is to minimize the total delay time, delivery costs and number of batches. However, due to the target sentences that are not similar in terms of gender, we use the scores of G and δ for delays as well as categorization, respectively. Therefore, the objective function is represented as Eq. (1). In the real world example, when there is a delay in delivery to the customer, there is a penalty unit for the delay, as well as a fee for the categories to be done by the relevant personnel, which is considered here . Waiting to complete other customer orders and placing orders in one batch can increase the delay of a number of orders and on the other hand reduce delivery costs. In fact, this study seeks balance between delay and delivery cost in the supply chain. Bank et al. showed that the two-machine flow shop problem with the objective function of the sum of delays without considering the reduction is Np-hard. Since the problem under study, in addition to being two-stage, also includes classification and capacity issues, and the objective function is the sum of delays, delivery costs and classification costs, so the complexity of this issue is more than the issue studied by Bank et al. and is considered NP-hard. It is worth noting that the machine that is available during the operation is not allowed to perform other operations at that time. 3.1 Model Assumptions 1. All machines are available in zero time. 2. Each work consists of components that must first be performed on the machine in the first stage and then transferred to the assembly machines for assembly in the second stage. 3. Preparation time is considered zero and all works can start at zero time. 4. Stoppage due to line failure or maintenance operations is not considered in the issue.

The Bi-level Assembly Flow-Shop Scheduling Problem

509

5. Each part of the work, when placed on a machine, is on that machine until the end of the operation, and it is not allowed to interrupt work either in the first stage or in the second stage. 6. The assembly machines in the second stage are of the same material and the part can be placed on the available assembly machine after the completion of all the components. 7. Categories that include tasks always include at least one task. Indices i j m q k b

Job Index i = 1, 2, 3, ..., nj Customer Index j = 1, 2, 3, ..., H Index of machines in the first stage m = 1, 2, 3, ..., M Index of assembly machines in the second stage q = 1, 2, 3, ..., Q Job position index in the first stage k = 1, 2, 3, ..., N Classification index b = 1, 2, 3, ..., N

Parameters Pijm Sij Dij Dj gi G l δ

Process time i for client j on machine m Assembly time work i for customer j Delivery time i for customer j Delivery time for the customer j Work size i Handle size The cost of categorizing each category Cost per delay unit

Decision variables  Pkm Xijk

Fkm FFk Sk Ckq CTkq Rkq Cij Ykq Cbj

Work process time on position k on machine m (in the first step) This variable is 1 when job i is related to client j in position k, otherwise it is zero Completion time of a work component that is on the k position on the m machine Completion time of all components of the work or when the work is ready for assembly Assembly time of the work that is on the k position Completion time k on assembly machine q (If not assigned to machine q, this time is zero.) Completion time of the last work before k th which is assigned to the assembly machine q The time it takes for assembly work k to start on machine q Completion time of task i, which belongs to customer j This variable will be equal to 1 when work k is assigned to the assembly machine q, otherwise it will be zero Completion time of category b belongs to customer j

510

H. S. Garjan et al.

RRij Zijb

Time to send work i to client j If the i-th job belongs to the j-th customer in category b, it is 1, otherwise it is zero If the  job belonging to customer j is in category b it is 1, otherwise it is zero (∀j b Vjb equal to the number of categories owned by customer j) Category b size belongs to the customer j Number of components of category b belonging to the customer j

Vjb SSVbj Vjb

3.2 Mixed Integer Nonlinear Programming Model

Min

nj H  

δ · Tij +

j=1 i=1

N H  

Vjb · Dj +

j=1 b=1

N 

l · Vb

(1)

∀i, ∀j

(2)

b=1

Tij ≥ 0

Tij ≥ RRij − dij

∀i, ∀j

(3)

RRij + M (1 − zijb ) ≥ Cbj

∀i, ∀j, ∀b

(4)

Cbj + M (1 − zijb ) ≥ Cij

∀j, ∀b, ∀i

(5)

Cij + M (1 − xijk ) ≥



Ckq

(6)

q

 ) C1q = Y1q × (FF1 + Skq

 Ckq = Ykq × (Rkq + Skq )

∀q

(7)

∀k, k = 1, ∀q

(8)

Rkq ≥ FFk

∀k, k = 1, ∀q

(9)

Rkq ≥ CTkq

∀k, k = 1, ∀q

(10)

The Bi-level Assembly Flow-Shop Scheduling Problem

CTkq ≥ C(e−1)q

Fkm =

∀k, k = 1, ∀q, e = 2, ..., k

k 

 Pkm

511

(11)

∀k, ∀m

(12)

∀k, ∀m

(13)

∀i, ∀j

(14)

∀K

(15)

∀K

(16)

a=1

FFk ≥ Fkm

N 

Xijk = 1

k=1 nj H  

Xijk = 1

j=1 i=1



Ykq = 1

q

 Pkm =

nj H  

(Xijk · Pijm )

∀k, ∀m

(17)

∀k

(18)

j=1 i=1

Sk =

nj H  

(Xijk · Sij )

j=1 i=1

H 

Zijb ≤ 1

∀i, ∀b

(19)

 ∀i, ∀j b2 − 4ac

(20)

j=1

N 

Zijb = 1

b=1

−

nj  i=1

Zijb + M · (Vjb ) ≥ 0

∀j, ∀b

(21)

512

H. S. Garjan et al.

nj 

Zijb + M · (1 − Vjb ) > 0

∀j, ∀b

(22)

∀b

(23)

i=1

SVb =

H N  

Zijb · gi

i=1 j=1

Vb =



SVb G

 ∀b

(24)

Objective function (1) shows the minimization of total delays as well as delivery costs and classification costs. Constraints (2) and (3) calculate the delay for each task. Constraint (4) Calculates the delivery time for each task, which is equal to the time it takes to complete the category. Constraint (5) Calculates the completion time of each category corresponding to the longest completion time in this category. Constraint (6) Calculates the time of completion of work i related to the customer j and related to the assembly machine assigned to it. Constraints (7) and (8) calculate the time to complete the task, which is in the first and kth position. Constraints (9), (10) and (11) are necessary restrictions for use. Constraint (13) according to the definition of the variable in the description of variables. Constraint (12) Calculates the completion time of a task related to its own process time. Constraints (14) and (15) each job is only for one sequence, and each sequence is occupied by only one job, respectively. Constraint (16) they guarantee that each job will be assigned to only one assembly machine. Constraint (17) calculates the process time of work in position k on machine m. Constraint (18) calculates the assembly time of the work in position k. In other words, constraints (17) and (18) convert the process and assembly time for work i to working time in position k. Constraint (19) prevents the tasks of different clients from being grouped together. Constraint (20) guarantee that each job falls into only one category. Constraints (21) and (22) ensure that each category is created when at least one job is assigned to it, otherwise Vjb will be zero. Constraint (23) specifies the total size of each category. The Constraint (24) specifies the number of batches after the batches are divided into parts so that if the batches have a size higher than the batch capacity, they can be divided into more batches. Constraints (7) and (8) are nonlinear, because the variable of decision Ykq is multiplied by the variables FFk and Rkq .

4 Solution Method The assembly machine is placed for delivery to the customers for stacking the parts on top of each other, which is done with the aim of achieving three criteria, minimizing delays, delivery cost and classification cost, so the above purpose is different from Non-homogeneous Delays We use the penalty rate to match the three criteria under

The Bi-level Assembly Flow-Shop Scheduling Problem

513

consideration. In this section, a mixed nonlinear integer model is formulated, in which Bank et al. (2012) showed that the two-step flow shop problem with the aim of minimizing delays is the NP-Hard problem [6, 28]. Therefore, the model of the problem we are investigating, which is much more complex, is of the NP-Hard type. And the innovation of this article is that it creates capacity for categories that are sent to customers.

5 Case Study and Performance Evaluation In this section, an example of a factory is mentioned that for the delivery of 7 types of work, each work consists of 10 sequences on 3 machines, and in the next stage, the parts are mounted on 2 assembly machines, and finally the classification is done on them. Shipped to 4 potential customers. (Capacity of each category is 20). The above problem has been solved with GAMS 24.8.2 software and with a computer with specifications, CORE i5 processor, RAM 4 and 1 GB graphics card and the BARON solver has been used. As illustrated in Table 1, if the capacity for work is considered and the categories have a certain capacity, the number of category will increase. And the model is formulated so that both results are visible once implemented by the software. Table 1. Computational results Objective function

Number of batches

Number of batches considering capacity

Total delay time

1018

4

16

133

Because the factory can use different batches but with the same capacity, the scenario by performing the problem in 150 times with different parameters to the capacity of the batch, while comparing the condition that the batches have capacity and unlimited capacity mode, the result was that Unlimited capacity for this problem is the capacity of the handle above 62 units, which is also evident in Fig. 1.

Fig. 1. Two-scale diagram of comparisons of the number of categories and the objective function to the capacity of the category

514

H. S. Garjan et al.

In fact, according to the problem data, the minimum number of categories and the objective function are 4 and 418 units, respectively, and the diagram from the capacity of 62 units upwards continues completely horizontally. According to the bubble diagram of Fig. 2, where the radius of the bubble indicates the number of categories, it is important to note that as the capacity of the category decreases, the number of categories as well as the objective function increases exponentially.

Fig. 2. Bubble diagram comparing the number of categories, the objective function and the capacity of the category

6 Conclusions In this paper, the problem of two-stage assembly workflow using batch delivery system in order to minimize the overall cost of delay and delivery cost is investigated. A new mathematical model of nonlinear integer programming is presented. Due to the fact that the software is not able to provide the model answer for large dimensions in a reasonable time, so to solve the small size of the model. Due to the approach to the reality of the mentioned issue, there was no restriction for the categories, which is not the case in reality. This is further achieved by giving volume to tasks and capacity to categories And the analysis on the results includes the points that by reducing the capacity of the category, increasing the number of categories and consequently increasing the objective function, which according to the type of factory and type of products, this model can be worked on a special case. In this regard, to get closer to reality, a routing model can be added for potential customers, and the costs of quality control or warehouse and inventory control can be very important for future studies. Acknowledgement. This research has been financially supported by The Analytical Center for the Government of the Russian Federation (Agreement No. 70-2021-00143 dd. 01.11.2021, IGK 000000D730321P5Q0002).

References 1. Lee, C.-Y., Cheng, T., Lin, B.M.T.: Minimizing the makespan in the 3-machine assembly-type flowshop scheduling problem. Manage. Sci. 39(5), 616–625 (1993)

The Bi-level Assembly Flow-Shop Scheduling Problem

515

2. Goodarzian, F., Wamba, S.F., Mathiyazhagan, K., Taghipour, A.: A new bi-objective green medicine supply chain network design under fuzzy environment: hybrid metaheuristic algorithms. Comput. Indust. Eng. 160, 107535 (2021) 3. Baker, K.R.: Introduction to Sequencing and Scheduling. John Wiley & Sons, Hoboken (1974) 4. Rahman, H.F., Janardhanan, M.N., Chuen, L.P., Ponnambalam, S.G.: Flowshop scheduling with sequence dependent setup times and batch delivery in supply chain. Comput. Indust. Eng. 158, 107378 (2021) 5. Jabbari, M., Tavana, M., Fattahi, P., Daneshamooz, F.: A parameter tuned hybrid algorithm for solving flow shop scheduling problems with parallel assembly stages. Sustain. Oper. Comput. 3, 22–32 (2021) 6. Zhang, Z., Tang, Q.: Integrating flexible preventive maintenance activities into two-stage assembly flow shop scheduling with multiple assembly machines. Comput. Indust. Eng. 159, 107493 (2021) 7. Goodarzian, F., Kumar, V., Abraham, A.: Hybrid meta-heuristic algorithms for a supply chain network considering different carbon emission regulations using big data characteristics. Soft. Comput. 25(11), 7527–7557 (2021). https://doi.org/10.1007/s00500-021-05711-7 8. Goodarzian, F., Kumar, V., Ghasemi, P.: A set of efficient heuristics and meta-heuristics to solve a multi-objective pharmaceutical supply chain network. Comput. Indust. Eng. 158, 107389 (2021) 9. Goodarzian, F., Taleizadeh, A.A., Ghasemi, P., Abraham, A.: An integrated sustainable medical supply chain network during COVID-19. Eng. Appl. Artif. Intell. 100, 104188 (2021) 10. Ghasemi, P., Goodarzian, F., Gunasekaran, A., Abraham, A.: A bi-level mathematical model for logistic management considering the evolutionary game with environmental feedbacks. Int. J. Logist. Manage. (2021). (ahead-of-print). https://doi.org/10.1108/IJLM-04-2021-0199 11. Goodarzian, F., Bahrami, F., Shishebori, D.: A new location-allocation-problem for mobile telecommunication rigs model under crises and natural disasters: a real case study. J. Ambient. Intell. Humaniz. Comput. 1−19 (2021). https://doi.org/10.1007/s12652-021-03461-w 12. Yang, S., Xu, Z.: The distributed assembly permutation flowshop scheduling problem with flexible assembly and batch delivery. Int. J. Prod. Res. 59(13), 4053–4071 (2021) 13. Basir, S.A., Mazdeh, M.M., Namakshenas, M.: Bi-level genetic algorithms for a two-stage assembly flow-shop scheduling problem with batch delivery system. Comput. Ind. Eng. 126, 217–231 (2018) 14. Ceylan, Z., Tozan, H., Bulkan, S.: A coordinated scheduling problem for the supply chain in a flexible job shop machine environment. Oper. Res. Int. Journal 21(2), 875–900 (2021). https://doi.org/10.1007/s12351-020-00615-0 15. Cheng, T., Kahlbacher, H.: Scheduling with delivery and earliness penalties. Asia-Pacific J. Oper. Res. 10(2), 145–152 (1993) 16. Goodarzian, F., Abraham, A., Fathollahi-Fard, A.M.: A biobjective home health care logistics considering the working time and route balancing: a self-adaptive social engineering optimizer. J. Comput. Des. Eng. 8(1), 452–474 (2021) 17. Kazemi, H., Mazdeh, M.M., Rostami, M.: The two stage assembly flow-shop scheduling problem with batching and delivery. Eng. Appl. Artif. Intell. 63, 98–107 (2017) 18. Potts, C.N., Sevast’Janov, S.V., Strusevich, V.A., Van Wassenhove, L.N., Zwaneveld, C.M.: The two-stage assembly scheduling problem: complexity and approximation. Oper. Res. 43(2), 346–355 (1995) 19. Mazdeh, M.M., Rostami, M.: A branch-and-bound algorithm for two-machine flow-shop scheduling problems with batch delivery costs. Int. J. Syst. Sci. Oper. Logist. 1(2), 94–104 (2014) 20. Garey, M.R., Johnson, D.S., Sethi, R.: The complexity of flowshop and jobshop scheduling. Math. Oper. Res. 1(2), 117–129 (1976)

516

H. S. Garjan et al.

21. Belabid, J., Aqil, S., Allali, K.: Solving permutation flow shop scheduling problem with sequence-independent setup time. J. Appl. Math. 2020, 1–11 (2020). https://doi.org/10.1155/ 2020/7132469 22. Taxidou, A., Karafyllidis, I., Marinaki, M., Marinakis, Y., Migdalas, A.: A hybrid firefly VNS algorithm for the permutation flowshop scheduling problem. In: Sifaleras, A., Salhi, S., Brimberg, J. (eds.) Variable Neighborhood Search: 6th International Conference, ICVNS 2018, Sithonia, Greece, October 4–7, 2018, Revised Selected Papers, pp. 274–286. Springer International Publishing, Cham (2019). https://doi.org/10.1007/978-3-030-15843-9_21 23. Wang, K., Luo, H., Liu, F., Yue, X.: Permutation flow shop scheduling with batch delivery to multiple customers in supply chains. IEEE Trans. Syst. Man Cybern. Syst. 48(10), 1826–1837 (2017) 24. Al-Behadili, M., Ouelhadj, D., Jones, D.: Multi-objective biased randomised iterated greedy for robust permutation flow shop scheduling problem under disturbances. J. Oper. Res. Soc. 71(11), 1847–1859 (2020) 25. Goodarzian, F., Hosseini-Nasab, H., Muñuzuri, J., Fakhrzad, M.B.: A multi-objective pharmaceutical supply chain network based on a robust fuzzy model: a comparison of meta-heuristics. Appl. Soft Comput. 92, 106331 (2020) 26. Ochi, H., Driss, O.B.: Scheduling the distributed assembly flowshop problem to minimize the makespan. Procedia Comput. Sci. 164, 471–477 (2019) 27. Pessoa, L.S., Andrade, C.E.: Heuristics for a flowshop scheduling problem with stepwise job objective function. Eur. J. Oper. Res. 266(3), 950–962 (2018) 28. Goodarzian, F., Ghasemi, P., Gunasekaren, A., Taleizadeh, A.A., Abraham, A.: A sustainableresilience healthcare network for handling COVID-19 pandemic Ann. Oper. Res. 1−65 (2021). https://doi.org/10.1007/s10479-021-04238-2

Building Trust with a Contact Tracing Application: A Blockchain Approach Tom´as Hon´ orio1(B) , Catarina I. Reis2 , Marco Oliveira1 , and Marisa Maximiano1 1

Computer Science and Communication Research Centre (CIIC), School of Technology and Management, Polytechnic of Leiria, Leiria, Portugal [email protected] 2 ciTechCare - Center for Innovative Care and Health Technology, School of Technology and Management, Polytechnic of Leiria, Leiria, Portugal [email protected] Abstract. On March 11, 2020, the novel coronavirus (COVID-19) was declared a global pandemic. With no treatment or vaccine available at the time, it was necessary to rely on non-pharmaceutical methods for case identification and contact tracing. This kind of approach has good results in detecting and preventing tuberculosis, sexually transmitted infections, and vaccine-preventable diseases. Contact tracing and keeping safe distances are crucial to containing the spread of COVID-19. Nonetheless, contact tracing is a complex intervention, it involves quarantining and investigating close contacts. Manual contact tracing methods are slow, require a large amount of effort, and more often than not rely on the memory or assumptions of individuals. To combat these downsides, contact tracing applications were developed, resulting in quicker and more reliable recognition of infected individuals. However, because of the complex nature of these applications and their lack of transparency, a large portion of the population started doubting the privacy of the data collected. Soon after, many of these applications started to dwindle in the user department, which caused a feedback loop. “If fewer people are using the application, the application itself becomes useless, and there is no longer a reason to use it.” Is clear that the main issue behind their downfall was an overwhelming lack of trust. In response, this paper will analyze how the use of blockchain technology can help the development of a more transparent application. And describe how a proof of work based on this concept was implemented. On the same note, it will also approach why was Hyperledger Sawtooth chosen, instead of more popular solutions such as Bitcoin or Ethereum. Keywords: COVID-19 · Contact tracing · Bluetooth · Distributed ledger · Blockchain · Bitcoin · Ethereum · Hyperledger · Hyperledger sawtooth · Directed acyclic graph

1

Introduction

Over the recent COVID-19 epoch, it has become clear that the once acclaimed tracking applications have mostly turned out to be a letdown [1]. That being c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 517–526, 2022. https://doi.org/10.1007/978-3-030-96299-9_49

518

T. Hon´ orio et al.

said, the question is why did it end up like that, and what led this type of application to be abandoned. The COVID-19 pandemic caused a global health crisis that no one was prepared to face, resulting in millions being infected and the appearance of new, more infectious strains [2]. It is now clear that there is a need for infection tracking, and since this is such a significant process, it needs to be done reliably and in real-time [3]. However, decision-making has proven to be a daunting challenge for both authorities and the public. To better understand the dynamics of this pandemic and provide effective countermeasures, data with quality is a must. That being said, the search for trustworthy data has led to the appearance of contact-tracing applications, apps that in theory should facilitate the management/tracking of new cases. However, it is apparent that in the vast majority of cases, the result was somewhat of a letdown. According to [4], the main reason for the low adherence was not technological limitations, but the concern with privacy and the negativity surrounding this type of application. One of the leading factors of this sentiment was the direct connection between these applications and big tech giants such as Google and Apple. A vast portion of contact-tracing applications was based on protocols provided by these companies, giving them the power to change them overnight. These are the same companies that come to blows repeatedly with entities like the EU Court of Justice because of privacy concerns. If possible, a user should be able to identify where and how their data is being used. The proposal is to solve the lack of trust by relying on transparent systems, such as solutions based on Distributed Ledger, more precisely Blockchain. This paper follows the order described below: – Introduction, describing the problem and the basis for the solution. – State of the art, describing how it is done and what other implementations exist. – Discussion, describing what is the reason Hyperledger Sawtooth was chosen, and how it competes with existing solutions. – Contact Tracing using Hyperledger Sawtooth, describing why and how a Contact Tracing Application based on Sawtooth works and what is their workflow. – Conclusion, the last section entails an overview of the solution proposed, and why it could be a valid solution for the problems initially raised.

2

State of the Art

To better understand the context behind this development, some key points need to be fleshed out, these are: What is a Distributed Ledger, what is Contact Tracing, what kind of application is currently available, and at last, is there any application that tries to accomplish the same goal.

Building Trust with a Contact Tracing Application: A Blockchain Approach

2.1

519

Distributed Ledger

The concept of Distributed Ledgers has been gaining attention over the past few years, in part because of the emergence/spread of new blockchain technologies based on Bitcoin [5] and Ethereum [6]. That being said, blockchain is just one of the many forms a Distributed Ledger may have, and while Bitcoin and Ethereum have been pretty popular, blockchain goes beyond that, an example of this, is something like Hyperledger. Hyperledger is an umbrella project focused on developing tools, frameworks, and libraries for enterprise-level blockchain implementations [7]. Some of the most well know projects that resulted from it are Hyperledger Fabric, Hyperledger Sawtooth, Hyperledger Grid, and many more. That being said, Hyperledger Sawtooth is of special interest, thanks to its flexibility and ease of use. Being the main advantages, the ability to specify business rules without requiring extensive knowledge of the underlying design, followed by having better parallel transaction scalability when compared with Bitcoin and Ethereum. 2.2

Contact Tracing

Contact Tracing can be described as using collected data from people diagnosed with an infectious disease, to recognize and provide support to new infectious individuals [8]. By enabling people to know that they may have been infected, it’s possible to monitor their health for symptoms. The World Health Organization (WHO) defines three crucial steps for any form of contact tracing [9]: – Identification - Upon confirmation of an infected individual, all of his contacts must be identified. This is achieved by analyzing the habits and activities of the infected individual. – Listing - All people who have come into contact with an infected person must be listed as a “Contact” and informed of their status. – Follow-up - Contacts should be observed regularly to monitor the onset of symptoms/complications. According to [10], close contacts who spent over 15 min in contact with an infected person are of special interest, since they are more likely to be infected. By testing these individuals, it’s possible to delay the progress of the transmission chain. As a benchmark for a successful contact tracing (COVID-19) operation, WHO suggests locating/quarantining 80% of close contacts within 3 days (after the first case is confirmed). According to Christophe Fraser (Oxford University), transmission is extremely fast and the virus can spread before any action is taken [11]. Even if all cases were discovered/isolated within three days, the pandemic will continue to grow. To prevent such an outcome, 70% of cases need to be isolated on the first day, only then, can the outbreak be significantly reduced. Mukhi et al. [12], advocates that traditional methods of surveillance and data collection (employing a paper-based method), pose many challenges, such as data loss, duplication, difficulty in managing databases, and lack of timely

520

T. Hon´ orio et al.

access to the data. That being said, the only way to gather such information in such a brief amount of time is by relying on technology. As mentioned by Mukhi et al. [12], “COVID-19 pandemic infection/death rates may slow down in countries with robust vaccination coverage. However, on a global scale, new mutations and new pathogens will continue disrupting society for ages to come, and therefore, it is important to keep advancing in this field”. Contact tracing applications allow for a quicker and more reliable identification of newly infected individuals. Thanks to this, the quality of the data collected is also greatly improved, being this the main purpose for the development of such apps in several countries. 2.3

Conventional Applications

Using the Portuguese application as an example. The development was a collaboration between INESC TEC, ISPUP, Keyruptive, and Ubirider, where companies and research institutions joined efforts to develop the application Stay Away Covid [13]. This application has the purpose of detecting (through the use of a smartphone) if users were near someone infected with COVID-19. It works by announcing its presence to all nearby devices, using random identifiers. By doing this, the application recognizes who has been in contact with the user, for a contact to be of interest, it needs to be within two meters and for at least 15 min [13]. In case of infection, the user will receive a code to be inserted into the application, after which they will notify the contacts. One of the key points is that it works without recording personal information, being the sole exception, infected users. However, according to the article “60% have already deleted StayAway Covid” [14], thus the application has been losing users at an alarming rate, being the peak of users during October 2020. Around the same time, a bill was drafted to mandate the use of the application [15], but later on, end up being rejected. Since then, the number of users has been steadily declining and in January 2021 only 39% of the nearly three million people who installed the application continued to use it. In hindsight, it is possible to point out two main issues: they are lack of trust/concerns with data privacy and a lack of coordination. Lack of trust can be pointed to as the fundamental issue behind the slow adherence and fast decline of the application. There are many reasons for this sentiment, but a glaring one is a reliance on technology developed by Google and Apple. Both of which are known for possessing their fair share of privacy issues [16,17]. Lack of coordination did also play a big role in the application’s failure, quoting [14]: “in the last five months, only 2708 codes were used”. The problem was caused by a lack of knowledge since many of the infected did not know where to get the codes needed for the app, and many health professionals dint know where to find them. Some examples of the overall mistrust are: – “People are losing confidence in the app because there are no codes”;

Building Trust with a Contact Tracing Application: A Blockchain Approach

521

– “And there are no codes because doctors are poorly informed about how the app works and where the codes are provided”; – “Since the application was launched, doctors have contacted us for help. It shouldn’t be like that.” 2.4

Contact Tracing Using Distributed Ledgers

According to [18], “engineers from the University of Glasgow outline how a trustworthy contact tracing system could be built on the unique properties of distributed ledger technology.” In the same article, [19] is mentioned how traditional methods pale in comparison with new, more technologically advanced ways of doing contact tracing. And how an application like the TraceTogether helped the government of Singapore to contain their coronavirus outbreak [20]. Simultaneously, it was proposed a new mobile-based system named BeepTrace [19], which hopes to “harness the unique properties of distributed ledgers to create a decentralized, potentially international system to help break the chains of virus transmission.” It works by assigning a randomly generated ID, which changes regularly to prevent tracking or identification. If a test gives a positive for COVID-19, it shares the IDs gathered by the application over the past 14 days with the ledger. Notifying any user that was in contact during the period. In the same paper, it is also mentioned how the application possesses two different modes, a passive and an active. The passive mode works by relying on GPS information to gather where the user was and who was the user in contact with. The active mode works by relying on Bluetooth to register users in close contact for a prolonged period.

3

Discussion - The Future of Contact Tracing

The Portuguese application is only one of many examples where social factors remain the driving force behind the lack of adherence. To combat the negative sentiment, some key points need to be taken into consideration, they are: – Data transparency: a user should be able to identify where and how their data is being used; – Individual control of data: a user should be able to choose what information is being shared; – Decentralized chain of power: it should not be fully reliant on the goodwill of a company (Google or Apple) to work. That being said, the proposed way to combat all these issues is by shifting the application from a centralized chain of power to a decentralized system. Where each user has control over the information and can freely interact with it, instead of it being a black box. A more effective system to accomplish this is most likely a Distributed Ledger or more accurately, an application that relies on Blockchain technology to store the user information.

522

T. Hon´ orio et al.

Although the use of distributed ledgers has already been approached in studies like the one mentioned in [19], there is still room for improvement. One of the key differences between the developed approach and the one described in the paper mentioned above is the underlying technology. This approach was based on a Chain-based solution, more precisely Hyperledger Sawtooth instead of a Directed Acyclic Graph (DAG) based solution. DAG Based Solutions - A DAG based solution represents an alternative to the traditional blockchain that aims to improve speed, scalability, and cost. However, it is still a distributed ledger. It relies on a graph structure with directed edges and where no vertex should lead back to himself. Usually, there are no blocks, being the most considerable difference the way they add transactions to the network [21], according to [22] because of the adoption of graph structure, it can do the processing of transactions in parallel. This is in direct contrast to the sequential manner used in chain-based solutions. Chain-Based Solutions vs Directed Acyclic Graph Based Solutions When comparing Chain-based solutions vs DAG-based solutions, it is possible to recognize some key differences. Chain-based solutions offer transparency and immutability, but lack scalability when it comes to performance since it often relies on consensus algorithms like Proof of Work [23]. These algorithms not only limit the number of transitions that can be processed in a given amount of time but also consume copious amounts of computing power and electricity. Establishing them as not the best-suited solution for high volumes of transactions. DAG offers more efficient scaling and the reduction/avoidance of fees, but as a result, this may make it vulnerable to attacks. That is why many of the DAG based solutions have to rely on centralized features like central co-ordinators, pre-selected validators, ‘witness’ nodes, or completely private network systems.

4

Contact Tracing Using Hyperledger Sawtooth

The goal behind this development is to have a contact tracing application capable of leveraging blockchain technology and with this, increase trust and adherence from the end-user. The key requirements for the development are: – Being able to detect if two users are in close contact; – In case of a user infection, share this information with the public ledger; – Fetch information about infections to notify its contacts. For this implementation, the blockchain technology chosen was Hyperledger Sawtooth. It offers freedom when choosing the permission level (either permissioned or permissionless). It gives the choice between a vast amount of different consensus algorithms, making sure that at least one suits the needs. Parallel processing of transactions provides a performance improvement for every workload by reducing overall latency effects, which occur when transaction execution

Building Trust with a Contact Tracing Application: A Blockchain Approach

523

Fig. 1. Sawtooth architecture diagram [24]

is performed serially. And at last, it provides an SDK in multiple languages, allowing the abstraction of much of the work. By looking at the Fig. 1, it is possible to see the proposed architecture for this development. On the left, there are two users in contact, both with the app installed and the Bluetooth enabled, followed by two arrows repressing the IDs being shared between one another. If at any moment, one of the users is diagnosed with COVID-19 and it decides to share the information, the application will send a request with all IDs gathered in the past 14 days into the Validator Node, being their final destination the Sawtooth Network. 4.1

Detection of Close Contacts

As the name entails, the primary function behind the application is to contact trace, something that is achieved by making use of proximity tracking. More accurately, it uses the AltBeacon protocol [25] in order to gather the distance between two users. For this to work, the app relies on a beacon advertising in the background, and a tracker to identify other users. The content of the emitted message allows the receiving device to get the user ID and compute the relative distance between each other [25]. This ID is generated every couple of hours to keep the user’s privacy. For a new contact to be recorded a couple of procedures need to happen: First, the user needs to have done the application setup (Application installed, Bluetooth turned on, etc.). Second, the user needs to be in contact with another user that has also completed the setup. Third, for a contract to be deemed of interest, it needs to be for an extended period (around 15 min) and within two meters of distance. Afterward, this information is stored locally in a database for at least 14 days. The reason that lead to this choice, was it being the minimum amount of time to ensure no symptoms were present, and also, being the suggested time for a person to be in self-quarantine after exposure to an infected individual [26].

524

4.2

T. Hon´ orio et al.

Identifying and Sharing in Case of Infection

For a user to be deemed infected, he’s required to have a positive test result for COVID-19. Then, a code provided by the health authorities allows the user to share its state information with the remaining network. This is achieved by providing the ledger with the IDs collected over time by the user. Something that is done by doing a request to the Sawtooth Rest API. After this, the request is validated by the Sawtooth Validator and afterward provided to the Transaction Handler. The Transaction Handler is what contains the business logic for a particular family of transactions. And on the same note, transaction families can be seen as the way Hyperledger Sawtooth implements smart contracts. That being said, once the Transaction Handler gets a hold of the information, it will confirm if everything is alright, and if so, save this new information into the ledger state. 4.3

Notifying About Possible Infections

Fig. 2. Notification process

After everything is set and done, ensuring that everyone that was in contact with the infected user is rightfully notified is of the utmost urgency. To achieve this, the application makes use of a handy feature provided by Sawtooth called Sawtooth Events. They occur when new blocks are committed and result from the validator broadcasting an event when a commit operation succeeds. By using Sawtooth Events, it’s possible to notify the user about what exactly was changed inside the ledger, and with this on the same note filter only the relevant information for this use case. That being said, once the information of a new infection is added to the Ledger, everyone can fetch it and see if they are part of the affected users. If so, the app will send a notification to the user, asking for this to go to a testing center, to be tested for COVID-19. As is shown on Fig. 2, the steps are, first an infected user is found, then it shares the collection of IDs with one of the Validator Nodes, followed by, them being added to the ledger. Once this is done, all users that need to be notified will be notified to get tested.

5

Conclusion

Although there are multiple examples of contact tracing applications, it is clear that there was a widespread lack of trust in them. This resulted from issues like

Building Trust with a Contact Tracing Application: A Blockchain Approach

525

the heavenly reliance on private companies like Google and Apple, the lack of transparency about how the data is being used, the inability to choose what was shared, and sometimes trying to impose its use without addressing the underlying issues. This lack of trust was the main reason for the lack of adoption by the general population, and the fast decline in users after the first issues became widespread. Resulting in a feedback loop where the fewer users actively using it, the less useful it is, followed by even fewer users using it. By using blockchain technology, it is possible to solve many of the concerns related to the mistreatment of data, allowing every user to check what information is being used and how is it being handheld. It also needs to be pointed out that Hyperledger is a significant performance improvement in the handling of parallel transactions when compared with Bitcoin and Ethereum. With the appearance of new alternatives like for example Hyperledger Sawtooth, it is now possible to develop large-scale applications without having to fear performance issues or suffering the shortcomings of alternative technologies like DAG-based solutions. By making use of the Sawtooth SDK, it is possible to ease the development cost, since many of the modules needed for such applications are already present. As a future improvement, the addition of location data would be an interesting challenge, since there are so many ways to go about his issue. He can approach this by gathering GPS data, like how it is done in the BeepTrace project [19], or by creating a location journal, where points of reference (e.g. Restaurants and malls) are stored by making use of QR-codes. In the end, the development of a contact tracing app allowed for a better understanding of the workflow involved in contact tracing and the challenges related to a distributed solution. Acknowledgments. This publication is funded by FCT-Fundac˜ ao para a Ciˆencia e Tecnologia, I.P., under the project UIDB 045242020.

References 1. Chan, E.Y., Saqib, N.U.: Privacy concerns can explain unwillingness to download and use contact tracing apps when COVID-19 concerns are high. Comput. Human Behav. 119, 106718 (2021) 2. CDC: What you need to know about variants, September 2021. https://www.cdc. gov/coronavirus/2019-ncov/variants/variant.html. Accessed 9 Oct 2021 3. Shelby, T., et al.: Lessons learned from COVID-19 contact tracing during a public health emergency: a prospective implementation study. Front Public Health 9, 721952 (2021) 4. Bambauer, J., Ray, B.: Covid-19 apps are terrible — they don ’ t have to be, November 2020. https://s3.documentcloud.org/documents/20424830/bambauerand-ray-final-2.pdf. Accessed 26 Sept 2021 5. Nakamoto, S.: Bitcoin: a Peer-to-Peer electronic cash system 6. Ethereum whitepaper. https://ethereum.org/en/whitepaper/. Accessed 9 Oct 2021 7. Bhanushali, H., Arthena, A., Bhadra, S., Talukdar, J.: Digital certificates using blockchain: an overview, April 2019

526

T. Hon´ orio et al.

8. COVID19-contact-tracer-508.pdf, July 2020 9. Infection prevention and control: contact tracing, May 2017. https://www.who. int/news-room/q-a-detail/contact-tracing. Accessed 26 Sept 2021 10. Lewis, D.: Why many countries failed at COVID contact-tracing — but some got it right, December 2020. https://www.nature.com/articles/d41586-020-035184. Accessed 5 Oct 2021 11. Abueg, M., et al.: Modeling the combined effect of digital exposure notification and non-pharmaceutical interventions on the COVID-19 epidemic in Washington state, September 2020 12. Mukhi, S., et al.: An innovative mobile data collection technology for public health in a field setting, September 2018 13. e ISPUP ..., V.D.I.T.: AIPD STAYAWAY v2.0 09 2020.pdf, August 2020 14. 60% j´ a apagaram a StayAway covid: s˜ ao 1,8 milh˜ oes de portugueses — sa´ ude ´ — PUBLICO, January 2021. https://www.publico.pt/2021/01/15/tecnologia/ noticia/60-ja-apagaram-stayaway-covid-sao-18-milhoes-portugueses-1946366. Accessed 5 Oct 2021 15. Proposta de lei n. 62/xiv, October 2020 16. Shead, S.: Apple accused of breaching European privacy law by French start-up group. https://www.cnbc.com/2021/03/09/apple-accused-of-breachingeu-privacy-law-by-french-start-up-group.html. Accessed 9 Oct 2021 17. Satariano, A.: Google is fined $57 million under Europe’s data privacy law. The New York Times, January 2019 18. Glasgow-University: University news, September 2020. https://www.gla.ac.uk/ news/headline 752925 en.html. Accessed 26 Sept 2021 19. Xu, H., Zhang, L., Onireti, O., Fang, Y., Buchanan, W.B., Imran, M.A.: BeepTrace: blockchain-enabled privacy-preserving contact tracing for COVID-19 pandemic and beyond, May 2020 20. Yuen-C, T.: More than 4.2m people using TraceTogether, token distribution to resume soon: Lawrence wong, politics news & top stories - the straits times, January 2021. https://www.straitstimes.com/singapore/politics/parliament-morethan-42m-people-using-tracetogether-token-distribution-to-resume. Accessed 5 Oct 2021 21. Das, V.K.: Role of directed acyclic graphs in the blockchain landscape, September 2020. https://www.blockchain-council.org/blockchain/role-of-directedacyclic-graphs-in-the-blockchain-landscape/. Accessed 27 Sept 2021 22. Yang, W., Dai, X., Xiao, J., Jin, H.: LDV: a lightweight DAG-Based blockchain for vehicular social networks. IEEE Trans. Veh. Technol. 69(6), 5749–5759 (2020) 23. Nehra, V., Sharma, A.K., Tripathi, R.K.: Blockchain Implementation for Internet of Things Applications, pp. 113–132. Unknown, January 2020 24. About sawtooth events — sawtooth latest documentation. https://sawtooth. hyperledger.org/docs/core/nightly/1-2/app developers guide/about events.html. Accessed 27 Sept 2021 25. spec: AltBeacon technical specification. Accessed 20 Aug 2021 26. CDC: Contact tracing for COVID-19, August 2021. https://www.cdc.gov/ coronavirus/2019-ncov/php/contact-tracing/contact-tracing-plan/contacttracing.html. Accessed 27 Sept 2021

Immunity Passport Ledger Digital Certificates Implemented on a Permissioned Blockchain Marco Oliveira1(B) , Tom´as Hon´ orio1 , Catarina I. Reis2 , and Marisa Maximiano3 1

Polytechnic of Leiria, Leiria, Portugal {2192406,2190338}@my.ipleiria.pt 2 ciTechCare - Center for Innovative Care and Health Technology, School of Technology and Management, Polytechnic of Leiria, Leiria, Portugal [email protected] 3 Computer Science and Communication Research Centre (CIIC), School of Technology and Management, Polytechnic of Leiria, Leiria, Portugal [email protected]

Abstract. The global outbreak of Coronavirus (SARS-CoV-2) which in 2020 reached pandemic scale, has been a central topic of debate in our society. Concerns over the ease of transmission of the infection led to the imposition of measures restricting freedom such as curfews, lockdown, general confinement, and closure of trade. Technology was one of the tools used to resist to the spread of the disease using applications that, on one hand, track contacts to warn users that were close to someone infected and, on the other hand, provide immunity digital certification. Despite the relevance of these options, end users have no confidence, transparency, and responsibility that the registration and use of their health data are ethical, secure, anonymous, and available through verifiable credentials and, most importantly, is being used for its main purpose. Consequently, a solution based on a distributed ledger technology, such as blockchain, is introduced to assure the trustworthiness and integrity of user’s data. Since the proposed application embraced user privacy, we conducted a comparative study between permissioned blockchains, that includes an authorization abstraction layer and ensures that certain actions can only be performed by identifiable participants. We concluded that Hyperledger Fabric was an option that fulfilled all the requirements to develop a platform for the immunity passport ledger. Its modularity and versatility accommodates the needs that were initially proposed for the development of a proof of concept. The work leads us to propose that further research be conducted regarding scalability and performance evaluation.

Keywords: Immunity passport ledger ledger technology · Hyperledger fabric

· Blockchain · Distributed

c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 527–536, 2022. https://doi.org/10.1007/978-3-030-96299-9_50

528

1

M. Oliveira et al.

Introduction

On 31 December 2019, the first case of a new type of coronavirus was reported in Wuhan, China. On 30 January 2020, World Health Organization declared coronavirus outbreak a public health emergency of international concern. On 11 March 2020, the outbreak reached a pandemic scale, with 118 000 cases reported in 114 countries [1]. Since then, there has been an ongoing debate about how best to overcome the pandemic. Concerns about the ease of spread of the infection led to the imposition of measures restricting freedom such as lockdowns, curfews, general confinement, and trade closure. Technology was one of the tools used to resist to the transmission of the disease through applications such as STAYAWAY COVID [2] that tracks contacts and notifies users that they were close to someone infected. These applications raised concerns about user’s data privacy and some countries were considering making these applications mandatory, which harmed their adoption by the population [3]. In December 2020 a global scale vaccination was started and a return to normality is expected. One of the options to help in this situation of “back to normal” was the creation of immunity passports that will allow people to prove their health status. This option was questioned and even contested by the World Health Organization [4]. This is why it is imperative to present an Immunity Passport that is transparent and secure that end users can trust, where the data is anonymous and used only for its purpose. This paper is organized as follows. Section 2 presents an overview of Distributed Ledger Technologies, review, and analysis of related work. Section 3 gives the specifications for the implemented proof of concept. Section 4 summarizes the concerns we aimed to solve, the problems with similar implementations, and future work on the subject.

2

Related Work

Digital health certification is already a reality worldwide and are used broadly in European flights, through the Digital Green Certificate [5]. Travelling in Asia demands the CoWIN application [6]. They both provide digital and interoperable ways of validating vaccination certificates. The private sector and other non-government organizations have urged the development of worldwide standardized vaccine certificates such as the CommonPass [7], and the IATA Travel Pass [8]. The use of blockchain technology is already gaining traction with the AOKpass [9] and Immunitee [10] as it is safe against tampered certificates or tests results.

Immunity Passport Ledger

2.1

529

Distributed Ledger Technologies

Distributed Ledger Technology (DLT) can be defined by the technological infrastructure and protocols that support concurrent access, validation, and state updating in an immutable manner across a network that’s spread across multiple participants or nodes [11]. A DLT system has multiple participants which reach a settlement over a set of distributed data and its validity, in the absence of a central authority. What separates a DLT system from a traditional distributed database are the core features capable of transacting data and maintaining data integrity in the presence of malicious actors actively attempting to attack the network [12]. DLT has great potential to disrupt the way governments, institutions, and corporations work. It can help governments with tax collection, the issuance of passports, recording land registries and licenses, and the outlay of Social Security benefits as well as voting procedures [13]. Industries such as finance, music and entertainment, art, and supply chains of various commodities (including diamonds and other precious assets) are the early adopters of this technology [14]. There are multiple types of DLT such as DAG, Hashgraph, Holochain, Radix [15], and one of the most well-known: Blockchain. Big corporations such as IBM, Intel and Microsoft are exploring the technology and some of the distributed ledger protocols. Ethereum, Hyperledger Fabric, R3 Corda, and Quorum are amongst the most popular alternatives [11]. Blockchain is a DLT where transaction records are registered and kept in the ledger as a chain of blocks. This technology is attracting a great deal of attention propelled by the success of Bitcoin [16], launched in 2009 and triggering a large number of projects in different industries, with finance being the one that leads the use of this technology due to the success of cryptocurrencies. This technology underlies Bitcoin and has the potential to support a wide variety of business processes. Ethereum launched in 2015 by Vitalik Buterin [17] extends Blockchain concepts from Bitcoin but introduces smart-contracts that make possible execution of code in a decentralized way [18]. Hyperledger was launched in 2016 and is an industry-wide open-source initiative to advance blockchain technology, governed by The Linux Foundation [19]. Permissioned and private Blockchain differ from public Blockchains of Bitcoin and Ethereum. These Blockchain applications rely on trust relationships between participant organizations, with the need to share data with a greater degree of security. Contract privacy is a mandatory requirement for enterprise applications, and using a permissioned or private blockchain ensures it. Data for these Permissioned blockchains can be used to record promises, trades, transactions, or any data that can’t be lost without needing to run a Proof-of-Work mechanism [20]. Hyperledger Fabric is an enterprise-grade open-source permissioned blockchain framework for developing solutions and applications with a modular architecture. Its modular and versatile design satisfies a broad range of industry

530

M. Oliveira et al.

use cases [21–23] and allows components, such as consensus and membership services, to be plug-and-play. The unique approach to the consensus mechanism enables performance at scale while preserving privacy [24]. The consensus mechanism has a fundamental role in the transaction flow of Fabric that goes through the process of a transaction proposal, endorsement, ordering, validation, and commitment [19]. Block sequencing and transaction sequencing within blocks are established when the ordering service first creates blocks. Each block contains a sequence of transactions representing a query or update to the world state. Fabric has been designed at its core to have a modular architecture and meet the diversity of enterprise use case requirements. It allows pluggable identity management protocols such as LDAP or OpenID Connect, and the ledger supports a variety of DBMSs, and pluggable endorsement and validation policy enforcement [25]. The performance of Fabric has been further improved [26], during the release of new versions, with a substantial increase in performance on Fabric 2.0 [27]. The performance of a blockchain is subjected to many implementation variables such as transaction size, block size, network bandwidth, hardware limitations, consensus algorithm, caching, and parallelism, among others [28,29]. Out of the box with simple configurations, Fabric can support 3000 transactions per second, and with advanced tweaks. A study [30], demonstrates that is possible to scale Fabric performance up to 20000 transactions per second. 2.2

Applications

Digital green certificate is a collaborative effort of health care authorities across the EU and consists of a Private Key Infrastructure (PKI) solution with a single authority. This solution is not ideal since the key pairs are not issued to single health care professionals but rather by big organizations. A single breach of the private key will bring down every certificate signed by that authority. Then, all the certificates will be mark as untrusted, which risks blocking traveling in Europe. Malicious hackers are targeting these keys because of their value to generate new fake certificates and sell them on the black market [5]. CoWIN is the platform created by the Indian government to manage the appointment of vaccinations and the emission and validation of vaccine certificates. This solution is also based on a PKI system and allows offline verification. The infrastructure is centralized and under a unique central authority [6]. CommonPass is a digital health pass developed by The Commons Project, a non-profit organization with support from the Rockefeller Foundation [7]. This application serves as proof of vaccination for a wide range of airlines. The certificate is stored on the local device. IATA Travel Pass is a mobile app that stores and manages verified certifications for tests or vaccines [8]. It gives the user a digital wallet to store all their travel documentation, including biometric passports. This application

Immunity Passport Ledger

531

allows travelers to plan their journeys accordingly to the health conditions of the destination country. AOKpass [9] is a platform and mobile application using blockchain technology, enabling users to verify their health status with third parties while preserving the privacy of their underlying personal health data. Users have exclusive control over their health data, such as health certificates or test results. They are stored only on the user’s mobile device and never on any external database or centralized system. AOKpass saves a hash on the Ethereum public blockchain network for the certificate data of the users. Immunitee Health Passport is the Malaysia health passport accepted into Singapore [10]. It was designed to store personal immunization records and vaccine data with the Unifier platform, allowing interoperability to securely share the necessary data with the various national health check systems being put globally. Immunitee stores all patient data hashed on a public blockchain system, ensuring that data cannot be tampered, is protected, and belongs to the user. Verifiers can only obtain information by scanning a QR Code that holds all the relevant testing and vaccination data and can only be unlocked using a secret key that belongs to the user. IBM Digital Health Pass uses a blockchain framework, Hyperledger Fabric. The implementations consist of administration authorities providing X.509 certificates to healthcare organizations. Then they use their private key to sign the user’s health certificates. Only the keys supplied to healthcare organizations are stored on the ledger. Verifier uses a QR code reader to extract and decode the issuer identifier. Verifier queries the ledger to obtain the public key associated with issuer identification. Using the latter, the verifier checks the certificate’s digital signature and gets assured that the claimed issuer indeed generated the health certificate [31,32]. A summary of the related work can be found in Table 2 from Appendix A.

3

Immunity Passport Ledger

The proposed solution for an Immunity Passport Ledger considers the chief requirement of allowing a consortium of government and private agencies to collaborate through a single system to store and validate the information. Most importantly, it requires keeping some information hidden from consortium participants while allowing data to be audited and managing access to the data. Hyperledger Fabric provides flexibility, modularity, scalability, and performance. These characteristics allow our main use cases: – Allow health organizations and vaccination centers to issue and verify certificates. – Allow individuals to choose with whom they share their information with. – Make information verifiable to a broad range of organizations. – Allow statistics organizations to get insights on the issued and expired certificates.

532

3.1

M. Oliveira et al.

Use Case

Application use cases start after the user physically visits an authorized issuer organization, provide a traditional mean of identification with health number and perform a test or vaccination. User data and validity of immunity are submitted to the ledger, and a hash is generated. Issuance of a QR Code in a digital and/or paper format containing the hash that validates the immunity of the user. A user presents the QR Code to an authorised verifier organisation member. Then, the verifier scans the QR Code and obtains the information on the validity of the immunity and the name of the person associated with the certificate, and the person presents an identifying document to prove certification ownership. If, by any chance, a person loses his certificate, a reissue request is available through the health organization platform. 3.2

Data Access

Access to data across the blockchain participants ranges accordingly to the needs of the work performed by each organization as shown on Table 1. Health organizations have access to all the data and are the only organization allowed to submit new transactions on the ledger. Verifier organizations will only perform read operations on the ledger and only need access to immunity status and name to validate if the certificate is valid or invalid and user ownership of the certificate. Statistic organizations will perform read operations to get insights on immunity status and validity of the population. Table 1. Data access Data

Health org Verifier org Statistics org

Immunity date

X

Immunity validity X

3.3

Name

X

Health number

X

X X

X

X

Network Architecture

The network for the application (Fig. 1) consists of a single channel common on all organizations with the main purpose of conducting transactions, an orderer service, a peer for each organization (that holds a copy of the ledger) with a world state database and the executable chaincode to perform actions on the ledger. The policy rules are specified on the channel configuration represented by the CC1 in the Fig. 1. All organizations in the network endorse these rules.

Immunity Passport Ledger

533

Health organizations will keep user-sensitive data in a private state database off-chain known as world state database. This Fabric functionality allows all channel participants to see a transaction while keeping a portion of the data private. Private data will be transmitted peer-to-peer between authorized organizations only, via the gossip protocol. A hash of the data, is endorsed, ordered, and written to the ledgers of every peer that participates on the channel. Hash is used as evidence of the transaction, is used for state validation, and for audit purposes.

Fig. 1. Proposed network architecture.

3.4

Data Structure

CouchDB holds the world state database and allows ledger states to be structured as JSON documents. This optimization enables JSON queries against stored data values instead of the default approach of using LevelDB, where its primary purpose is to query the keys. Regarding the data structure, it consists of a JSON object that represents the user certificate. The data ensures the connection between the individual and the immunity validation (certificate status). It also allows statistic organizations to perform metric reports based on dates and health organizations to send certificate expiration notices in advance. 3.5

Application Structure

The proposed application has two main components: a web and a mobile application for organizations to manage and issue certificates and an additional component that holds the chaincode installed at the organization’s nodes. The two applications invoke the chaincode to update and read the ledger.

534

M. Oliveira et al.

Applications use Hyperledger Fabric SDK to interact with the Fabric blockchain network. It provides a gateway module to manage the network interactions and an API to submit transactions or query the ledger. The application stores multiple wallets on the filesystem for each health professional with identities to connect to the network. Health organizations’ platforms allow health care professionals to register user data on the blockchain and generate certificates. The interface provides authentication to health care professionals and a mechanism to generate certificates. After submission to the blockchain, a QR code is generated from the hash and sent to the patient. Verifiers organizations chaincode search on blockchain by a hash and return valid and health number or not valid. Verifiers have an Android or iOS application that reads QR code and sends the hash to a backend that invokes the smart contract. Then it shows the health number and validity status, or not found in the case certificate doesn’t exist on the ledger. Statistics organizations chaincode receives a date interval and returns the number of certificates expiring between that date also returns the number of certificates issued between that date. This web platform generates reports and allows analysts to visualize data about issued and expired certificates.

4

Conclusion

Providing privacy to a massive number of certificates issued while ensuring scalability can be challenging. Some of the existing centralized solutions on the cloud can be costly and delegate the data ownership to third parties. Infrastructures with a single point of failure are a significant issue that can suspend global traveling. Private Key Infrastructure solutions aim to solve this problem and verify certifications without an Internet connection, but this raises a more significant issue with a flood of tampered certificates online due to leaks of private keys. Usage of public blockchains can be expensive due to transactions costs and do not assure privacy. We developed a proof of concept that aims to solve these issues using permissioned blockchain technology with a distributed ledger between multiple organizations that agree on a single source of truth. In future work, we must test the performance and scalability of the proof of concept application simulating a real scenario of use at a global scale.

Immunity Passport Ledger

535

Appendix A Table 2. Comparison table of related work Solutions

Available Governmental Infrastructure

Data privacy Offline mode

D.G.C.a

Yes

Yes

PKI

Yes

Yes

CommonPass Yes

No

Local

Yes

Yes

IATA T.P.

Yes

No

Local

Yes

Yes

AOKpass

Yes

No

Blockchain

Yes

Yes

Immunitee

Yes

Yes

Blockchain

Yes

No

Blockchain/PKI Yes

Yes

IBM D.H.P. Yes No a Digital Green Certificate

References 1. WHO: WHO Director-General’s opening remarks at the media briefing on COVID19, 11 March 2020 (2020). https://www.who.int/director-general/speeches/detail/ who-director-general-s-opening-remarks-at-the-media-briefing-on-covid-19---11march-2020 2. STAYAWAY COVID - Fique longe da COVID num clique (n.d.). https:// stayawaycovid.pt/. Accessed 2 Aug 2021 3. Miranda Ramos, L.F.: Digital contact tracing and data protection: assessing the French and Portuguese applications. UNIO - EU Law J. 6(2), 35–48 (2020). https://doi.org/10.21814/unio.6.2.2767 4. “Immunity passports” in the context of COVID-19 (n.d.). https://www.who.int/ news-room/commentaries/detail/immunity-passports-in-the-context-of-covid-19. Accessed 13 Oct 2021 5. EU Digital COVID Certificate — European Commission (n.d.). https://ec. europa.eu/info/live-work-travel-eu/coronavirus-response/safe-covid-19-vaccineseuropeans/eu-digital-covid-certificate en#does-it-matter-which-vaccine-citizensreceived. Accessed 24 Aug 2021 6. CoWIN (n.d.). https://www.cowin.gov.in/. Accessed 24 Aug 2021 7. CommonPass — Digital Health App (n.d.). https://commonpass.org/. Accessed 24 Aug 2021 8. IATA - Travel Pass Initiative (n.d.). https://www.iata.org/en/programs/ passenger/travel-pass/. Accessed 24 Aug 2021 9. AOKpass: Secure - Private - Portable (n.d.). https://www.aokpass.com/. Accessed 24 Aug 2021 10. Malaysia’s Immunitee Health Passport gains Singaporean verification (n.d.). https://focusmalaysia.my/malaysias-immunitee-health-passport-gainssingaporean-verification/. Accessed 24 Aug 2021 11. Frankenfield, J.: Distributed Ledger Technology Definition. Investopedia, October 2018. https://www.investopedia.com/terms/d/distributed-ledger-technologydlt.asp 12. Rauchs, M., et al.: Distributed ledger technology systems: a conceptual framework. SSRN Electr. J. (2018). https://doi.org/10.2139/ssrn.3230013 13. Hj´ almarsson, F., Hreioarsson, G.K.: Blockchain-Based E-Voting System (n.d.) 14. McLean, S., Deane-Johns, S.: Demystifying blockchain and distributed ledger technology - hype or hero? Comput. Law Rev. Int. 17(4), 1–8 (2016). https://doi.org/ 10.9785/cri-2016-0402

536

M. Oliveira et al.

15. Different types of DLTs and how they work — by TerraGreen — Medium (n.d.). https://medium.com/@support 61820/different-types-of-dlts-and-howthey-work-cfd4eb218431. Accessed 25 Aug 2021 16. Velde, F.R.: Bitcoin - A Primer. Chicago Fed Letter, 1–4 December 2013. http:// search.ebscohost.com/92563197&site=ehost-live 17. Ethereum Whitepaper — ethereum.org (n.d.). https://ethereum.org/en/ whitepaper/. Accessed 13 Oct 2021 18. Buterin, V.: A next-generation smart contract and decentralized application platform. Etherum, 1–3 January 2014. http://buyxpr.com/build/pdfs/ EthereumWhitePaper.pdf 19. Hyperledger - Open Source Blockchain Technologies (n.d.). https://www. hyperledger.org/. Accessed 16 Aug 2021 20. Androulaki, E., et al.: Hyperledger fabric: a distributed operating system for permissioned blockchains. In: Proceedings of the 13th EuroSys Conference, EuroSys 2018. ACM (2018). https://doi.org/10.1145/3190508.3190538 21. Uddin, M.: Blockchain Medledger: hyperledger fabric enabled drug traceability system for counterfeit drugs in pharmaceutical industry. Int. J. Pharm. 597, 120235 (2021). https://doi.org/10.1016/j.ijpharm.2021.120235 22. Spengler, A.C.F., Souza, P.S.L.de.: Avalia¸ca ˜o de desempenho do Hyperledger Fabric com banco de dados para o armazenamento de grandes volumes de dados m´edicos, pp. 61–72 (2021). https://doi.org/10.5753/wperformance.2021.15723 23. Pajooh, H.H., Rashid, M., Alam, F., Demidenko, S.: Hyperledger fabric blockchain for securing the edge internet of things. Sensors (Switzerland) 21(2), 1–29 (2021). https://doi.org/10.3390/s21020359 24. Chacko, J.A., Mayer, R., Jacobsen, H.A.: Why do my blockchain transactions fail?: A study of hyperledger fabric. In: Proceedings of the ACM SIGMOD International Conference on Management of Data, vol. 221, pp. 221–234 (2021). https://doi.org/ 10.1145/3448016.3452823 25. Mohammed, A.H., Abdulateef, A.A., Abdulateef, I.A.: Hyperledger, Ethereum and Blockchain Technology: A Short Overview, 1–6 June (2021). https://doi.org/10. 1109/hora52670.2021.9461294 26. The Ordering Service - hyperledger-fabricdocs master documentation (n.d.). https://hyperledger-fabric.readthedocs.io/en/release-2.2/orderer/ordering service.html. Accessed 16 Aug 2021 27. Dreyer, J., Fischer, M., T¨ onjes, R.: Performance analysis of hyperledger fabric 2.0 blockchain platform. In: CCIoT 2020 - Proceedings of the 2020 Cloud Continuum Services for Smart IoT Systems, Part of SenSys 2020, pp. 32–38 (2020). https:// doi.org/10.1145/3417310.3431398 28. Nguyen, M.Q., Loghin, D., Tuan, T., Dinh, A.: Understanding the scalability of hyperledger fabric (n.d.). https://github.com/quangtdn/caliper-plus 29. Xu, X., Sun, G., Luo, L., Cao, H., Yu, H., Vasilakos, A.V.: Latency performance modeling and analysis for hyperledger fabric blockchain network. Inf. Process. Manage. 58(1), 102436 (2021). https://doi.org/10.1016/j.ipm.2020.102436 30. Gorenflo, C., Lee, S., Golab, L., Keshav, S.: FastFabric: scaling hyperledger fabric to 20 000 transactions per second. Int. J. Netw. Manage. 30(5) (2020). https:// doi.org/10.1002/NEM.2099 31. Digital Health Pass — IBM (n.d.). https://www.ibm.com/products/digital-healthpass. Accessed 28 Aug 2021 32. Androulaki, E., et al.: IBM digital health pass: a privacy-respectful platform for proving health status whitepaper (i), 1–10 (2021)

Computer Graphics Rendering Survey: From Rasterization and Ray Tracing to Deep Learning Houssam Halmaoui1,2(B) and Abdelkrim Haqiq2 1

ISMAC - Higher Institute of Audiovisual and Film Professions, Rabat, Morocco [email protected] 2 Hassan First University of Settat, Faculty of Sciences and Techniques, Computer, Networks, Mobility and Modeling Laboratory: IR2M, 26000 Settat, Morocco [email protected]

Abstract. In this article we present a survey of the different techniques of rendering of 3D computer generated images. We start with the principles and advances of the traditional methods of rasterization and ray tracing. Then, we discover the new techniques based on deep learning, which are now part of a new discipline of computer graphics called neural rendering, allowing the synthesis and rendering of 3D images, thanks to generative adversarial network and variational auto encoder models. Finally, we compare theses approaches according to different criteria. Keywords: Generative adversarial networks · Variational autoencoders · Neural rendering · Rasterization · Ray tracing Photorealistic images

1

·

Introduction

Computer graphics or computer generated images are terms used interchangeably to designate 3D images created on a computer. Today, one can hardly imagine a world without this type of images which have become omnipresent in the media and our daily means of communication, but also of crucial importance in several fields ranging from industry such as aviation or automotive, to the visualization of scientific calculation results or medical images, as well as virtual reality, video games, animated films and visual effects. For a long time, the generation of these images necessarily involved the use of modeling tools. But thanks to the progress made in deep learning, a new discipline called neural rendering [26] allows to automatically generate photorealistic 3D images with a certain degree of control over the geometric and photometric parameters. However, as we will discover, these techniques do not yet completely replace the traditional methods of rasterization [4] and ray tracing [5] which in turn have taken advantage of the parallel computing advances of the new GPUs to improve rendering quality and computation time. c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 537–548, 2022. https://doi.org/10.1007/978-3-030-96299-9_51

538

H. Halmaoui and A. Haqiq

The main challenge in computer graphics is to create photo-realistic images at low cost in terms of computing [17]. For this, the different techniques try to approximate the solution of the transport equation, which models the radiance reflected from a point of an object to the observer. The exact computer modeling of this equation is still impossible despite the computing power we have today, because of the very large number of photons whose interaction with all objects of the scene before reaching the camera must be simulated. However, with approximations, it is possible to obtain images indistinguishable from photographs. In the following, we will introduce the rasterization pipeline and how it approximates light with ambient, diffuse and specular shading models for different types of materials and lighting. We will then discover the pipeline and the evolution of ray tracing techniques, and how they simulate lighting effects to increase the level of photo-realism. Concerning the methods based on deep learning, we will present the generative adversarial networks (GAN), which have revolutionized synthetic images, and their use with variational auto encoders (VAE) in the field of neural rendering in order to generate 3D models and render them in a photo-realistic way from a new point of view or a new lighting position. A comparison is made according to various criteria of photo-realism, controllability, computation and working times, accessibility, quality of details and applications. In Sect. 2, we present the transport equation. In Sect. 3, we discover the rasterization pipeline. In Sect. 4, we present the ray tracing pipeline. In Sect. 5, we discuss the principles of GANs and VAEs. In Sect. 6, we discover neural rendering. In Sect. 7, we make a comparison of the different approaches.

2

Transport Equation

Rendering is the process of projecting a computer-generated 3D model onto a 2D screen. The color of each pixel of the 2D rendered image will depend on the position, direction and type of lighting, the geometry and materials of the scene, and the position and direction of observation (camera). The transport equation (or rendering equation) is used to calculate for an object point P , the radiance Lo (pixel color) in a direction wo (towards the camera):  Lo (P, wo ) =

S

f (P, wo , wi )Li (P, wi )| cos(θi )|dwi ,

(1)

with f the bidirectional reflectance distribution function (BRDF) describing the surface properties of the object (material), wi the direction of the incident light, and θi the angle between the incident light and the surface normal. The term cos(θi ) models the geometric attenuation as a function of the angle. To obtain a photo-realistic rendering, it is thus necessary to integrate the incident lights coming from all sources S, from all directions, towards each point of the scene in order to estimate its color (radiance). In practice, despite the computing power we have today, it is impossible to simulate on a computer the movement of each photon that leaves the lighting and interacts with several elements of the scene before reaching the observer, because of the very large number of photons generated by all the lighting. However, it is possible to create photorealistic scenes using approximate models.

Computer Graphics Rendering Survey

3

539

Rasterization

Rasterization is the rendering technique that for each vertex of the 3D model, visible by the camera, calculates its pixel color using a model of shading (lighting). The rasterization pipeline is illustrated in the Fig. 1.

Fig. 1. Rasterization pipeline.

Shaders correspond to GPU programmable steps. The vertex shader receives the geometric data from the CPU and transforms it into the position relative to the camera. The tessellation and geometry shaders are optional. The tessellation shader allows to generate more polygons by vertex combinations. The geometry shader allows to manipulate the primitives (vertexes, polygons and lines) in several ways (stretching, holes, etc.) to create more complex models. The rasterizer transforms the polygons into pixel fragments by projecting them onto the image with the camera model, and also allows to interpolate any output variable of the vertex shader (position, color, etc.). The fragment shader calculates the color of the pixels according to the lighting, materials and texture, using a shading model. During display, a hidden surface removal (HSR) algorithm (non-programmable) allows to display only visible vertices by comparing their distances to the camera. OpenGL is the standard API for shader programming. Texture Mapping: It is the simplest way to add realism, often used in video games and animations to paint faces, skin and clothes. Textures to be applied to 3D models are stored in images. We call texels the pixels of the textures to differentiate them from the pixels of the rendered image. For each vertex (x, y, z) of the 3D model, we assign the corresponding texel coordinates (s, t) of the color to display. The texel coordinates can be assigned manually in the case of simple shapes (cube, pyramid), or algorithmically for curved geometries (sphere, torus), or in the case of complex models (human, objects) with “UV-Mapping” tools in dedicated software (Maya, Blender). Texture mapping often produces blurring, distortion and aliasing artifacts, due to the difference in resolution and aspect ratio between the model and the texture [21]. Mipmapping uses textures at different resolutions to apply the closest to each image region. Blur and distortion are corrected with anisotropic filters and perspective correction algorithms. Lighting and Materials: To give more realism to the rendering, rasterization uses shading models (or reflection models) that calculate the color of pixels according to the lighting and materials of the objects. The most commonly used shading model to approximate light (transport equation) is the ambient, diffuse and specular model (see [6] and [15]). An ambient reflection affects all elements in the scene equally. A diffuse reflection reflects light in all directions depending on

540

H. Halmaoui and A. Haqiq

the angle of incidence. A specular reflection gives the object a shiny appearance depending on the angle of the observer and the angle of incidence. The lighting and the material both have properties that reflect these three characteristics that need to be specified. The material has in addition a shininess characteristic. On the lighting side, there are four types of lighting sources: – Global ambient: lighting whose source and direction are indeterminate, its contribution is equal at any point of the scene. It corresponds in fact to the light that has bounced several times through the objects of the scene. It is modeled by an RGBA value. – Directional (or distant:) lighting whose position is also indeterminate but it has a direction. It corresponds to very distant lighting (e.g. the sun), whose rays are considered parallel. It is modeled by four vectors: the ambient, diffuse and specular RGBA values, and the direction vector. – Positional: lighting whose position is known but its direction will depend on each vertex of the scene. It corresponds to nearby sources (lamps, candles). It is modeled by the ambient, diffuse and specular RGBA values and the position vector. We can also add an attenuation factor. – Spotlight: light that has both a direction and a position. It is modeled by a cone of half angle θ between 0 and 90◦ . The intensity factor is computed by cosf (φ), with φ the angle between the direction D of illumination and the direction V to the vertex, and f is the falloff exponent that simulates the variation of the illumination along the cone. The intensity factor is multiplied by the illumination intensity to simulate the spotlight cone effect. We therefore specify for the spotlight the ambient, diffuse and specular RGBA values, and the direction, position, falloff f and angle θ. Lighting interacts differently with each material. On the material side, we must specify the ambient, diffuse and specular RGBA values as well as the shininess. Some materials that emit their own light may have an emissiveness parameter. Lighting of the Scene: We calculate the color of each pixel by: I = Iamb + Idif f + Ispec ,

(2)

where Iamb (Idif f or Ispec ) is the sum of the ambient (diffuse or specular) intensis s s (Idif ties Iamb f or Ispec ) of the different light sources s. The ambient, diffuse and specular contributions of the light sources and materials are included according to Eqs. 3, 4 and 5. The Fig. 2 illustrates the different parameters. s = Lsamb Mamb , Iamb

(3)

with Lsamb and Mamb the ambient lighting and material coefficients. s s Idif f = Ldif f Mdif f cos(θ),

with Lsdif f and Mdif f the diffuse lighting and material coefficients, θ the angle between the surface normal N and the incident light L , and cos(θ) = max((N • L), 0).

(4)

Computer Graphics Rendering Survey

s Ispec = Lsspec Mspec max((R • V )n , 0),

541

(5)

with Lsspec and Mspec the specular coefficients of lighting and material, R the direction of reflection, V the direction of observation, and n the shininess coefficient which models the attenuation as a function of the observation angle.

Fig. 2. Light reflection illustration.

The calculation of the normal for each pixel is time consuming. Smooth shading techniques (Gouraud or Phong) compute the color only for the pixels corresponding to the vertex, and deduce by interpolation the color of the other pixels. There are several ways to combine the color Itex of the texture image with the shading model. The simplest way is to use a weighted average as in the Eq. 6. There are several variants depending on the desired effect. I = 0.5 · Itex + 0.5 · (Iamb + Idif f + Ispec )

(6)

Shadow, Bump Mapping and Height Mapping: Shadows mapping are important for depth perception and photo-realism (see [13] and [9]). Hard shadow can be estimated with projective shadow matrix, or shadow volumes that reduce the intensity of shadow regions, or shadow mapping which uses the HSR algorithm by positioning the camera in place of the lighting. Soft shadows are usually calculated by percentage closer filtering (PCF) which estimates the percentage that a region is in shadow. In the case of irregular surfaces, texture mapping is ineffective when the camera and lighting move [22]. Bump mapping consists in simulating the effect of lighting on irregular surfaces, without modifying the geometry of the model, but by modifying the normal to the polygons. Height mapping consists in using an image map of the height of the irregularities to be applied on the model to modify the position of the vertices. However, bump mapping allows to capture more details because it is executed at pixel level.

4

Ray Tracing

The power of ray tracing lies in its simplicity. The basic principle illustrated in the Fig. 3 consists of sending a ray from the pixel and finding the point of intersection with the nearest object. For shading, another ray is sent from the hit point to the lighting to determine whether the object is lighted or in shadow.

542

H. Halmaoui and A. Haqiq

Fig. 3. Ray tracing illustration.

A ray is a half 3D line starting from a point A (the origin) towards a point B. It is defined by the parametric equation: R(t) = (1 − t)A + tB,

(7)

The parameter t is used to determine the hit point. This equation is often written: R(t) = O + t · d,

(8)

with O the origin (point A) and d the direction (B-A). So in order to shade a point P of the scene with lighting at position L, we send a shading ray from the origin P in the direction d = L − P . If the intersection occurs in the interval t ∈ [0, 1], the ray hits a geometry that blocks the light. In practice, we consider an interval [tmin , tmax ] because of the imprecision of float calculations, with tmin =  and tmax = 1 − . To take into account more real effects of lighting and materials, there are different variants and implementations of ray tracing [23]. The goal is to obtain approximations of the transport equation. Whitted’s algorithm considers punctual light sources. If the surface is diffuse, we send a shading ray to each lighting source. If the surface is specular, we send a reflection ray. Generally, we send several rays per pixel for anti-aliasing. Cook’s algorithm (or stochastic ray tracing) is the one that revolutionized cinema by adding several effects. The rays are launched randomly out of the specular direction (Fig. 4 on the left) to create motion blur effects, and randomly towards the lighting to simulate non punctual lighting. The brute force aspect of launching several rays per pixel (anti-aliasing), and several rays per hit point, and towards each lighting source, makes the technique very time consuming. Kajiya’s algorithm (or path tracing) modified Cook’s algorithm, by applying the principle of random ray launching for diffuse surfaces (Fig. 4 in the middle), and introduced the notion of path tracing which consists in sending from each hit point a single ray but in a direction chosen by the monte carlo algorithm. A path (Fig. 4 on the right) corresponds to the path followed by the ray from the pixel to the lighting. For each pixel several paths are generated to improve the radiance estimation, a compromise must be found between the number of rays and the denoising to speed up the processing and to have a photo-realistic rendering. Denoising remains important for ray tracing acceleration despite the computing power of recent GPUs. The GPU programming of ray tracing is done on five shaders [5]: – Ray generation shader: transformation of pixels into rays.

Computer Graphics Rendering Survey

543

Fig. 4. Random ray casting for specular and diffuse surfaces, and generation of 2 paths.

– Intersection shader: intersection code of the rays with the geometry. Each geometry (sphere, triangle, bezier curve, etc.) requires a different code. – Miss shader: what to do if no intersection is found. – Closest Hit shader: what to do in case of intersection (shading model). – Any hit shader: adding effects such as transparency.

5

Generative Adversarial Network and Autoencoders

In deep learning, we distinguish between discriminative models that allow to separate data into classes, and generative models that allow to generate data of a certain class. Given an image X of an object and Y its class, a discriminator allows to predict the probability P (Y /X) that X belongs to Y. Conversely, a generator allows to generate a synthetic image X belonging to a class Y by maximizing the probability P (X/Y ). The most popular generative models are the variational autoencoder (VAE) and the generative adversarial networks (GAN) whose architectures are illustrated on the Fig. 5. During the training, VAE encodes the image in a “Latent” vector which in turn feeds a decoder to generates a realistic image. During the test phase we only use the decoder by supplying it with a random vector which allows us to generate new random realistic images. The encoders and decoders are generally convolutional neural networks (CNN). GANs have revolutionized the field of image synthesis by allowing the generation of ultra-realistic images automatically and quickly. They are used in many image and video applications such as face creation [8], image translation [16] which allows the creation of photorealistic scenes from a coarse segmentation, portrait animation [30], or 3D objects generation [28]. A GAN consists of two models: a generator G and a discriminator D, which are generally CNN models. An analogy from the initiator of GAN [3] compares G to an art forger and D to an inspector who tries to detect forgeries. The two models, with a competitive game, learn from each other and improve their performance. After a while, it becomes impossible to differentiate the counterfeit from the original. The GAN system is illustrated in Fig. 5. G is equivalent to the decoder in VAE, except that it is not guided by an encoder and receives a random noise at its input. We

544

H. Halmaoui and A. Haqiq

Fig. 5. VAE architecture and the training of the generator and the discriminator.

alternate the training of D and G. To train D, we first feed G with a noise vectors  to generate fake images X ∗ . D is then fed with X ∗ as well as real images (with the corresponding classes). The θd parameters of D are optimized by minimizing a binary cross entropy (BCE) loss function: m

1  i J =− [Y log(h(X i , θ) + (1 − Y i )log(1 − h(X i , θ)], m i=1

(9)

with m the number of samples in the batch, X i the image sample and Y i the corresponding label, and h(X i , θ) the prediction of X i with the parameters θ. The training of G (Fig. 5 at the bottom right) is performed as before, except that D is only provided with fake images X ∗ but with “real” labels, because we want G to learn to fool D in order to generate realistic images. Therefore, the θg parameters of G are optimized by maximizing the BCE. In order to avoid the collapse mode and vanishing gradient problems, due to the fact that D learns faster than G, and thus the training of G does not progress any more, we can use a deep convolutional GAN (DCGAN) [19] or replace the BCE by Wasserstein loss function (WCGAN) [1]. The generation can be done in three different ways: – Unconditional: image synthesis of a random class. – Conditional: the class to generate is specified at the input of G and D. – Controllable: specification of a characteristic in the content of the image to be generated (object position, gaze direction, lighting direction, etc.). The controllable GANs [12] are the ones we are interested in here, in order to be able to control the object, camera and lighting parameters in the same way as rasterization or ray tracing. To modify a feature in the generated image, we only need to modify a value in the latent (noise) vector. But we need to know which value of the latent vector corresponds to the feature we want to modify. Usually, a classifier is used at the output of the generator to detect the feature we want to control, and penalize the generator for each image not containing this feature. To avoid the features being correlated, a large latent vector is used.

Computer Graphics Rendering Survey

6

545

Neural Rendering

Neural rendering algorithms allow, by combining machine learning techniques with the physical aspect of computer graphics, to generate controllable photorealistic renderings and 3D models accessible in virtual environments. The principle is to train models to generate images from different specified parameters of lighting, geometry, material and camera. Novel View Synthesis: Novel view synthesis consists to perform, from one or more images of a scene, a photo-realistic rendering by simulating a camera motion. In [2], a generative query network (GQN) takes as input images of a scene acquired from different viewpoints, and allows to generate a new point of view. The GQN consists of an encoder that extracts a representative vector of the scene from the observed images, and a generator that takes as input the representative vector, as well as the desired camera position and orientation, and generates the corresponding viewpoint. In [14], a VAE is trained to generate, from images of different viewpoints, a 3D model which is used to generate an image of a new viewpoint by ray tracing given the camera parameters. Relighting and Materials: Relighting consists to generate, from images acquired with different lighting positions, an image with any lighting position. This will allow, in combination with novel view synthesis, to control both the lighting and the camera position. The simplest approach is to first reconstruct the geometry and the properties of the materials in the scene, and then perform the rendering by changing the lighting position. However, the reconstruction approach is difficult for complex materials and geometries. These methods (see [20] and [27]), use the transport equation 1, where function f can be sampled using different lighting positions. The image illuminated from a new viewpoint can be obtained by interpolating f . This methods allows to obtain very photorealistic results, but requires several input images (between ten and hundreds). In [29], a model allows to estimate the image of a scene illuminated from a new point of view, from only five images of different lighting. For this, a VAE takes as input the images of the different lighting, and the latent vector is modified with the new lighting direction, in order to generate the corresponding image as output. Also, in order to select only the relevant viewpoints (not very close) a first network is optimized to perform this operation. Other methods based on VAE allow a more specific relighting of portrait [25], body [7], or outdoor scene [18]. Content Control: For geometry control, lot of work are done for 3D faces control. In [10] the used model takes in input two video portraits, source and target, and allows to transfer from the source to the target, both the head pose, the facial expressions, and the gaze movement. For this, an input model is used to extract from the source and the target, the parameters to be transferred as well as the fixed parameters of illumination and identity. A model use the new parameters to generate by rasterization a 3D model of the target. Finally, a VAE is trained to generate a photo-realistic image from the 3D model. A similar technique with a new architecture is presented in [11]. In [30], a GAN-based model allows to generate new face positions from landmarks. In [24] the same principle of using a VAE allows to generate different types of objects.

546

7

H. Halmaoui and A. Haqiq

Comparison

Rasterization allows a good quality, but is inferior to ray tracing because of the approximations made on the shading model which generate artifacts and do not allow to simulate all the reflection effects needed for photorealism. However, rasterization is easier to implement in real time on any hardware, and it is often used in video games and the web, but also for animated movies because of the faster rendering. Ray tracing has long been difficult to implement in real time, so it has been used offline for applications such as movies, but nowadays, it is possible thanks to the power of recent GPUs. Ray tracing still remains the main solution to create photorealistic content. We have also seen that rasterization and ray tracing allow to control all the parameters of the scene (lighting, object and camera) but require a lot of work time (geometrical and material modeling, lighting, texturing, animation, etc.) and rendering calculation time. On the other hand, GAN techniques are faster and automatic. To make GAN controllable, classifiers are used to detect the characteristics (age, hair, etc.) that we want to control, but the problem is that we need a classifier for each characteristic and it must be reliable to not mislead the GAN. Neural rendering allows to generate 3D models from real images, and to perform photo-realistic rendering from any camera position or lighting. However, this technique requires very large datasets for training. Also, quality losses are observed when moving the camera and lighting simultaneously. Therefore, these methods do not allow a total control of the scene as in rasterization and ray tracing. However, neural rendering is very relevant for real-time applications such as video-conferencing or virtual reality, or for off-line applications such as visual effects. The Fig. 6 shows a comparison of the four approaches according to the criteria discussed above.

Fig. 6. Comparative table: rasterization, ray tracing, GAN and neural rendering.

Computer Graphics Rendering Survey

8

547

Conclusion

We presented the different approaches of computer graphics rendering, explaining the pipelines of traditional rasterization and ray tracing techniques, as well as the recent GAN and neural rendering techniques based on deep learning that have revolutionized and changed the paradigms in the field of image synthesis, thanks to the ultra-realistic quality and the automatic generation of images. However, deep learning techniques are still limited in terms of scene control, and do not allow to manipulate objects, camera and lighting with the same ease of traditional techniques. The advantage of these methods is that the work time is much lower than traditional methods that require learning a modeling tool and creating geometric models manually. Therefore, deep learning methods are very relevant for some applications such as virtual reality. In the same way rasterization remains the most used technique in video games, and ray tracing is the main technique for photorealistic scenes and which is constantly developing thanks to the power of new GPUs.

References 1. Arjovsky, M., Chintala, S., Bottou, L.: Wasserstein generative adversarial networks. In: International Conference on Machine Learning, pp. 214–223. PMLR (2017) 2. Eslami, S.A., et al.: Neural scene representation and rendering. Science 360(6394), 1204–1210 (2018) 3. Goodfellow, I., et al.: Generative adversarial nets. Adv. Neural Inf. Process. Syst. 27 (2014) 4. Gordon, V.S., Clevenger, J.L.: Computer Graphics Programming in OpenGL with C++. Stylus Publishing, LLC (2020) 5. Haines, E., Akenine-M¨ oller, T.: Ray Tracing Gems: High-Quality and Real-Time Rendering with DXR and Other APIs. Apress (2019) 6. Halmaoui, H., Haqiq, A.: Matchmoving previsualization based on artificial marker detection. In: International Conference on Advanced Intelligent Systems and Informatics, pp. 79–89. Springer (2020) 7. Kanamori, Y., Endo, Y.: Relighting humans: occlusion-aware inverse rendering for full-body human images. arXiv preprint arXiv:1908.02714 (2019) 8. Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen, J., Aila, T.: Analyzing and improving the image quality of stylegan. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8110–8119 (2020) 9. Kessenich, J., Sellers, G., Shreiner, D.: OpenGL Programming Guide: The Official Guide to learning OpenGL, version 4.5. Addison-Wesley Professional, Boston (2016) 10. Kim, H., et al.: Deep video portraits. ACM Trans. Graph. (TOG) 37(4), 1–14 (2018) 11. Koujan, M.R., Doukas, M.C., Roussos, A., Zafeiriou, S.: Head2head: video-based neural head synthesis. In: 2020 15th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2020), pp. 16–23. IEEE (2020) 12. Lee, M., Seok, J.: Controllable generative adversarial network. IEEE Access 7, 28158–28169 (2019)

548

H. Halmaoui and A. Haqiq

13. Liu, N., Pang, M.Y.: A survey of shadow rendering algorithms: projection shadows and shadow volumes. In: 2009 Second International Workshop on Computer Science and Engineering. vol. 1, pp. 488–492. IEEE (2009) 14. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.: Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751 (2019) 15. Marschner, S., Shirley, P.: Fundamentals of Computer Graphics. CRC Press, New York (2018) 16. Park, T., Liu, M.Y., Wang, T.C., Zhu, J.Y.: Semantic image synthesis with spatially-adaptive normalization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2337–2346 (2019) 17. Pharr, M., Jakob, W., Humphreys, G.: Physically Based Rendering: From Theory to Implementation. Morgan Kaufmann, San Francisco (2016) 18. Philip, J., Gharbi, M., Zhou, T., Efros, A.A., Drettakis, G.: Multi-view relighting using a geometry-aware network. ACM Trans. Graph. 38, 1–14 (2019) 19. Radford, A., Metz, L., Chintala, S.: Unsupervised representation learning with deep convolutional generative adversarial networks. preprint arXiv:1511.06434 (2015) 20. Ren, P., Dong, Y., Lin, S., Tong, X., Guo, B.: Image based relighting using neural networks. ACM Trans. Graph. (ToG) 34(4), 1–12 (2015) 21. Segal, M., Akeley, K.: The Opengl Graphics System: A Specification. version 4.6, core profile. The Khronos Group Inc., 2006-2018 (2020) 22. Sellers, G., Wright Jr, R.S., Haemel, N.: OpenGL superBible: comprehensive tutorial and reference. Addison-Wesley (2013) 23. Shirley, P.: Ray Tracing in One Weekend. Amazon Digital Services LLC 1, Seattle (2016) 24. Sitzmann, V., Thies, J., Heide, F., Nießner, M., Wetzstein, G., Zollhofer, M.: Deepvoxels: learning persistent 3d feature embeddings. In: Proceedings of the IEEE/CVF CVPR, pp. 2437–2446 (2019) 25. Sun, T., et al.: Single image portrait relighting. ACM Trans. Graph. 38(4), 1–12 (2019) 26. Tewari, A., et al.: State of the art on neural rendering. Comput. Graph. Forum 39, 701–727 (2020) 27. Wang, J., Dong, Y., Tong, X., Lin, Z., Guo, B.: Kernel nystr¨ om method for light transport. In: ACM SIGGRAPH 2009 Papers, pp. 1–10 (2009) 28. Wu, J., Zhang, C., Xue, T., Freeman, W.T., Tenenbaum, J.B.: Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In: Proceedings of the 30th International Conference on Neural Information Processing Systems, pp. 82–90 (2016) 29. Xu, Z., Sunkavalli, K., Hadap, S., Ramamoorthi, R.: Deep image-based relighting from optimal sparse samples. ACM Trans. Graph. 37(4), 1–13 (2018) 30. Zakharov, E., Shysheya, A., Burkov, E., Lempitsky, V.: Few-shot adversarial learning of realistic neural talking head models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 9459–9468 (2019)

A Sentiment-Based Approach to Predict Learners’ Perceptions Towards YouTube Educational Videos Rdouan Faizi(B) ENSIAS, Mohammed V University in Rabat, Rabat, Morocco [email protected]

Abstract. Given the increasing proliferation of educational resources on YouTube, this social media platform has become an optimal space to which students usually resort to improve their learning in various subjects. It has also turned into an online community in which all users can voice their opinion about videos they have watched. However, getting the perceptions of learners on these learning materials is not always an easy task. Our objective in this paper is, therefore, to propose a machine learning approach that can classify learners’ feedback as being positive or negative. For this purpose, textual data was retrieved from YouTube educational videos, pre-processed and then classified using different machine learning algorithms. Results of the present study demonstrated that the Support Vector Machines algorithm performs better than other models. In fact, with an accuracy score of 92.82% on combinations of unigrams and bigrams and 92.67% for associations of unigrams and trigrams, this supervised machine learning approach outperforms other implemented models, namely Naïve Bayes, Random Forest and Logistic Regression. Keywords: Sentiment analysis · Education · Machine learning · YouTube · Learners · Feedback

1 Introduction Given the growing amounts of data that have been generated on the Web over the last two decades, sentiment analysis has gained a lot of attention among researchers and businesses. Sentiment analysis (also termed opinion mining) denotes the process of retrieving subjective information from textual data by resorting to various techniques, namely Natural Language Processing (NLP), text mining and computational linguistics [1, 2]. Its main task is to analyze the reviewers’ opinions or perceptions towards different entities such as products, services, events or topics. Specifically, it aims at detecting whether an opinion is positive, negative or neutral. Sentiment analysis is carried out by using three major approaches: (i) machine learning (ii) lexicon-based and (iii) hybrid [3]. Machine learning approaches are used to predict the polarity of sentiments on the basis of train and test datasets. Lexicon-based approaches do not require any prior training but just a predefined list of words, where © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 549–556, 2022. https://doi.org/10.1007/978-3-030-96299-9_52

550

R. Faizi

each item is related to a specific sentiment. As for hybrid approaches, they combine machine learning and lexicon-based techniques. Sentiment analysis is mostly used by businesses in domains such as e-commerce, marketing and hospitality so as to improve customer experience and, thus, gain a competitive advantage [4, 5]. Recently, sentiment analysis applications have started finding their way into more and more domains, one of which is education [6–9]. In fact, given the growing proliferation of social media platforms in education settings [10], the comments and questions/answers posted by learners on these virtual spaces has witnessed an increasing growth [11, 12]. This textual feedback has, thus, become valuable data that can be used to improve the teaching/learning experience [13–16]. In an attempt to extract value from the textual data that students or learners post on YouTube, our objective in the present article is to propose a sentiment-based classifier that can predict the learners’ perceptions towards educational videos. This analysis will enable us to measure the students’ satisfaction of YouTube courses and estimate the performance of teachers or content creators. The remainder of this article is structured as follows. Section 2 justifies our choice of YouTube as source of students’ feedback. Section 3 presents our sentiment-based approach. Finally, Section 4 summarizes the findings of this work.

2 YouTube as an Educational Platform Although YouTube has been widely used as an entertainment outlet, via which users can watch, share and upload a wide range of video content, this video sharing platform has lately developed into an optimal destination to which nearly every student or teacher have recourse to thanks to the numerous educational opportunities that it presents [17–20]. Probably the most important educational benefit of YouTube is that it hosts a huge amount of valuable learning and teaching resources simply at the click of a button. In fact, with the shift of the Web from a ‘read-only’ website into a ‘read-write’ channel, any internet user can create content on YouTube [21, 22]. Accordingly, this social media platform has turned into a precious virtual source of open educational resources that can meet the needs of students or teachers. Thus, for any student aspiring to acquire a new language or enhance his computer skills, or for any faculty member who would like to enrich his teaching materials and enhance his pedagogical methodologies, YouTube is undoubtedly the optimal venue. These easily accessed educational videos can be used as autonomous tutorials or as a supplement to teaching or learning materials within and beyond classroom walls. The second educational asset of YouTube is that the videos, which this platform is bristling with, have been found out to be a compelling method of learning as they improve the transmission of knowledge and facilitate the comprehension of complex concepts. Given the fact that some topics or issues are not usually easy to explain or comprehend, resorting to learning via videos provides a dynamic and stimulating learning environment where students can better understand and retain information longer. Following Moss [23], a learning video benefits from the power of visual perception in new ways as it helps the learner or student visualize how a process works. For their parts, Goodyear and Steeples [24] noted that video resources provide vibrant descriptions to articulate

A Sentiment-Based Approach to Predict Learners’ Perceptions

551

unspoken knowledge that is difficult to describe. Given all these potentials, YouTube learning videos can be highly effective in capturing the students’ attention, increasing their understanding and in boosting their learning experience [25]. Such Videos can also enable both students and teachers to overcome the practical limitations of the real world and explore the boundless opportunities of digital spaces. An additional advantage of using YouTube as an educational space is that it provides students with a flexible and convenient learning environment. Unlike conventional classroom-based learning/teaching which requires the actual presence of a teacher and students, YouTube-based instruction provides students with a flexible learning experience, allowing them to get engaged with their studies from anywhere and at any time. [26, 27]. Hence, regardless of time and space constraints, students have the total freedom to access YouTube educational videos and as many times as they want until they fully grasp the information being presented and better achieve the desired learning outcomes. Consequently, rather than being forced to attend a course at a fixed time on a specific day, the student can study at his own pace. This learning flexibility offers students more opportunities to fit their studies around other activities. It also gives them a higher level of autonomy in regulating their learning, which could help them become independent learners. This students’ learning independence can free faculty members up to support individual students who may need customized attention. Besides the opportunities that have been mentioned above, YouTube can also operate as a strong e-learning community where students and all other members of this web-based learning platform can share their thoughts [28, 29]. Users can not merely upload, watch or share video educational resources but are also allowed to leave feedback on video content. Given these potentials, this video sharing platform can promote interaction and communication not only between students and instructors but also among students. Thus, after viewing a video learning resource, the student can voice his opinion about it. He can also ask questions, seek help if he has trouble understanding a particular part or issue in the video or can assist his peers who have learning difficulties. Similarly, instructors can, via the comments’ feature that YouTube incorporates, provide the required explanation or answer to any question the students may ask. All these interactions, be they between teachers and students or amongst learners, are likely to improve the learning process. In addition, the feedback from students can be made use of in a formative and developmental way so as to encourage reflection and enhance teaching practice [30, 31]. Given all the educational opportunities that it offers for both students and instructors, YouTube has turned into an open space in which a lot of textual feedback is generated. To extract value from this valuable data, we propose, in the section that follows, a sentiment classifier of YouTube comments.

3 Proposed Classifier In the present section we present a detailed description of the methodology used for the sentiment classification of YouTube comments. This includes the various natural language processing techniques used to eliminate less informative data from the dataset, feature extraction, the different machine learning algorithms used to classify sentiments as well as the results of the implementation of these models on the comments written about educational YouTube videos.

552

R. Faizi

3.1 Data Collection and Preprocessing To perform sentiment analysis on YouTube educational textual data, we automatically scraped comments that have been posted on different educational videos. Before proceeding with data annotation, we started by eliminating duplicated comments as well as reviews that are not written in English. These comments were then manually annotated as positive or negative. After annotation, we managed to retrieve 5969 valid comments: 3340 positive and 2599 negative. Once the comments had been annotated, this textual data had to go through the text pre-processing stage. Pre-processing is one of the essential steps in any NLP task. In fact, to ensure the performance of sentiment classifiers, data should be cleaned and transformed into a format that can be easily processed and predicted by machine learning algorithms. Pre-processing includes the following tasks: • Tokenization: Breaking sentences into tokens. • Letter casing: Converting all letters to lower case. • Noise removal: deleting unwanted characters (e.g. HTML tags, punctuation, special characters, numbers, white spaces, etc.) • Stop word removal: Eliminating words that do not impact sentiment and do not considerably contribute to the machine learning model. • Stemming: chopping off affixes from words in order to get stems. • Lemmatization: reducing words to their bases or dictionary forms. After pre-processing, data needs to be converted into numerical feature vectors for the model to easily process it. For this purpose, we used Term Frequency-Inverse Document Frequency (TF-IDF) as a feature extraction technique. This calculates how significant a word is in a document compared to the whole corpus. 3.2 Classification Algorithms Machine learning algorithms are often divided into supervised and unsupervised. Supervised learning is a category of machine learning algorithms that go through two stages: a learning phase in which the classifier is trained on a given dataset, and the evaluation phase in which the performance of that classifier is tested. In what follows, we will shed light on some classification learning algorithms that have been widely in sentiment analysis and implemented in the present work. • Naive Bayes (NB) is a probabilistic classification method which is built on the principle of class conditional independence of the Bayes Theorem. In simple terms, the Naive Bayes classifier assumes that all the features are unrelated to each other. The absence or presence of a feature is totally independent of the existence of any other feature. This algorithm predicts membership probabilities for every class, and the one that has the maximum probability is regarded as the most probable class. • Random Forest (RF) is a learning method that is composed of a large number of decision trees functioning as an ensemble. Random Forest fits decision trees on indiscriminately selected subsets of the data, gets prediction from each decision tree and

A Sentiment-Based Approach to Predict Learners’ Perceptions

553

then selects the results that are more accurate and stable by means of voting. The class with the highest votes becomes the model’s prediction. • Support Vector Machine (SVM) is a supervised machine learning algorithm which is founded on the concept of decision planes that define decision boundaries. This algorithm performs classification by finding the hyperplane that best splits a given dataset into two classes with the help of support vectors. The best hyperplane or decision boundary is the one with the maximum margin between both classes when input data is separated by a linear function. • Logistic Regression (LR) is a classification algorithm that is used to attribute observations to a discrete group of classes. It is most generally used when the data in question has binary output (e.g., “yes/no” “positive/negative”). As opposed to linear regression that produces continuous number values, logistic regression transforms the output by making use of the sigmoid or (logistic) function to give a probability value that can be mapped to two or more discrete classes. 3.3 Results and Discussion Before feeding the data into the machine learning algorithms examined above, our data was split into training and testing sets. 80% of the data was used as a training set to fit and tune our model and 20% as a testing set to create predictions on and to evaluate the model at the very end. Moreover, to assess the performance of these algorithms, we made use of accuracy, which is the ratio of the number of correct predictions over the total number of predictions, as an evaluation metric. The implementation of these supervised techniques on our dataset has yielded the following results. Table 1. Results based on accuracy and time Model

N-grams

Accuracy

Time

NB

Unigram + Bigram

92.34%

139 ms

Unigram + Trigram

92.00%

146 ms

Unigram + Bigram

90.82%

350 ms

Unigram + Trigram

90.74%

338 ms

SVM

Unigram + Bigram

92.84%

45.1 s

Unigram + Trigram

92.67%

45.1 s

RF

Unigram + Bigram

89.22%

6.57 s

Unigram + Trigram

89.98%

6.15 s

LR

As shown in Table 1 above, the highest accuracy score is associated with SVM algorithm. In fact, with a score of 92.84% for combinations of unigrams (one word) and bigrams (two contiguous words) and 92.67% grouping unigrams and trigrams (three adjacent words), this model has outperformed all other algorithms. With a value not far from that of SVM, the multinomial Naïve Bayes classifier gives an accuracy of

554

R. Faizi

92.34% for unigrams and bigrams and 92% for unigrams and trigrams. Despite the good performance of SVM, its training time is longer than that of other models. In addition to SVM and NB, Logistic Regression has also shown good results in classifying YouTube comments. Despite the outstanding accuracy scores of the three classifiers (i.e. SVM, NB and LR), data processing has revealed that the non-removal of stop words negatively impacts the performance of all models. Moreover, though stemming and lemmatization both serve to generate the canonical form of inflected words, it has been found out that while the second text normalization technique shows better results with SVM, stemming outperforms lemmatization in other models.

4 Conclusion Our objective in this paper was to propose a sentiment classifier that can predict learners’ perceptions towards YouTube educational videos. In this respect, based on the implementation of different supervised machine algorithms on the learners’ comments that we retrieved from YouTube, it has been revealed that SVM does well compared to other machine leaning algorithms. The proposed classifier can, thus, enable teachers to gain valuable insights into how their students are learning. Indeed, they will not be able to get the opinions or sentiments of students on the video learning materials, but also get information on issues that their students master and on the specific areas that they need to work on. Using these valuable insights are likely to enhance teaching strategies and students’ learning experience.

References 1. Liu, B.: Sentiment Analysis: Mining Opinions, Sentiments, and Emotions. Cambridge University Press, Cambridge (2020) 2. Pozzi, F.A., Fersini, E., Messina, E., Liu, B.: Sentiment Analysis in Social Networks. Morgan Kaufmann, Burlington (2016) 3. Berry, M.W., Mohamed, A., Yap, B.W. (Eds.): Supervised and Unsupervised Learning for Data Science. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-22475-2 4. Faizi, R., El Fkihi, S., Ezzahid, S.S., El Afia, A.: Using sentiment analysis to derive business value. In: Proceedings of the 32nd International Business Information Management Association (IBIMA), 15–16 November 2018, Seville, Spain (2018). ISBN: 978-0-9998551-1-9 5. Faizi, R., El Fkihi, S., El Afia, A., Chiheb, R.: Extracting business value from big data. In: Proceedings of the 29th International Business Information Management Association (IBIMA), 3–4 May 2017, Vienna, Austria, pp. 997–1002 (2017). ISBN: 978-0-9860419-7-6 6. Colace, F., Casaburi, L., De Santo, M., Greco, L.: Sentiment detection in social networks and in collaborative learning environments. Comput. Hum. Behav. 51, 1061–1067 (2015) 7. Barrón Estrada, M.L., Zatarain Cabada, R., Oramas Bustillos, R., Graff, M.: Opinion mining and emotion recognition applied to learning environments. Expert Syst. App. 150, 113265 (2020) 8. Zhou, J., Ye, J.M.: Sentiment analysis in education research: a review of journal publications. Interact. Learn. Environ. 1−13 (2020). https://doi.org/10.1080/10494820.2020.1826985

A Sentiment-Based Approach to Predict Learners’ Perceptions

555

9. Barrón Estrada, M.L., Zatarain Cabada, R., Oramas Bustillos, R., Graff, M.: Opinion mining and emotion recognition applied to learning environments. Expert. Syst. App. 150, 113265 (2020) 10. Faizi, R., El Fkihi, S.: Analyzing students’ perceptions towards using Facebook as a learning platform. In: Proceedings of IADIS International Conference e-Learning 2018, Madrid, Spain (2018) 11. El Fkihi, S., Ezzahid, S.S., Faizi, R., Chiheb, R.: Formative assessment in the era of big data analytics. In: Proceedings of the 32nd International Business Information Management Association (IBIMA), ISBN: 978-0-9998551-1-9, 15–16 November 2018, Seville, Spain (2018) 12. Souali, K., El Afia, A., Faizi, R., Chiheb, R.: A new recommender system for e-learning environments. In: 2011 International Conference on Multimedia Computing and Systems, pp. 1–4. IEEE (2011) 13. Menaha, R., Dhanaranjani, R., Rajalakshmi, T., Yogarubini, R.: Student feedback mining system using sentiment analysis. IJCATR 6, 1–69 (2017) 14. Eng, T.H., Ibrahim, A.F., Shamsuddin, N.E.: Students’ perception: student feedback online (SuFO) in higher education. Procedia Soc. Behav. Sci. 167, 109–116 (2015) 15. Sangeetha, K., Prabha, D.: Sentiment analysis of student feedback using multi-head attention fusion model of word and context embedding for LSTM. J. Ambient. Intell. Humaniz. Comput. 12(3), 4117–4126 (2020). https://doi.org/10.1007/s12652-020-01791-9 16. Singh, L. K., & Devi, R. R.: Student feedback sentiment analysis: a review. In: Materials Today: Proceedings (2021) 17. Faizi, R.: Moroccan higher education students’ and teachers’ perceptions towards using web 2.0 technologies in language learning and teaching. Knowl. Manage. E-Learn. Int. J. (KM & EL). 10(1), 86–96 (2018) 18. Snelson, C.: The Benefits and Challenges of YouTube as an Educational Resource. The Routledge Companion to Media Education, Copyright, and Fair Use (2018) 19. Vieira, I., Lopes, A.P., Soares, F.: The potential benefits of using videos in higher education. In: Proceedings of EDULEARN 2014 Conference, pp. 0750–0756. IATED Publications (2014) 20. Faizi, R.: Teachers’ perceptions towards using Web 2.0 in language learning and teaching. Educ. Inf. Technol. 23(3), 1219–1230 (2017). https://doi.org/10.1007/s10639-017-9661-7 21. Schicchi, D., Marino, B., Taibi, D.: Exploring learning analytics on YouTube: a tool to support students’ interactions analysis. In: International Conference on Computer Systems and Technologies 2021, pp. 207–211 (2021) 22. Faizi, R., Chiheb, R., El Afia, A.: Students’ perceptions towards using web 2.0 technologies in education. Int. J. Emerg. Technol. Learn. (iJET) 10(6), 32 (2015). https://doi.org/10.3991/ ijet.v10i6.4858 23. Moss, R.: Video, the Educational Challenge. Croom Helm Ltd., London and Canberra (1983) 24. Goodyear, P., Steeples, C.: Creating shareable representations of practice. Adv. Learn. Technol. J (ALT-J). 6(3), 16–23 (1998) 25. Whatley, J., Ahmad, A.: Using video to record summary lectures to aid students’ revision. Interdisc. J. E-Learn. Learn. Objects 3(1), 185–196 (2007) 26. Wang, H.C., Chen, C.W.Y.: Learning English from YouTubers: English L2 learners’ selfregulated language learning on YouTube. Innov. Lang. Learn. Teach. 14(4), 333–346 (2020) 27. Blakey, N., Guinea, S., Saghafi, F.: Research and Development in Higher Education: Curriculum Transformation Volume 40 (2017) 28. Pecay, R.K.D.: YouTube Integration in Science classes: understanding its roots, ways and selection criteria. Qual. Rep. 22(4), 1015 (2017)

556

R. Faizi

29. Faizi, R., El Fkihi, S.: Could social media replace formal education? In: Proceedings of the 28th International Business Information Management Association Conference - Vision 2020: Innovation Management, Development Sustainability, and Competitive Economic Growth 2016, pp. 3380–3384. International Business Information Management Association (2016) 30. El Fkihi, S., Ezzahid, S., Faizi, R., Chiheb, R.: Formative assessment in the era of big data analytics, In: Proceedings of the 32nd International Business Information Management Association (IBIMA), 15–16 November 2018, Seville, Spain (2018) 31. Irons, A., Elkington, S.: Enhancing Learning Through Formative Assessment and Feedback. Routledge, London (2021). https://doi.org/10.4324/9781138610514

ChatBots and Business Strategy Teresa Guarda1,2,3(B)

and Maria Fernanda Augusto1,2,3

1 Universidad Estatal Península de Santa Elena, La Libertad, Ecuador 2 CIST – Centro de Investigación en Sistemas y Telecomunicaciones, Universidad Estatal

Península de Santa Elena, La Libertad, Ecuador 3 BiTrum Research Group, Leon, Spain

Abstract. New technologies emerge every day, and existing ones are improving. Due to this improvement, chatbots with artificial intelligence appeared. A chatbot is a robot (bot), built through rules defined by a conversational flow and/or Artificial Intelligence, which meets a certain demand of the person who is chatting with it and delivers pre-defined responses. Chatting with a chatbot allows you to solve most customer requests without the intervention of a human, and applications can rely on a powerful ally, artificial intelligence. There are chatbots with and without artificial intelligence; the adoption of one or the other depends on the business of each company. The greater the degree of difficulty in identifying its presence, the greater its efficiency. With artificial intelligence, chatbots can develop in a surprising way, going far beyond a customer service solution, and even becoming indispensable in business strategy. In this article, we will carry out a qualitative bibliometric analysis of works on this topic, with the objective of evaluating the impact of bots on business strategy. Keywords: Chatbot · Artificial intelligence · Neural networks · Natural language processing · Machine learning · Business strategy

1 Introduction A few decades ago, digital transformation took its place in our lives. We depend on the computer to work, the smartphone to communicate and relate, the apps to solve everyday issues, and the internet for basically everything. People spend a good part of their day connected, and so the technological resources are developed with a focus on optimizing processes and improving the user experience. In this context, Bots appeared, allowing companies to manage more efficiently, with greater productivity, and as a consequence, more profit. A chatbot is a software that talks to a person as naturally as you can imagine, serving their needs quickly and assertively [1]. The bot appeared with the aim of allowing the establishment of an immediate communication, which would also allow the reduction of operational costs, and also the expansion of the organizations’ responsiveness to their users. Among the reasons for the growth of chatbots, within different types of companies, is their ability to provide immediate and quality support to customers, while reducing © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 557–566, 2022. https://doi.org/10.1007/978-3-030-96299-9_53

558

T. Guarda and M. F. Augusto

the need to maintain a large team for this service. This allows employees to focus on tasks that require planning and strategy. Overall, the goal of chabots is then to allow the service work to be scalable and truly contribute to the customer experience [2]. Chatbot is a robot (bot) that interacts via chat by messaging applications [3] such as WhatsApp, and Facebook Messenger, Skype and Telegram. The chatbot is the evolution of a system that answers questions through natural language processing (NLP), in addition to using conversation automation to maintain a relationship with customers and leads [4]. For the survey carried out in January 2021, the Scopus platform (www.scopus.com) was chosen, the international database responsible for scientific publications of an interdisciplinary nature. The term used for the search was “chatbot”, having presented 2044 documents, distributed between 2011 and 2020. Analyzing the results presented, the originality of the theme clear (see Fig. 1). In the year, 2016 there was a slight increase in the number of publications, with the year 2017 being the number in which the highest indices appeared at first. From 2018 onwards, the numbers became more expressive.

Fig. 1. Chronology of research related to the Chatbot topic between the years 2011–2020.

Analyzing publications by topic, in 2011 it was distributed by Computer Science (45.5%), Mathematics (18.2%), Social Sciences (18.2%), Engineering (13.6%), and Medicine (4.5%), being the topic with highest indices Computer Science (Fig. 2). Between 2011 and 2020, the publications on a greater variety of topics were increasing. Analyzing publications by topic, in 2020 it were distributed by Computer Science (37.5%), Engineering (15.0%), Mathematics (9.0%), Social Sciences (7.0%), Decision Sciences (6.5%), Medicine (6.4%), Business, Management and Accounting (3.5%), Physics and Astronomy (2.9%), Energy (1.6%), others (7.0%). The area with the highest number of publications is Computer Science. From the year 2018, there was a progressive increase in the number of publications in the following areas: Computer Science; Engineering; Mathematics; Social Sciences; Decision Sciences; Medicine; Business, Management and Accounting; Physics and Astronomy; and Energy (Fig. 3).

ChatBots and Business Strategy

559

Fig. 2. Chronology of research with the most recurrent topics since 2011.

Fig. 3. Chronology of research with the most recurrent topics since 2018.

2 ChatBots The term Chatterbot was created by Michael Mauldin in 1994 [5]. The word Chatbot is an abbreviation for Chatterbot, which in turn can be divided into two other words: the first of these is the word chatter, which designates a conversation or the act of conversing; and the word bot, an abbreviation of the word robot. Joseph Weinzenbaum developed the first chatbot application, ELIZA in the year 1966 at MIT. Matches between keywords were used to simulate a consultation with a psychologist through responses provided to users. With the advancement of graphical interfaces and studies [6]. In the field of artificial intelligence chatbot architectures were created and optimized from the simplest to complex architectures with advanced levels of machine learning and artificial intelligence.

560

T. Guarda and M. F. Augusto

The word Chatbot is used to designate all software or robot that communicates with the user through a conversation, whether through writing or speaking [7]. Chatbots are automated communication software. They are an online message exchange channel that allows the exchange of instant information between the organization and the customer, which consists in creating a harmony between automation and customer service, creating an integrated human service strategy based on chatbots; and enabling the strengthening of the relationship with the consumer, since it is a customizable multi-channel system, which can greatly contribute to the brand’s reputation gain. 2.1 Types of ChatBots Chatbots can be classified according to their purpose, and according to how they are developed [8]. This differentiation happens because the construction of a chatbot varies according to the business objectives it represents. If the purpose is to receive objective commands and provide simple and direct answers, for example, there is no need to implement AI in the chatbot. As for their purpose, they can be classified into (Fig. 4): optimizers; conversational; proactive; shield; and social. Optimizers, take on more concrete challenges and execute them efficiently. They are the type of solution that reduces the friction that exists in other channels, such as applications. Conversational, as the concept suggests, they are used to converse with users. The Proactive, tend to be more objective as they avoid unnecessary notifications. In other words, go straight to the point. Proactive bots track user activities across devices. In this way, they provide accurate information at the most opportune moments. Shield, work in more routine tasks, such as handling complaints and customer support, And social, perform specific activities, as they are structured based on crowds. According to the way they are developed, they are classified into: rules-based chatbots; and chatbots with artificial intelligence. Rules-based bots are built with closed commands and predefined questions [7]. Thus, navigation is structured from a flowchart. In this system, the targeting occurs based on the responses provided by the user. But if chatbots are difficult to understand, the consumer is referred to human service. Chatbots with Artificial Intelligence, or Smart chatbots are more sophisticated and based on machine learning methodology. They provide more complete information than basic models as they are able to learn from every interaction with the customer. The idea of formatting mechanisms based on artificial intelligence is interesting, but it is far from being a common reality in the market due to the high investment and complexity required. However, there are more accessible options, which also meet the criteria for managing the companies’ internal processes [13]. Rule-based chatbots, are more limited than an AI chatbot, because they can only respond to a set number of requests, (Fig. 5). Its use has to do with the objectives that are intended to be achieved. They are an option where technology and implementation are simpler, faster, with a lower cost, good control of chatbot behavior and responses. Despite not detecting the types, and the interaction with the chatbot and the human

ChatBots and Business Strategy

561

Fig. 4. Types of ChatBots.

being very robotic, which can lead to some frustration for the human, this solution is used efficiently in situations that require less complexity. AI-based chatbots, due to their complexity, require more time to implement, and also a greater investment. Despite being more unpredictable to control the behavior of the chatbot, it has a more natural interaction with the human, and the more it interacts with the human, the more it learns.

Fig. 5. Benefits and limitations of Chatbots according to the way they are developed.

2.2 Use of ChatBots Chatbots are still associated with customer support, although this is not their only functionality of a chatbot. The use of chatbots in companies is transforming the relationship

562

T. Guarda and M. F. Augusto

between customers and organizations on different fronts, such as marketing, sales, and, of course, customer service. The chatbot for marketing aims to increase the engagement of the public that relates to the brand [8]. It is also possible to use the chatbot to convert users into leads [9], and from there, establish a relationship close to that established by email marketing, but much more effective. In marketing, chatbots can become extremely effective tools for capturing and qualifying leads; manage the sales funnel steps; provide quick service, thus improving the company’s image before customers; and collect data about the market. In this context of chatbots to capture leads, the software provides explanations about the characteristics of products or services. The objective is to involve the visitor with a strategic dialogue, that is, to make him/her provide as much information as possible to lead him to the sales team. In addition, it is possible to trigger e-mail and integrate automatic chat with other commercial force tools, which organize the scheme of generating and converting leads more effectively. The integration of automatic chat allows us to qualify the information of registered people. In this way, it is possible to adapt the dialogue, direct specific actions, and optimize the entire chatbot process for customer service. In the case of customer service, the user experience must be treated as a priority in today’s market. Therefore, the service needs to be efficient and of quality. In this scheme, it is possible to format the script for the language of a virtual attendant. With this type of triage, it is possible to reduce the waiting list time, and enhance the team’s work. The most complex subjects are directed to human contact. This being the most common type of application.

3 Chatbots and Artificial Intelligence Artificial Intelligence (AI) is an intelligence simulated by computers, which try to approximate the way we humans think, trying to imitate the ability to solve problems and reason, among others. AI is a very old concept, started in 1940. At that time, new features for the computer were already thought, but they were still just projects. With the advent of World War II, the need arose to develop this technology to boost the arms industry, and from that came its great advances. One of the many lines of study of AI is Neural Networks (NN), which are computer models that try to imitate the way a human brain thinks. NN are capable of machine learning and pattern recognition, facilitating the intelligent chatbot, which understands and approaches informal language. This type of learning, through artificial intelligence in bots, is called Natural Language Processing (NLP), which aims to understand people’s language in a more human way to be able to judge what the person is talking about and give an answer according to the subject. Non-AI chatbots can accomplish their purpose by programming specific rules, created in a structure known as a navigation tree or dialog tree. This means that the bot will only respond to specific commands, for which it expects to receive data in a certain format, and will only give previously established and mapped responses. Note that this does not mean that a chatbot without AI is bad. There is no problem with having such

ChatBots and Business Strategy

563

a bot, as long as it fulfills the purpose for which the company built it and works as intended. Objective customer service tasks via chatbots, such as tracking the status of an order, and online ordering services are some chatbot applications that do not require the use of artificial intelligence [11, 12]. Chatbot with AI, have features that are more complex. Among them, we can refer to Machine Learning (ML) and systems that enable the understanding of human language [8]. Through ML techniques, chatbots are able to learn as interactions with users take place, as if they gained experience with each one [8]. This is done using technologies for a better understanding of the requests they receive, such as the case of NLP, which are not limited to ready-made commands, improving engagement with users. The bot is programmed to learn while interacting with users. That is, while we are talking to a system like this, it learns our language and to look for solutions to our doubts. While a rules-based bot has a limited number of sentences and responses that it can identify, with AI, the more the bot interacts, the more it learns and the more accurately it responds. To start its work the bot receives an initial load of information to interact. Some elements are fundamental to the overall functionality of the chatbot with AI, among them: the text classifiers; suitable algorithms; Artificial neural networks; and natural language understanding (NLU), which is a type of NLP. A chatbot with AI can and should offer a hybrid service, in which the chatbot can refer the user to a human agent, in cases of request for information that he still knows how to respond [8]. The difference is that after this first contact with content, the chatbot will learn about what just happened and will have enough material to respond to the next user without needing human help for that problem. When chatbots are connected to technologies like NLU, they can learn more complex ways to simulate human conversations, such as: maintaining context; manage a dialog; adjust responses based on what appears in the conversation. An AI chatbot can also be trained to actively learn from any interaction with a customer in order to improve performance during the next interaction [8]. In an AI chatbot, NLP aims to help the user get to their goal faster, with phrase processing to figure out their intention, or subject. AI systems can be used for a number of purposes, such as: streamline interactions, clarify doubts, transmit knowledge, respond to demands, provide guidance, and control processes, between others. AI chatbots are a link between humans and search engines. Best of all, they are very easy to use: just ask questions and they will answer. Because they work connected to the internet, they are able to provide a huge amount of answers capable of facilitating people’s daily lives [10, 17].

4 Importance of Chatbots in Business Strategy Chatbots provide companies with a real evolution in the relationship between people, technology and companies. Through these systems, companies have access to advantages

564

T. Guarda and M. F. Augusto

such as: quick feedback; efficient interaction; the user does not need to call or even download any application to contact the company. A chatbot is software capable of having a conversation with a human user in natural language through messaging applications, websites, and other digital platforms. Currently, it is possible to create chatbots integrated with platforms such as: WhatsApp; Facebook Messenger; Telegram; websites in general; call center systems; virtual stores, among others. Gartner predicted that by 2020, more than 85% of customer interactions will be handled without humans [14]. This shows that everything we’ve talked about so far about chatbot is already extremely relevant to any business. AI chatbots are already well advanced. That is, they are a reality in the present, and not just a future expectation. In practice, this means that, with the right database and investment, a company can use this type of technology [15, 16]. When we talk about the use of chatbots, we are talking about an irreversible interaction process, since the consumer profile has changed with the advent of new technologies. The use of chatbots aggregates, facilitates, adds and optimizes, regardless of the degree of artificial intelligence involved. Upon hearing the word robot, the trend may be to imagine a cold and impersonal communication. Thankfully, in practice it’s very different. Chatbots can adopt the style of your best service employees, preserving the company’s identity and creating positive experiences that captivate customers. They are friendly, customizable and engaging. For companies Chatbots are ways to improve this experience and offer quick service and solutions to the needs and problems of consumers and prospects. In addition, a company can benefit in different ways, by adopting the bots in their sectors. Chatbots allows minimizing the time used by attendants to perform repetitive tasks, and thus, they can spend more time creating new forms of relationships, loyalty strategies, among others. In addition to making the team more strategic, with company chatbot you increase the speed of service, improving the customer experience with your brand, which increases their satisfaction with your company. Reinforcing the maxim a satisfied customer is a loyal customer. According Intercom, Chatbots although they are used for a wide range of automated tasks, the highlighted areas are mostly used for sales (41%), followed by customer support (37%), and marketing (17%) [8]. Another positive factor for using the chatbot is the scalability of the service, after all, unlike humans, who can only communicate with one person at a time; bots can simultaneously chat with thousands of people. Along with scalability comes cost reduction. With chatbots it is not necessary to maintain a fixed service team all the time. Your business is able to provide assistance to thousands of people, 24 h a day, 7 days a week, at a lower cost than expected with personnel. From the customer’s perspective, some of the benefits that can be mentioned: the ease of getting answers to their demands immediately and at any time; reduced waiting time for service and support; and efficient resolution of support cases [9].

ChatBots and Business Strategy

565

5 Conclusions The need for personalized service is growing. Chatbots are designed to save billions of businesses by 2022, simplifying customer service without the need for a human at the other end of communication, incurring expenses for the business. Unlike rule-based bots, artificial intelligence-based bots do not just respond to commands, they can capture information and desires through the context of a sentence. Most Chatbots are based on AI, which allows them to learn from a pattern in the data, and this is making them more real than ever, increasing their ability with humans in a more natural, effective and intelligent way. The most complex chatbots can interpret and analyze the user’s levels of emotion and from this analysis make predictions and recommendations, making conversations more and more human. These bots have high natural language processing capacity and unstructured data and once they have these characteristics, and can capture and interpret the user’s intentions, emotions and patterns, aiding in the bot’s decision making and in turn making it smarter. Chatbots are being used in companies to improve customer service; reduce costs; generate more friendly, agile and efficient contacts. Despite the challenges that still exist, the chatbot has everything to make life easier for both customers and companies.

References 1. Khan, R., Das, A.: Build Better Chatbots: A Complete Guide to Getting Started with Chatbots. Apress, New York (2019) 2. Kumar, K.N., Balaramachandran, P.R.: Robotic process automation-a study of the impact on customer experience in retail banking industry. J. Internet Bank. Commer. 23(3), 1–27 (2018) 3. Araujo, T.: Living up to the chatbot hype: the influence of anthropomorphic design cues and communicative agency framing on conversational agent and company perceptions. Comput. Hum. Behav. 85, 183–189 (2018) 4. Nagarhalli, T.P., Vaze, V., Rana, N.K.: A review of current trends in the development of chatbot systems. In: 6th International Conference on Advanced Computing and Communication Systems (ICACCS) (2020) 5. Mauldin, M.L.: Chatterbots, tinymuds, and the turing test: entering the loebner prize competition. AAAI 94, 16–21 (1994) 6. Colby, K.: Human-computer conversation in a cognitive therapy program. In: Machine Conversations, pp. 9–19. Kluwer Academic Publishers, Boston (1999) 7. Hasal, M., Nowaková, J., Saghair, K. A., Abdulla, H., Snášel, V., Ogiela, L.: Chatbots: security, privacy, data protection, and social aspects. Concurr Comput. Prac. Exper. 33(19),(2021). https://doi.org/10.1002/cpe.6426 8. Adamopoulou, E., Moussiades, L.: An overview of chatbot technology. In: Maglogiannis I., Iliadis L., Pimenidis E. (eds.) Artificial Intelligence Applications and Innovations. AIAI 2020. IFIP Advances in Information and Communication Technology, vol 584. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-49186-4_31 9. Cheng, Y., Jiang, H.: Customer–brand relationship in the era of artificial intelligence: understanding the role of chatbot marketing efforts. J. Prod. Brand Manage. 31(2) 1–13 (2021). https://doi.org/10.1108/JPBM-05-2020-2907

566

T. Guarda and M. F. Augusto

10. Illescas-Manzano, M.D., López, N.V., González, N.A., Rodríguez, C.C.: Implementation of chatbot in online commerce, and open innovation. J. Open Innov. Technol. Mark. Complex. 7(2), no. 125, 1–20 (2021) 11. Khan, R., Das, A.: Introduction to Chatbots. In: Build Better Chatbots. Apress, Berkeley, CA (2018) 12. Smutny, P., Schreiberova, P.: Chatbots for learning: a review of educational chatbots for the Facebook Messenger. Comput. Educ. 151, 1–11 (2020) 13. Gao, Z., Jiang, J.: Evaluating human-AI hybrid conversational systems with chatbot message suggestions. In: CIKM 2021, Virtual Event, QLD, Australia, 1–5 Nov 2021 14. Berezina, K., Ciftci, O., Cobanoglu, C.: Robots, artificial intelligence, and service automation in restaurants. In: Robots, artificial intelligence, and service automation in travel, tourism and hospitality, pp. 185–219. Emerald Publishing Limited (2019) 15. Gartner: Gartner Customer 360 Summit 2011 (2018). https://www.gartner.com/imagesrv/sum mits/docs/na/customer-360/C360_2011_brochure_FINAL.pdf 16. Yin, S.: Where chatbots are headed in 2021 (2019). https://www.intercom.com/blog/the-stateof-chatbots/. Accessed 2019 17. Brandtzaeg, P.B., Følstad, A.: Why people use chatbots. In: International Conference on Internet Science (2017)

A Modified Feature Optimization Approach with Convolutional Neural Network for Apple Leaf Disease Detection Vagisha Sharma(B) , Amandeep Verma, and Neelam Goel IT-Department, University Institute of Engineering and Technology - Panjab University, Chandigarh, India {amandeepverma,erneelam}@pu.ac.in

Abstract. Monitoring plant health and detecting various plant diseases is an important discipline of plant pathology. Some diseases can even be detected by observing pathological characteristics. However, it requires high expertise and is very time consuming when inspecting large farms. Therefore, automated disease detection methods are required for rapid analysis. In this paper, an automated method for detecting apple leaf disease is proposed, and image processing techniques are used to analyze leaf samples. The paper uses the approach of Contrast Limited Intensity Adjustment (CLIA) for image enhancement followed by segmentation of the diseased region using k-means. Features of this region are extracted using Speeded Up Robust Features (SURF) that is further optimized using Particle Swarm Optimization (PSO) to select the best feature set among the extracted features. At training and classification stage, CPU based Convolutional Neural Network (CNN) is implemented whose performance is evaluated on a dataset comprising of 3,171 leaf images. Simulation analysis demonstrates that the proposed work outperforms existing works, with an average precision of 0.97 and accuracy of 99.26%. Keywords: Apple leaf disease · k-means · SURF · PSO · Convolutional neural network (CNN)

1 Introduction Agriculture is a major source of revenue in many countries. Many plants are harvested depending on land and environmental conditions [1]. However, farmers face several challenges, such as no proper amenities, soil erosion, natural catastrophes, plant diseases, and so on. Although, some of these issues can be resolved by offering technical support to farmers [2, 3]. An automated and accurate plant disease detection system is very beneficial for farmers as it helps to increase production rate by minimizing the losses in crop production [4]. Traditionally, visual inspection methods were used by experts for detecting plant leaf diseases. But risk factor in these methods was high as inspection had been performed through naked eyes [5]. Several strategies have been integrated into traditional systems to obtain a dynamic automatic plant disease detection system. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 567–578, 2022. https://doi.org/10.1007/978-3-030-96299-9_54

568

V. Sharma et al.

Apple is one of the most productive fruits in the world, with considerable nutritional and therapeutic value. Various diseases, however, frequently damage apple production, resulting in significant economic losses. To prevent the loss of harvest many machine learning approaches for disease detection have been developed. To detect the diseased part in apple plant, K-Means based image segmentation is performed and features are extracted. A Multi-class SVM classifier is used for training and classification which gives an accuracy of 93% [6]. To identify disease in apple plant, Fuzzy C-means Algorithm, and Nonlinear Programming Genetic Algorithm in combination with a multivariate image analysis approach has been used on a dataset of 2,000 images which gave an overall detection accuracy of 98% [7]. An improved automated computer-based method which uses lesion spot contrast stretching, lesion segmentation, and features selection & recognition for detection of apple diseases. The best features are selected using Genetic algorithm, which are used by M-SVM for classification [8]. The key contributions of this work can be summarized as follows: A novel deep convolutional neural network model in combination with feature extraction and optimization is proposed to accurately identify the four common categories of apple leaf diseases at an early stage and thereby improving the rate of apple production. The proposed approach improves detection accuracy and is completely automated and outperforms other common apple disease diagnosis models. The paper has been organized as follows: Sect. 2 of paper describes background of existing work in apple leaf disease detection. In Sect. 3, methodology of proposed work is presented. The detailed description of results is discussed in Sect. 4. Finally, in last section, the conclusion of work is presented, and future work has been discussed which is followed by references.

2 Related Work The primary cause for loss in agricultural productivity are plant diseases. Several researchers have worked to detect plant leaf disease at an early stage to minimize the losses for farmers. This section discusses various techniques used by researchers for detecting diseases in apples. Omrani et al. [9] used three classes of apple leaf diseases from agricultural research institution at University of Tehran. Training and testing were performed using ANN and SVM. Zhang et al. [10] used I-RELIEF algorithm for feature extraction and weighted relevance vector machine (RVM) classifier for detecting apple diseases. They had an accuracy of 95.63% for 160 samples. Nachtigall et al. [11] studied apple diseases caused by nutritional imbalances, such as potassium and magnesium deficiencies, as well as diseases such as apple scab, Glomerella stains, and herbicide damage. They trained their data using CNN and achieved 97.3% accuracy using a data set of 2,539 images of apple leaves. Chuanlei et al. [12] developed a model using Genetic Algorithm, and correlationbased feature selection approach. GA has been used to minimize the dimension of feature space by selecting desired features of image based on CFS function. The desired features are then used as training data to train SVM classifier with 90% disease detection accuracy. Wang et al. [13] presented a deep learning approach where performance of deep models fine-tuned by transfer learning and shallow networks are evaluated for

A Modified Feature Optimization Approach

569

apple leaf features. The network is trained using CNN using four different models. The VGG16 model trained with transfer learning performs best, with an accuracy of 90.4%. Liu et al. [14] presented a model for identifying four types of apple leaf diseases based on deep CNN. They used a dataset of 13,689 images of diseased apple leaves and achieved an accuracy of 97.62%. Baranwal et al. [15] proposed a deep learning-based CNN approach to detect diseases in apple trees. The model is trained on a dataset which has apple black rot, apple cedar, apple rust and healthy apple leaf images. The model achieves an accuracy of 98.42%. Jiang et al. [16] have used deep CNN to detect apple leaf disease. By introducing GoogLeNet Inception structure with Rainbow concatenation, a new apple leaf disease detection model based on deep-CNN is proposed. Around 26, 377 leaf images of four different categories have been considered. The designed model has a detection performance of 78.80%. Li et al. [17] have used data sets consisting of apple grey-spot, black star, and cedar rust diseases. For classification SVM classifier, as well as the ResNet and VGG convolutional neural network models, were used. Their model achieved 98% accuracy rate. Al-bayati et al. [18] proposed a novel experimental analysis for detection and classification of disease in apple leaves by using deep neural network. The dataset has been taken from Plant Village Dataset to perform simulation. Features have been extracted using Grasshopper Optimization Algorithm. An accuracy of 98.28% is achieved using this model. Alam et al. [19] proposed a deep learningbased automatic classification for apple diseases. A dataset produced from the set of photos acquired from the agricultural-based industrial fruit sorting process is used to train the classifier. The system had an accuracy of 96% for detecting apple diseases. Zhong et al. [20] have developed a method to identify apple leaf diseases, based on DenseNet-121 deep convolution network. For predictive analytics and evaluation, a data set of 2,462 apple leaf images with six apple leaf diseases, was used. The proposed methods achieved an average accuracy of 93.5%. Chao et al. [21] proposed a model for detecting diseases in apple based on deep convolutional neural network in combination with DenseNet and Xception. They used a dataset of 2,970 images of diseased and healthy apple leaves shot in the lab and in real cultivation fields in China. Their work achieved an accuracy of 98.82%. Yu et al. [22] have presented a technique using a deep convolutional neural network with a region-of-interest-aware architecture. Apple leaf images were used from Apple Research Institute in South Korea and categorized into three groups, i.e., two diseased and one healthy class. Two subnetworks were designed that were trained separately through transfer learning with a new training set. MATLAB was used for implementation. The accuracy of this model was 84.3%. Song et al. [23] proposed a model using GoogLeNet’s inception for detecting diseases on small scale datasets. The data set is augmented using various techniques and then deep convolution generative adversarial network is used. The model had a accuracy of 98.5%. Rehman et al. [24] proposed a model for real-time apple leaf disease identification using RCNN classifier. Features were selected using the Kapur’s entropy along MSVM approach achieving an overall accuracy of 96.6%.

3 Methodology In this section, proposed work is discussed along with techniques used. In pre-processing stage image enhancement is performed to distinguish the background and foreground

570

V. Sharma et al.

images, and K-means is used for image segmentation. In this model we have, used SURF as a feature extraction approach and integrated it with PSO technique to optimize the extracted features. This allowed the training phase to be completed with minimal error and thus achieving high classification accuracy during system validation process when CNN is applied as classifier. 3.1 Dataset The dataset used is a subset of Plant Village Dataset by Mohanty et al. 2016 [25]. In this research, we have considered the apple leaf dataset consisting of 630 images of apple scab, 621 images of black rot, 275 images of cedar apple rust and 1,645 healthy apple leaves. The images are in JPG format with 8–20 KB size. 3.2 Pre-processing The initial step is to enhance the quality of image, which is performed by separating background from the foreground of image. K-means technique is then used to segment the diseased regions in the leaf image. The changes in the leaf image as a result of pre-processing are represented in Fig. 1.

Original Image

Enhanced Image

Segmented Image

Fig. 1. Pre-processed image

Image Enhancement is the process of improving quality of an image, which is performed by enhancing intensity of image in terms of red, green, and blue colour. Highquality images are useful in accurately extracting the diseased region in an image. To improve visual quality of an image, we have used Contrast Limited Intensity Adjustment (CLIA) technique that uses restricted contrast to map an image’s intensity values to a new range and generate higher-quality image. A window size is set for the intensity values that do not require improvement. After image enhancement, image segmentation has been performed using k-means. It divides image pixel into two clusters namely, foreground and background. Based on nearest distance criteria, the regions in the enhanced leaf image are marked as regions of interest that represent the diseased areas. The segmented region of the leaf is returned as foreground cluster and extra part in the leaf is returned as a background cluster. 3.3 Feature Extraction The features of enhanced and segmented leaf images are point of interest. The features representing the diseased region are extracted from enhanced leaf image using SURF.

A Modified Feature Optimization Approach

571

It works on the diseased region of leaf image that has been segmented and labelled as region-of-Interest. The best features are retrieved and returned as feature points for the segmented portion of leaf image. 3.4 Feature Optimization Many features are obtained in the previous stage however all features are not feasible to be used for classification purpose. Therefore, Particle Swarm Optimization (PSO) is used to optimize the features extracted in the previous stage using SURF. It is a stochastic optimization technique based on the social behaviour of swarms of birds or schools of fish. In this technique the particles, or a swarm of individuals, flow through swarm space [26]. It demonstrates the ability to analyze the search space while tracking the coordinates with the help of fitness function. It is a meta-heuristic approach that optimizes the results based on position and velocity parameters. The optimization of extracted features is done using following algorithm.

The above algorithm represents a form of decision-making step influenced by particle behaviour and architecture to select the best features from a vast number of features that is significant for distinguishing between different disease types. It returns the optimized feature data OTdata with reduced data size to enable fast processing in the next step. 3.5 Training and Classification Convolutional Neural Network (CNN) architecture consists of several convolutional layers and pooling layers. In the present work, CNN is used for training and classification

572

V. Sharma et al.

of the optimized feature data. The training data is distributed into four groups corresponding to three disease classes and one healthy class. New leaf image is passed to the network in the input layer for classification purpose and the probability is calculated while passing the prediction results to the output layer. The steps involved in training and testing of leaf images for disease prediction are given below.

In the above algorithm, the trained CNN database is further used to test the target data for disease recognition. If the test leaf image belongs to any of the disease categories it is

A Modified Feature Optimization Approach

573

classified into that particular class and the performance parameters in term of precision, recall, f-measure, accuracy and error rate are computed.

4 Results and Discussion In the proposed CNN model, the classifier’s input consists of optimized features that are obtained from a fitness function optimization technique, and the final output layer is responsible for classifying apple leaf disease. Fully connected layers are used, with each neuron providing a full connection to all optimized features given by the preceding layer. The ReLu nonlinear activation function is used in each convolutional layer. Max pooling is used after the convolutional layer for enhancing efficiency of the activation function and improving convergence speed. Cross-entropy loss function is also used. The parameters of the proposed model used in the study are given in Table 1. Table 1. The parameters of the proposed model Convolution layers

3 (with [3 × 3] filters each)

Max-pooling layers

3 (with [2, 2] pool size each)

Convolution layer activation function

ReLu

Epochs

1000

Loss function

Cross entropy

The metrics that are used to evaluate performance of our model are Precision, Recall, and F-Measure. Precision is the ratio of the true positives to the sum of true positives and false positives. Recall is the ratio of the true positives to the sum of true positives and false negatives. F-Measure is the harmonic mean of Precision and Recall whose value lies in the range of [0, 1]. A greater F-Measure score indicates a better balance of Precision and Recall which suggests a more reliable model [27]. These are given in mathematical equations as: Precision = Recall =

True Positives True Positives + False Positives

True Positives True Positives + False Negatives

F−Measure =

2 1 precision

+

1 recall

(1) (2) (3)

The performance analysis in terms of these metrics is shown in Table 2. Performance of the proposed CPU based CNN used for classification of leaf images into four classes is analyzed in terms of precision, recall and f-measure in Fig. 2. The number of leaf images used in the study is plotted on X-axis against the parametric values of three performance parameters. It is observed that the precision of the proposed

574

V. Sharma et al. Table 2. Performance analysis in terms of Precision, Recall and F-measure Number of leaf images

Precision

Recall

F-measure

10

0.965

0.945

0.954

20

0.968

0.948

0.957

50

0.971

0.954

0.962

100

0.976

0.964

0.969

200

0.979

0.969

0.973

500

0.982

0.971

0.976

1000

0.986

0.975

0.980

2000

0.987

0.979

0.982

3000

0.990

0.981

0.985

3171

0.991

0.982

0.986

Parametric Values

work lies between 0.96 and 0.99 which increases from 0.965 with increase in the sample size used for evaluation, demonstrating an average value of 0.979. Similarly, recall value of 0.945 is observed corresponding to smaller sample size consisting of 10 leaf images which increases to 0.982 when 3,171 leaf images are included for experimentation achieving an average value of 0.966. It is observed that f-measure also increased from 0.954 to 0.982 with an average value of 0.973. This shows that CPU based CNN demonstrated a better precision for disease classification.

1 0.99 0.98 0.97 0.96 0.95 0.94 0.93 0.92 10

20

Precision

50

100 200 500 1000 2000 3000 3171 Number of Leaf Images Recall F-measure

Fig. 2. Performance analysis in terms of Precision, Recall and F-measure

The parametric values of accuracy and error achieved during simulation are summarized in Table 3. It is observed that as the numbers of leaf images are increased for classification, accuracy also increases with corresponding decrease in the error.

A Modified Feature Optimization Approach

575

Table 3. Performance analysis in terms of Accuracy and Error Number of leaf images

Accuracy

Error

10

98.55

1.45

20

98.75

1.25

50

98.98

1.02

100

99.13

0.87

200

99.25

0.75

500

99.38

0.62

1000

99.46

0.54

2000

99.52

0.48

3000

99.72

0.28

3171

99.9

0.1

Figure 3, shows that initially when small image dataset comprising of 10 leaf images is used the proposed work achieved accuracy of 98.55% with error of 1.45%. Further as the number of leaf images are increased to 100, accuracy increases to 99.13% with an error of 0.87% and when a larger dataset comprising of 3,171 images is used accuracy of classification increases to 99.9% with 0.1% error. It shows that proposed work demonstrated an average accuracy of 99.26% and average error rate of 0.736%.

Values in percentile

100 99 98 97 96 95 10

20

50 Error

100 200 500 1000 2000 3000 3171 Number of Leaf Images Accuracy

Fig. 3. Performance analysis in terms of Accuracy and Error

The performance of the proposed work is also evaluated against the existing leafbased disease classification models based on CNN architecture. Table 4 shows the classification accuracy of existing and proposed work with the techniques implemented by researchers for classifying apple leaf disease and it is plotted in Fig. 4, for graphical comparison. Deep learning architecture of CNN was implemented by Wang et al., to classify leaf diseases into five diseased classes with an accuracy of 90.4%. Liu et al. used Deep CNN

576

V. Sharma et al. Table 4. Accuracy comparison with existing work Research work

Techniques used

Accuracy (%)

Proposed

CPU based CNN

99.2

Wang et al. [13]

deep-learning CNN

90.4

Liu et al. [14]

Deep CNN

97.62

Jiang et al. [16]

CNN

78.8

Al-bayati et el. [18]

DNN

98.28

Rehman et al. [24]

MASK RCNN

96.6

Accuracy %

based on AlexNet to obtain an accuracy of 97.62%. Then Jiang et al., had used CNN for training and classification of leaf images with an accuracy of 78.8%. Al-bayati et al. used DNN for disease detection and had accuracy of 98.28%. Rehman et al. made use of MASK RCNN to achieve an accuracy of 96.6%. In contrast to these works, the proposed work uses CPU based CNN to achieve an average accuracy of 99.2% which reflects that the proposed work outperformed the existing works.

120 100 80 60 40 20 0 Proposed Wang et el. Liu et al. (2021) (2017) (2018)

Jiang et el. Al-bayaƟ et Rehman et (2019) el. (2020) al. (2021)

Fig. 4. Accuracy comparison with existing work

5 Conclusion and Future Work The paper introduces a Contrast Limited Intensity Adjustment technique for image enhancement and then, k-means is used for distinguishing foreground and background pixels of leaf image. Then using SURF features are extracted and PSO is applied to identify best feature set to be fed to CNN for classification of leaf samples into three disease classes and one healthy class. Simulation analysis demonstrates an average precision

A Modified Feature Optimization Approach

577

of 0.97, recall of 0.96, f-measure of 0.97 and accuracy of 99.26% with a small error of 0.736%. The outperformance of proposed work against existing models demonstrates effectiveness of proposed work in detecting apple leaf disease. However, there are some limitations of this work such as only a few diseases of apple plant have been discussed. Another issue is that the proposed approach cannot detect many diseases in a single image. Also, there is a problem of limited number of images in the available dataset. An expansion of the database is required, and more training samples are needed for improving disease prediction accuracy. In future, other deep neural network model architectures, can be used for identifying apple leaf diseases in real time. The model described here is a generic model, and more effort will be required to turn this into a real-world application that farmers can use to reduce apple production loss. Also, high-quality natural images of apple leaf diseases can be obtained from the fields to identify more diseases in timely and accurate manner. The authors also look forward to extending the presented segmentation and classification architecture to detect leaf disease in other crops.

References 1. Wang, Z., Li, H., Zhu, Y., TianFang, X.: Review of plant identification based on image processing. Arch. Comput. Meth. Eng. 24(3), 637–654 (2017) 2. Singh, V., Misra, A.K.: Detection of plant leaf diseases using image segmentation and soft computing techniques. Inf. Process. Agric. 4(1), 41–49 (2017) 3. Sandeep, K., Sharma, S., Sharma, V.K., Sharma, H., Bansal, J.C.: Plant leaf disease identification using exponential spider monkey optimization. Sustain. Comput.: Inf. Syst. 28, 100283 (2018) 4. Keunho, P., Hong, Y.K., Kim, G.H., Lee, J.: Classification of apple leaf conditions in hyperspectral images for diagnosis of Marssonina blotch using mRMR and deep neural network. Comput. Electron. Agric. 148, 179–187 (2018) 5. Barbedo, J.G.A.: Factors influencing the use of deep learning for plant disease recognition. Biosyst. Eng. 172, 84–91 (2018) 6. Shiv Ram, D., Jalal, A.S.: Detection and classification of apple fruit diseases using complete local binary patterns. In: 2012 Third International Conference on Computer and Communication Technology, pp. 346–351. IEEE (2012) 7. Zhang, W., Juan, H., Zhou, G., He, M.: Detection of apple defects based on the FCM-NPGA and a multivariate image analysis. IEEE Access 8, 38833–38845 (2020) 8. Khan, M.A., et al.: An optimized method for segmentation and classification of apple diseases based on strong correlation and genetic algorithm based feature selection. IEEE Access 7, 46261–46277 (2019). https://doi.org/10.1109/ACCESS.2019.2908040 9. Omrani, E., Khoshnevisan, B., Shamshirband, S., Saboohi, H., Anuar, N., Nasir, M.H.N.M.: Potential of radial basis function-based support vector regression for apple disease detection. Measurement 55, 512–519 (2014). https://doi.org/10.1016/j.measurement.2014.05.033 10. Zhang, B., et al.: Computer vision detection of defective apples using automatic lightness correction and weighted RVM classifier. J. Food Eng. 146, 143–151 (2015) 11. Lucas, G.N., Araujo, R.M., Nachtigall, G.R.: Classification of apple tree disorders using convolutional neural networks. In: 2016 IEEE 28th International Conference on Tools with Artificial Intelligence (ICTAI), pp. 472–476. IEEE (2016) 12. Chuanlei, Z., Shanwen, Z., Jucheng, Y., Yancui, S., Jia, C.: Apple leaf disease identification using genetic algorithm and correlation-based feature selection method. Int. J. Agric. Biol. Eng. 10(2), 74–83 (2017)

578

V. Sharma et al.

13. Wang, G., Sun, Y., Wang, J.: Automatic image-based plant disease severity estimation using deep learning. Comput. Intell. Neurosc. 2017, 1–8 (2017). https://doi.org/10.1155/2017/291 7536 14. Liu, B., Zhang, Y., He, D., Li, Y.: Identification of apple leaf diseases based on deep convolutional neural networks. Symmetry 10(1), 11 (2018) 15. Saraansh, B., Khandelwal, S., Arora, A.: Deep learning convolutional neural network for apple leaves disease detection. In: Proceedings of International Conference on Sustainable Computing in Science, Technology and Management (SUSCOM), Amity University Rajasthan, Jaipur-India (2019) 16. Jiang, P., Chen, Y., Liu, B., He, D., Liang, C.: Real-time detection of apple leaf diseases using deep learning approach based on improved convolutional neural networks. IEEE Access 7, 59069–59080 (2019) 17. Xin, L., Rai, L.: Apple leaf disease identification and classification using ResNet models. In: 2020 IEEE 3rd International Conference on Electronic Information and Communication Technology (ICEICT), pp. 738–742. IEEE (2020) 18. Al-bayati, J.S.H. Burak, B.Ü.: Artificial intelligence in smart agriculture: modified evolutionary optimization approach for plant disease identification. In: 2020 4th International Symposium on Multidisciplinary Studies and Innovative Technologies (ISMSIT), pp. 1–6. IEEE (2020) 19. Nur Alam, M.D., Saugat, S., Santosh, D., Sarkar, M.I., Al-Absi, A.A.: Apple defect detection based on deep convolutional neural network. In: Pattnaik, P.K., Sain, M., Al-Absi, A.A., Kumar, P. (eds.) SMARTCYBER 2020. LNNS, vol. 149, pp. 215–223. Springer, Singapore (2021). https://doi.org/10.1007/978-981-15-7990-5_21 20. Yong, Z., Zhao, M.: Research on deep learning in apple leaf disease recognition. Comput. Electron. Agric. 168 105146 (2020) 21. Chao, X., Sun, G., Zhao, H., Li, M., He, D.: Identification of apple tree leaf diseases based on deep learning models. Symmetry 12(7), 1065 (2020) 22. Hee-Jin, Y., Son, C.-H., Lee, D.H.: Apple leaf disease identification through region-ofinterest-aware deep convolutional neural network. J. Imaging Sci. Technol. 64(2), 20507– 20507 (2020). https://doi.org/10.2352/J.ImagingSci.Technol.2020.64.2.020507 23. Song, C., Wang, D., Bai, H., Sun, W.: Apple disease recognition based on small-scale data sets. Appl. Eng. Agric. 37(3), 481–490 (2021) 24. Zia ur, Z., Khan, M.A., Ahmed, F., Damaševiˇcius, R., Naqvi, S.R., Nisar, W., Javed, K.: Recognizing apple leaf diseases using a novel parallel real-time processing framework based on MASK RCNN and transfer learning: an application for smart agriculture. In: IET Image Processing (2021) 25. Mohanty, S.P., Hughes, D.P., Salathé, M.: Using deep learning for image-based plant disease detection. Front. Plant Sci. 7, 1419 (2016) 26. Verma, A., Kaushal, S.: A hybrid multi-objective particle swarm optimization for scientific workflow scheduling. Parallel Comput. 62, 1–19 (2017) 27. Bansal, P., Kumar, R., Kumar, S.: Disease detection in apple leaves using deep convolutional neural network. Agriculture 11(7), 617 (2021)

Ontology Based Knowledge Visualization for Domestic Violence Cases Tanaya Das1(B) , Abhishek Roy1 , and Arun Kumar Majumdar2 1 Adamas University, Kolkata, India 2 JIS University, Kolkata, India

Abstract. Ontology of any domain formally represents the knowledge based on concepts along with its relationship and properties. Legal ontology has a significant contribution in capturing knowledge of legal domain. The legal professionals analyse legal documents to extract knowledge based on the parameters mentioned in legal sections to prepare their draft. This research work mainly contributes in capturing the knowledge of legal section via ontology. The proposed ontology has mapped the entities and their relationship of textual rule related to Criminal major Act on Domestic Violence as stated in Indian Penal Code Section 498A. It mainly focuses on relevant portions on which legal professionals give major emphasis in Domestic Violence cases. Authors have extracted major entities present in Domestic Violence cases using Natural Language Processing techniques like named entity recognition, regex, etc. and generated a parse tree. The entities and relationship generated by the parse tree is further represented as classes, object properties, instances, etc., using Protege. Authors have tried to represent the concepts and their relationship through ontology visualizer i.e. OntoGraf. Keywords: Artificial Intelligence · Natural language processing · Knowledge base · Indian Penal Code Section · Domestic violence

1 Introduction The concept of domain knowledge is essential for the development of any Artificial Intelligence based system. It represents basic understanding about the area or subject under its consideration. This capturing of knowledge leads to the identification of ontology. There are two major theories of artificial intelligence i.e., mechanism and content theories [1]. Content theories describe the types of objects, the attributes of objects, and the relationships between objects that are feasible in a specific domain of knowledge. The primary goal of ontology is to aid in the development of knowledge base applications [2]. It mainly assists in the formal modelling, reasoning, and querying the knowledge base. Ontology helps in structuring and relatively represent the knowledge to develop the knowledge base that assist professionals in faster delivery of services. Many fields like education [3], healthcare [4], business [5], legal [6] etc. are capturing the specific domain knowledge via ontology. Legal ontology-based system has the potential to capture the legal knowledge. Legal professionals till today are comfortable and rely on methods and solutions that were © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 579–589, 2022. https://doi.org/10.1007/978-3-030-96299-9_55

580

T. Das et al.

developed years ago. Over the past few years still lawyers are frequently using search tools that help to identify relevant laws, statutes, rules, judgments associated with different court cases which is time consuming. Legal cases are related to facts based on certain rules or charges i.e., Indian Penal Code Sections against the accused persons. In this paper the authors mainly focused on capturing the major entities and their relationships present in legal rule or section with the help of ontology. Instead of focusing on all Indian Penal Code Sections, the authors have emphasized on sections relating to crime against women. Indian Government study [7] shows that reported crime against women is rising on a daily basis. The rate of violence against women is high in India. A research report [8] has recorded significant growth in domestic violence cases that during the lockdown phases of COVID-19 outbreak. For better analysis of these cases, ontology incorporates textual rules related to criminal major acts on domestic violence. Hence, the authors have focused to present the concepts and their relationships present in Indian Penal Code Section 498A i.e., Domestic Violence [9]. The proposed ontology concentrated on pertinent aspects of Domestic Violence that legal professionals look while reviewing case materials. Section 2 describes the related literatures over legal ontologies. Section 3 and its subsection discusses proposed methodology with the help of natural language processing techniques to find the entities and relations in form of parse tree. Section 4 and its sub sections describe the representation of entities, relationship and finally visualize the ontological structure of Indian Penal Code Section 498A using Protege. Section 5 describes the performance evaluation of the proposed work with the help of some legal documents. Section 6 finally concludes the paper and the future work of this study.

2 Literature Review Domain-specific ontology building is a complex task. It includes the identification of many terms related to that domain along with their interrelationships between the concepts. Among several domain-specific ontologies, legal ontologies have rich literature. In 2005, Artificial Intelligence and Law literature focused on fundamental types of legal ontologies and their roles [10]. The role of legal ontologies was reasoning, semantic indexing, integration, and interoperation to understand the domain. Any domain conceptualization includes individuals, classes, properties, relations, rules, and axioms. Domain conceptualizations in legal information system was discussed, with the help of an example from United Kingdom security law [11]. The core ontology played a role in translating existing legal knowledge bases to other alternative representation styles. Legal core ontologies help to conceptualize legal domain knowledge. The LKIF Core Ontology was a research work based on combined principle and common-sense ideas that articulate legal knowledge [12]. Another literature explained a case-based reasoning using case ontology with the help of instances from a particular case [13]. The literature focused on the ontology of general legal concept like hearing, decision, jurisdiction, etc. but required clarity on specific concepts to make it more applicable to judgments. From the existing literatures, the authors have found that interpreting judicial decision includes ontology on overall judgment structure but does not provide any conceptualization that interpret the legal act or legal rule. Interpreting legal rule is an important task that helps

Ontology Based Knowledge Visualization

581

in analyzing judicial decisions. Legal professionals analyze the segments in any legal case first identify the concepts present in legal rule or section. To prepare any knowledge base on legal section or rule the first step is to identify the entities along with their relationships present in that particular legal rule. At present, no ontology-based work exists on Indian Judicial System, particularly on domestic violence against women. Hence, in this paper the authors have proposed knowledge-based ontology on Indian Penal Code Section 498A.

3 Proposed Work The proposed methodology includes natural language processing techniques like tokenization, pos tagging, regex,etc. Further, the entities, their relationship and instances are represented using protege tool to provide a comprehensive solution to legal professional in analyzing domestic violence cases. The stepwise proposed methodology is shown in Fig. 1.

Fig. 1. Stepwise proposed methodology

Step I describes the text form of domestic violence section written in law journals, books, statutes, etc. Step II involves preprocessing techniques like tokenization and parts-of-speech tagging to represent the parts of speech tagged with the words present in legal rule. Step III discusses the way to generate the parse tree of entities and relationships with the help of grammar rule, regex parser and domain expert. Step IV discusses the way to build ontology using Protege tool. Step V describes the evaluation of the ontology with the help of documents on domestic violence annotated by legal professionals. The detailed description of the steps are described in subsequent sections. 3.1 Text Analytics Techniques in NLP Natural Language Processing Techniques are comprised of some text analytics methodologies that are used to extract insights from the text [14]. In this paper, authors have applied techniques like parts of speech tagging and regex to generate parse tree of Indian Penal Code Section 498A. Figure 2 and Fig. 3 show the code snippet and represent the entity and relation present in legal rule in the form of parse tree. Further the entity and the relationship are described with the help of Protege tool.

4 Ontology of Indian Penal Code Section 498A Using Protege The formal defination of ontology along with ontological representation of Indian Penal Code Section 498A are discussed further.

582

T. Das et al.

Fig. 2. Code Snippet of rule to generate parse tree of IPC 498A

Fig. 3. Snapshot of Parse tree representation of IPC 498A

An ontology is a formalized unified conceptualization of a domain. An ontology should be formal i.e., machine readable. A formal ontology is defined as the tuple [15]. O = (C, H, I, R, P, A). According to the above formal definition of ontology, tuples like C, H, and I i.e., concepts, sub-concepts or subclasses and instances are identified and explained in Sect. 4.1. The remaining tuples R, P and A i.e., set of non-taxonomic relations, data properties of entities and rules or axioms are explained in SubSects. 4.2 and 4.3. The open-source framework Protege 5.5.0 is applied to illustrate the detailed framework of inter-relationships that exists between the major concepts. 4.1 Representation of Concepts as Entities Along with individuals As shown in Fig. 3, major words are represented as entities and relation are presented in form of parse tree. Domestic Violence consists of major entities or concepts like husband, wife, unlawful demand, wilful conduct, harassment, punishment. Further, the some classes or concepts are divided into subclasses or subconcepts for more clarity. For example, Unlawful_Demand is categorised as Property or Valuable Security such as gold, cash, vehicles, furniture, land, residential flat and many more. Punishment class is subdivided in sub clase Fine_Amount and Term_Years. Likewise, Individuals constitute the fundamental unit of ontology. Individuals define entities and their property values as a formal aspect of semantics (Table 2).

Ontology Based Knowledge Visualization

583

Table 1. Formal defination of tuples and its description Formal defination

Description

C = C c U C i [15]

Set of entities mentioned in the ontology. C c denotes the classes i.e. the concepts that represent entities. C i represents the class instances

H = {kind_of(c1 ,c2 )|c1 C c ,c2 C c } [15]

Set of taxonomic relations between concepts, which define a concept hierarchy and are denoted by kind_of(c1 ,c2 ) meaning that c1 is a subclass of c2

I = {is_a (c1 ,c2 ) | c1 ∈ C I , c2 ∈ C c } ∪ {propI (ck , value) | ck ∈ C I } ∪ { rel k (c1 , c2 … cn )|∀i, ci ∈ C I } [15]

Set of instances related to entities or concepts

R = {rel k (c1 , c2 … cn ) | ∀i, ci ∈ C c } [15]

Set of non-taxonomic ontology relationships between the concepts or entities

P = {propC (ck , datatype )| ck ∈ C c } [15] Represents set of data properties of entities A = {conditionx ⇒ conclusiony (c1, c2… cn) Axioms and rules that helps to check |∀ j, cj ∈ C c } [15] consistency of the ontology and infer new knowledge using inference techniques

Table 2. Some major entities of domestic violence Concept/Classes

Definition

Husband

A male person who is legally married and in relation with his spouse

Wife

A female person legally married and in relation with her spouse

Wilful conduct

A nature which force a women to self-immolation or invoke herself to injury or any self destruction (physically or mentally)

Valuable security

The word “Valuable Security” mainly related to any documented property with legal rights [16]

Figure 4 represented the snapshot of the entities along with their disjoint property and their visualization using OWLViz editor. According to the formal definition of ontology as mentioned in Table 1 and OWLViz of major classes as shown in Fig. 4, the set of concepts that represent entities (Cc ) along with the set of taxonomic relations between concepts(H) and some individuals are identified as: Cc = {Husband, Section_Charges, Cruelty, Marriage, Punishment, Harassment, Unlawful_Demand, Wilful_Conduct, Relative_Husband, Wife}.

584

T. Das et al.

Fig. 4. OWLViz of major classes

H = {kind_of (Term_Years, Punishment), kind_of(Fine_Amount, Punishment), kind_of(Physical_Harassment, Harassment), kind_of(Mental_Harassment, Harassment), kind_of(Property, Unlawful_Demand), kind_of(ValuableSecurity, Unlawful_Demand)}. I = {is_a(Gold, ValuableSecurity), is_a(PhysicalTorture, Harassament), is_a(Self_Injury, Wilful_Conduct) represents Gold is an instance of ValuableSecurity. Similarly, PhysicalTorture and Self_Injury is an instance of Harassment and Wilful_Conduct. The above-mentioned classes and their subclasses represent the knowledge behind the major entities present in cases of domestic violence. 4.2 Representation of Object Properties as Relationships Knowledge is needed to be represented in a format that allows the type of non taxonomic relationships. In this section the relationships are added between the classes to present the conectivity among them. For example, as per domain expert in domestic violence cases the wife is coerced into making an illegal demand for more dowry, whether in form of money or property. As a result, in order to express the relationship, the authors linked the entities “Wife” and “Unlawful_Demand” together using the relation “hasAskedFor”. Similarly, connection between the entities “Marriage”, “Husband” and “Wife” is

Ontology Based Knowledge Visualization

585

represented by the relationship “isHeldBetween”. Some of those relationships are clearly mentioned in Table 3. Table 3. Some relationships along with their descriptions Object properties/Relationships

Relationship type

Descriptions

HasAskedFor

Has-a

Has Wife ever been asked for any unlawful demand?

HasChargedWith

Has-a

What section has been charged?

HasForced

Has -a

Has any one forced Wife for any unlawful demand?

HasProvedThenLeadsTo

Has-a

IsLikelyToDrive

Is-a

If evidence showing any harassment then it leads to Section 498A ˙Is there any force to drive Wife for some wilful conduct as described in Table 3

According to the formal ontology definition as described in Table 1, some of the relationships are identified as: R = {hasAskedFor(Wife, Unlawful_Demand), hasChargedWith(Husband or Relative of Husband, Section_Charges), hasForced(Husband or Relative of Husband, Wife), isHeldBetween(Marriage, Husband and Wife)}. 4.3 Representation of Data Values and Axioms Data properties are features of Protege that allows for the association of individuals with their respective values. For example, Women who have been victims of domestic abuse often endure humiliation in front of their families, which can be represented as data property type_Harassment associated with entity Mental_Harassment. Figure 5 represents snapshot of the data property feature like type_Harassment linked with entity “Mental_Harassment” having datatype String. Similarly, other data properties of the individuals can also be represented. According to the formal ontology definition as described in Table 1, some of the relation with data property and datatype along with some axioms are identified as: P = {hasAskedFor(Dowry_Given,String), isLikelyToDrive (type_Wilful_Conduct,String)}where. hasAskedFor(Dowry_Given,String) representing Dowry_Given data value having String datatype associated with a relationship hasAskedFor. Likewise, other data values can be defined with some datatype. Some axioms are identified as: Instance (I, Ci ) => is_a( I, Ci ) representing I instance of a class Ci and is_a(I,Ci ) representing the class Ci having an instance I. Object_Property(I, R, Cc ) representing I as an instance associated by R nontaxonomic relationship of a class Cc . The condition along with conclusion of the rules is described as:

586

T. Das et al.

Fig. 5. Snapshot of data properties of Indian Penal Code Section 498A

Instance(I1 ,Ci ) Instance(I2 ,Ci )…… Object_Property(I, R, Cc ) => is_a(I1 ,Ci ) is_a(I2 , Ci ) …. Object_Property(I, R, Cc ) representing Instances I1 , I2 , I3 , etc., of classes Ci associated having some non-taxonomic relationship with class Cc . The rule is represented with an example from a case document of domestic violence [17] as follows:

V

V

V

V

V

Instance(Upendra Rai, Husband) => is_a(Upendra Rai, Husband) representing Upendra Rai as Husband. Instance(Meena Devi, Wife) => is_a(Meena Devi, Wife) representing Meena Devi as Wife. Instance(I2 , Cc ) Object_Property(I, R,Cc ) => is_a(Upendra Instance(I1 , Cc ) Rai, Husband) is_a(Meena Devi, Wife) Object_Property(Upendra Rai, Meena Devi, isHeldBetween, Marriage) representing class Marriage is associated with a relationship isHeldBetween in between two instances Upendra Rai and Meena Devi as Husband and Wife. Similarly, the proposed ontology is able to contribute rules that represents the legal section of domestic violence in a more structured form. V V

V

V

4.4 Ontological Visualization of Indian Penal Code Section 498A Figure 6 depict the final visualization of the legal framework representing Indian Penal Code Section 498A. The final ontology framework represents the knowledge connected with the major components of Indian Penal Code Section 498A.

Ontology Based Knowledge Visualization

587

Fig. 6. Snapshot of Ontology of Indian Penal Code Section 498A

5 Performance Analysis There are numerous methods for evaluating an ontology. In this paper, the authors have gathered various documents on domestic violence like judgments, reports, case studies, etc. Some of the documents have been annotated by legal professionals to identify the major components and their associated individuals. Further, the proposed ontology is manually mapped with the annotated segments mentioned by legal professionals. Figure 7 depicted the mapping of some prime components to the annotated segments indicated by legal professionals. Authors have collected approximately 10 legal case documents. Overall, it has been found that proposed ontological visualization of legal section of domestic violence is able to encompass the entities, object properties, individuals, etc. found within the documents. The proposed ontology can only be used to map cases under Indian Penal Code 498A. As

588

T. Das et al.

Fig. 7. Snapshot of mapping of proposed ontology with annotated document by Legal Professionals

we have focused on Indian Penal Code Section 498A, our proposed ontology successfully represents the scenario till the victim (i.e. Wife) is alive, because death of the victim will attract other legal sections as per Indian Penal Code. As a result, our proposed ontology can serve as a starting point for numerous research work aimed at developing legal ontologies on Indian Law which may be extended over other criminal sections.

6 Conclusion and Future Work The major role of our proposed ontology is to present the concepts and the relationships pertaining to the Indian Penal Code Section 498A and its related attributes mentioned in criminal statutes, acts, magazines, and other legal documents. This structured representation of Indian Penal Code (IPC) Section 498A will help legal professionals to study these legal cases at a faster pace for prompt delivery of justice to the victim. It also contributes to the advancement of the method of examining the major components of legal rule that are present domestic violence cases. Furthermore, the future scope of this research work can be stated as below: a. The proposed ontology can be expanded to encompass the major portions found in domestic abuse cases, and further analyzed the judgments of domestic violence using SPARQL query language. b. It can be expanded by including dowry death section of Indian Penal Code 304B [18] and related dowry death judgments. c. It can be integrated with Artificial Intelligence based solutions to assist legal professionals in order to reduce the amount of time they spend in studying and analysing legal cases.

Ontology Based Knowledge Visualization

589

References 1. Chandrasekaran, B., Josephson, J.R., Benjamins, V.R.: What are ontologies, and why do we need them? IEEE Intell. Syst. Their Appl. 14(1), 20–26 (1999). https://doi.org/10.1109/5254. 747902 2. Bench-capon, T.: Ontologies and legal knowledge-based systems development. Knowl.Based Syst. 1, 65–86, 1992 (1993) 3. Sankat, M., Thakur, R.S, Jaloree, S.: A framework for building ontology in education domain for knowledge representation. J. Comput. Sci. IJCSIS 14(3), 3154033 (2016) 4. Spoladore, D., Pessot, E.: Collaborative ontology engineering methodologies for the development of decision support systems: case studies in the healthcare domain. Electron 10(9) (2021). https://doi.org/10.3390/electronics10091060 5. Sahri, Z., Shuhidan, S.M., Sanusi, Z.M.: An ontology-based representation of financial criminology domain using text analytics processing. IJCSNS Int. J. Comput. Sci. Netw. Secur. 18(2), 56 (2018). http://paper.ijcsns.org/07_book/201802/20180208.pdf 6. El Ghosh, M., Naja, H., Abdulrab, H., Khalil, M.: Ontology learning process as a bottom-up strategy for building domain-specific ontology from legal texts. In: ICAART 2017 - Proceedings of the 9th International Conference on Agents and Artificial Intelligence, vol. 2, pp. 473–480 (2017). https://doi.org/10.5220/0006188004730480 7. Delhi, N.: Tackling Violence Against Women: A Study of State Intervention Measures Investigator: Bhartiya Stree Shakti (2017) 8. Krishnakumar, A., Verma, S.: Understanding domestic violence in india during COVID-19: a routine activity approach. Asian J. Criminol. 16(1), 19–35 (2021). https://doi.org/10.1007/ s11417-020-09340-1 9. Reddy, P.S., Reddy, P.R.: Of cruelty by husband or relatives of husband. In: Criminal Major Acts, 11th edn., ch. XX-A, pp. 192–193 Asia Law House, Hyderabad, India (2010) 10. Valente, A.: Types and roles of legal ontologies. In: Benjamins, V.R., Casanovas, P., Breuker, J., Gangemi, A. (eds.) Law and the Semantic Web. LNCS (LNAI), vol. 3369, pp. 65–76. Springer, Heidelberg (2005). https://doi.org/10.1007/978-3-540-32253-5_5 11. Bench-Capon, T.J.M., Visser, P.R.S.: Ontologies in legal information systems: the need for explicit specifications of domain conceptualizations. In: Proceedings of the International Conference on Artificial Intelligence and Law, pp. 132–141 (1997). https://doi.org/10.1145/ 261618.261646 12. Hoekstra, R., Breuker, J., Di Bello, M., Boer. A.: LKIF core: principled ontology development for the legal domain. Front. Artific. Intell. Appl. 188(1), 21–52. https://doi.org/10.3233/9781-58603-942-4-21 13. Wyner, A., Hoekstra, R.: A legal case owl ontology with an instantiation of popov v. Hayashi. Artif. Intell. Law 20(1), 83–107 (2012). https://doi.org/10.1007/s10506-012-9119-6 14. Lame, G.: Using NLP techniques to identify legal ontology components: concepts and relations. Artif. Intell. Law 12(4), 379–396 (2004). https://doi.org/10.1007/s10506-0054160-3 15. Faria, C., Serra, I., Girardi, R.: A domain-independent process for automatic ontology population from text. Sci. Comput. Program. 95(P1), 26–43 (2014). https://doi.org/10.1016/j.scico. 2013.12.005 16. Reddy, P.S., Reddy, P.R.: Valuable security. In: Padala Srinivasa Reddy, ch. 1, p. 9. Asia Law House, Hyderabad (2010) 17. The, I.N., Court, H., Judicature, O.F., Patna, A.T.: Patna High Court Upendra Rai vs State of Bihar & Anr on 12 January, 2017, no. 12180, pp. 1–21 (2017) 18. Khastgir, J.: Criminal Manual, vol. 469 (2012)

Application for the Management of Sports Performance in CrossFit Supported by an Artificial Intelligence Cognitive Service J. Oliveira, S. Nicola(B) , P. Graça, S. Martins, and T. Gafeira Instituto Superior de Engenharia do Porto, Porto, Portugal {joao.oliveira,paula.graca,silvia.martins, tiago.gafeira}@doitlean.com, [email protected]

Abstract. Nowadays, there is a massive concern with resource management and monitoring the status and evolution of performance over time. The same is true in the domain of health, namely in the sporting aspect, once it is important to have tools that enable the analysis and measurement of the evolution of the physical condition over time. Considering specifically the market of management and performance monitoring platforms in CrossFit, their contents do not reveal unique and innovative features that improve the sporting experience of those who use them. Thus, the main goal of this project was the development of a digital platform (application) that would allow the control, analysis, and recording of sports progress over time. To this end, OutSystems’s low-code development platform was used to build an application that met the expectations of gym owners, coaches, members, and athletes. This new solution stands out from the rest of the competition, as a new way of registering daily workouts was implemented, using artificial intelligence to extract handwritten text. In user acceptance tests, it was possible to verify that the app proved to be a solid and helpful option in improving CrossFit’s sporting experience and analysing and monitoring the performance over time. Keywords: CrossFit · Sports performance management · Data analysis · Low-code · OutSystems · Artificial Intelligence

1 Introduction In the digital age we live in, the organisation and visualisation of data that quantify or assess our behaviour in a given spectrum become increasingly relevant and recurrent. When it refers to health, namely the sporting spectrum, one of the decisive factors for success is a training plan suited to the athlete in question and its correct execution. However, the most significant difficulty inherent in this factor lies in identifying the correct stimulus, at the right time, by the coach when carrying out the training plan for his athletes (M¸eyk and Unold 2011). © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 590–598, 2022. https://doi.org/10.1007/978-3-030-96299-9_56

Application for the Management of Sports Performance

591

As Harrison and Bukstein (2016, p. 3) refer, converting the raw data of physical exercise into something measurable allows those who schedule the training and those who perform it to understand which performance points need improvement. The development of this process involves a certain complexity and scientific knowledge in all traditional sports (e.g., weightlifting, gymnastics, athletics, swimming, among others). However, this complexity increases in the presence of “hybrid” sports such as CrossFit. In basic terms, CrossFit consists of a high-intensity functional training methodology encompassing several traditional sports exercises. Its origin took place around 1996, in the United States, by Greg Glassman. Since then, it has been the sport with the highest growth rate, in terms of practitioners globally (Wagener et al. 2020). 1.1 Problem Identification There is a need to solve a research problem in this area, since the solutions on the market that allow the management and monitoring of the performance of CrossFit practitioners are expensive and not very agile. Each platform has similar features that make it difficult to stand out from the other alternatives. From a general perspective, the market has a great monotony of solutions for this type of domain, which is why it is relevant to develop a solution that meets the real needs of the users, allowing the registration, analysis, and control of sporting progress over time in a balanced and intuitive way. On the other hand, Artificial Intelligence (AI), particularly Machine Learning (ML), is increasingly accentuated in people’s daily lives. According to Silva (2020, pp. 26–27), AI “is central to data analysis, both in quantity and in-depth” reaching “an incredible level of accuracy”, since it “provides almost human interactions with different software and offers support in decisions and specific tasks”. In this way, it is worth applying this type of technology to the CrossFit area since, besides being an innovative proposal in terms of market, it improves the interaction of users with the platform, motivating them to achieve their goals. 1.2 Research Questions and Objectives Given the research problem set out in the previous point, the following research questions arise: – What kind of functionalities should a digital application contain so that it is possible to manage, analyse and monitor sports developments in CrossFit? – What or which cognitive services of Artificial Intelligence are most appropriate for improving the sporting experience in CrossFit? – What added value does the introduction of the new solution on the market bring to its users? Thus, to answer the research questions presented, the general objective inherent to them is defined by the development of a digital platform (application) that allows the

592

J. Oliveira et al.

registration, analysis, and control of sporting progress over time, supported by one or more AI cognitive services. In turn, several specific goals will have to be met for the overall goal to be achieved. Thus, it will be necessary: i)

Identify the solutions with the largest number of users in the market and analyse the content they offer; ii) Carry out a value analysis from the customer’s perspective through a quantitative and qualitative survey; iii) Implement the most relevant features to create value, according to the user’s real needs; iv) Test the solution developed in a real-environment, through a field study, and prove its usefulness through acceptance tests.

2 Methodologies 2.1 Design Science Research (DSR) For a given problem to be solved, several paths can be adopted to reach its solution, although some are clearly more suitable than others. Avison et al. (1999) mentioned that the most appropriate approach depends on the problem area, theme, and research questions raised. Hevner et al. (2004) claim that DSR constitutes strong advantages in solving practical problems, due to its pragmatic aspect, as it seeks to use theory to improve practice. According to the authors, “knowledge and understanding of a design problem, as well as its solution, are acquired in the process and application of an artifact” (Hevner et al. 2004, p. 82). DSR should address problems that have never been solved uniquely and innovatively or improve existing solutions, making them more efficient and effective. Hevner et al. (2004) also define three processing cycles that must be considered by the investigator, namely: relevance cycle, design cycle and rigour cycle. Figure 1 shows an adaptation of the scheme developed by them, considering the theme of this specific project.

Fig. 1. Design science research processing cycles

Application for the Management of Sports Performance

593

The relevance cycle addresses the connection of design activities with its application environment, considering requirements that combine the affected people and organisations, and the technologies used. In turn, the rigour cycle connects design activities with the knowledge base, according to appropriate information bases and contributions. At the centre is the design cycle that represents the interaction of construction and evaluation of the artefact(s) to be developed (Brendel et al. 2018). In general, the DSR method can be seen as an instrument aimed at structuring artefacts that enable the resolution of practical problems, in a real environment, and, in this way, lead to the generation of knowledge (Ribeiro 2015). 2.2 Agile Methodology - Scrum According to Nerur et al. (2005), the software development process is characterised by being a complex activity composed of tasks and requirements that encompass a high degree of variability. In this specific project, the software development was guided by the framework Scrum. Scrum is an agile methodology for planning, developing, and managing software projects (Schwaber e Sutherland 2017), and it promotes an iterative and incremental approach, enabling, at any point in the project, the occurrence of structural or functional changes in the solution to be implemented.

3 Solution Design and Implementation 3.1 Value Analysis In order to develop something innovative that could effectively stand out from the rest of the market, were investigated the main needs, not only of gym owners, but also of coaches and athletes. CrossFit’s current performance management and monitoring solutions appear to follow a certain pattern that does not give them a clear competitive advantage over the rest of the competition. The functionalities are the same from application to application, with few differentiating factors that allow for added value to be created. Thus, a survey was made available for a period of two weeks, to find out which needs were not being met by users, which features (of the application they use daily) needed significant improvement and what unique content could be added. In total, 197 responses were collected. Upon analysing the results of the surveys, it was concluded that around 40% of respondents were not satisfied with the application they use (Fig. 2), and of this 40%, 17.3% showed a neutral (indifferent) opinion. On the other hand, when analysing the answers to the qualitative questions, it was found that the main weaknesses highlighted by the respondents were related to: – the unattractive and confusing design/layout of the application they use; – the frequent appearance of bugs and, consequently, slow loading of application screens;

594

J. Oliveira et al.

Fig. 2. Degree of satisfaction of respondents with the application they use

– the fact that the user does not have access to all the historical data for a personal record or the day they register something new; – the impossibility of recording personal workouts in a freeway, without being dependent on the options present in the app; – the inefficiency of resources when creating the service calendar, as it is not possible to create the necessary time slots at once for a given day; – the fact that it presents a list with little detail about the movements/exercises, and it is necessary to use YouTube to view their demonstration; – the restrictions it imposes on the registration of personal records, since it does not allow registering a new PR for a movement/exercise that is not on the app’s list; – the cancellation of services, since to cancel a registration in a given class, the user must individually consult the days of the week in the services calendar, there being no list where all scheduled attendances are registered; – the lack of graphs that show the evolution over time. Considering the facts highlighted in the previous points, it became clear that it would be necessary to develop two variants of the same application: a web application for gym owners and coaches and a mobile application intended only for gym members and athletes. In general, these are the basic contents for an application of this kind to become functional in the real world. However, the goal was also to implement something that, in a way, would revolutionise and innovate what existing platforms have on the market. In this sense, two possibilities emerged that could provide the desired innovation. The first was related to the recognition of voice commands, giving users the possibility to enter their results, comment on their performances and perform quick searches in the application using voice commands. On the other hand, the second option involved a feature that would help the user insert his own workout into the application. In CrossFit, it is common for workouts to be written on a board, so when inserting the daily workout into the app, users would be able to register it through a photo, and the application would automatically transcribe the content of the submitted image into the appropriate field, removing the task of typing it manually.

Application for the Management of Sports Performance

595

Thus, in the surveys conducted, respondents were asked which functionality is most useful for their day-to-day life, and what benefits could improve their interactive experience with the app. It was then concluded that the training digitisation was the most favoured feature by the respondents, as it obtained 76.1% of the votes (Fig. 3).

Fig. 3. Respondents’ preference regarding the innovative functionality to be implemented

3.2 Architecture In IS projects, one of the first steps to be taken is the architectural design. This design allows the connection between the design and the requirements engineering, involving a high-level understanding of the functioning of a given system, its constituent components, and how they interact with each other. This specific project consists of 4 components: the web component (BackOffice), the mobile component, the OutSystems Platform, and Microsoft Azure (Fig. 4).

Fig. 4. Project architecture

For the system to be operational it is necessary to have an internet connection, since both the web and mobile aspects are supported by the data present on the OutSystems

596

J. Oliveira et al.

server. Whenever the user intends to insert new data or edit/delete existing ones in the database, this procedure is carried out through server actions, hence the bidirectional relationship between the server and the web and mobile devices. In turn, communication with Microsoft Azure services is carried out through a REST API, where the user, on his smartphone, sends an image file to the cloud and receives from it the text present in the image he has submitted.

4 Solution Validation To validate the developed app, it was necessary to run a series of tests to identify possible errors and improvement opportunities. With that being said, the tests performed were: unit tests, integration tests, system tests, and user acceptance tests. In the case of the last type of test mentioned, to analyse in detail the experience lived during the user acceptance tests, a satisfaction survey was distributed to a sample of 31 people who had the opportunity to test the app. This survey consisted of 10 close-ended questions and 3 open-ended questions. Since the developed solution is intended solely and exclusively for the practice of CrossFit, and since the target audience is formed only by gym owners, coaches, and athletes/members of a CrossFit gym, the tests were developed at CrossFit Leça do Balio. Here, the gym owner had the opportunity to test the web app and 30 members the mobile one over a period of approximately three weeks. It should also be noted that the mobile application was only tested on Android smartphones, since the OutSystems platform only allows the generation of free trial versions for this operating system. Regarding the survey, 90.3% of respondents considered that the application is a solid and valid solution to meet the daily needs of users, as CrossFit practitioners (Fig. 5).

Fig. 5. Answers to the question: “Given the solutions already on the market, do you think the app presents itself as a solid and valid option to respond to user’s needs?”

The appealing design, the fast response times and the clean and professional UI were mentioned as key points for the success of the developed solution. On the other hand, respondents also highlighted the “Scan WOD” feature as the main factor that distinguishes the developed app from the other solutions on the market.

Application for the Management of Sports Performance

597

As mentioned before, the involvement of cognitive services in CrossFit’s performance monitoring and control platforms is virtually non-existent. Thus, the workout digitisation, through Computer Vision, presents itself as a very powerful and revolutionary tool to optimize the user’s sporting experience. Considering all the facts presented above, 83.9% of respondents rated the developed application with a grade of 8 or more on a scale from 1 to 10 (Fig. 6).

Fig. 6. Answers to the question: “On a scale of 1 to 10, how do you rate the application?”

5 Conclusions and Future Work In a global point of view, the project was successfully completed as all research questions were answered. The main objective was not only to build just an application for control and management of the sports performance, but also to promote the user’s commitment to physical exercise, in this specific case, through CrossFit. This sport not only makes its practitioners fitter, but also improves their health all around and their sense of well-being. With that being said, by analysing all the answers given in the survey of the acceptance tests, it was possible to conclude that the app obtained a high degree of satisfaction, being considered by over than 90% of respondents as a solid and valid solution to manage their workout routine and analyse their performances over time. In the future, it would be relevant to implement more features that can improve the user experience, such as smartwatch/smartwatch connection, calories intake management, percentage automatic work calculation, private chat to enable the communication between athletes and coaches and provide a fitness level based on a specific result in a certain workout. In short, it is possible to say that the structure of the developed solution promotes an attractive environment for the recording, control, and analysis of sports performance in CrossFit, offering a significant market innovation through the integration of cognitive AI services (never before implemented in this type of platforms). The product’s design combined with the unique features that it presents constitute a strong asset for the fitness industry, as it materialises the user’s actual needs in a practical and intuitive way.

598

J. Oliveira et al.

References Avison, D., et al.: Action research. Commun. ACM 42(1), 94–97 (1999). https://doi.org/10.1145/ 291469.291479 Brendel, A.B., Zapadka, P., Kolbe, L.: Design Science Research in Green IS: Analyzing the Past to Guide Future Research (2018) Harrison, C.K., Bukstein, S.: Sports Business Analytics. 1st edn. CRC Press, Boca Raton (2016). https://doi.org/10.1201/9781315367613 Hevner, A.R., et al.: Design science in information systems research. MIS Quar. Manage. Inf. Syst. 28(1), 75–105 (2004). https://doi.org/10.2307/25148625 M¸eyk, E., Unold, O.: Machine learning approach to model sport training. Comput. Hum. Behav. 27(5), 1499–1506 (2011). https://doi.org/10.1016/j.chb.2010.10.014 Nerur, S., Mahapatra, R., Mangalaraj, G.: Challenges of migrating to agile methodologies. Commun. ACM 48(5), 72–78 (2005). https://doi.org/10.1145/1060710.1060712 Ribeiro, L.R.A.: Data Analytics: Abordagem para Visualização da informação. Master Thesis, Universidade do Minho (2015). http://hdl.handle.net/1822/40314 Schwaber, K., Sutherland, J.: The Scrum Guide (2017) Silva, B.: Inteligência artificial e suas implicações ético-jurídicas. Master Thesis, Faculdade de Direito da Universidade NOVA de Lisboa (2020). http://hdl.handle.net/10362/104098 Wagener, S., et al.: CrossFit® – development, benefits and risks. Sports Orthopaed. Traumatol. 36(3), 241–249 (2020). https://doi.org/10.1016/j.orthtr.2020.07.001

Perceptions of Cloud Computing Risks in the Public Sector Bonginkosi Mkhatshwa and Tendani Mawela(B) University of Pretoria, Hatfield, South Africa [email protected], [email protected]

Abstract. The transformation of public services through the adoption of cloud computing has shown greater advancement with service delivery efficiencies and reduction of costs. Cloud computing has been in existence for some time. However, the adoption has been seemingly low and stagnant in the South African (SA) government context. This study investigates the stakeholders’ perceptions in the SA public sector regarding the risks associated with cloud computing adoption. The study was exploratory, qualitative and interpretative in nature. Data was collected via semi-structured interviews and a questionnaire with IT representatives within government departments. The findings highlight that the perceived risks include security breaches, lack of supporting network infrastructure and concerns on redundancy. Additionally, poor skills and contract management were noted. This was coupled with concerns about compliance with relevant legislation and risks associated with transborder information flows. The study contributes to the literature on cloud computing in government. Keywords: Cloud computing · IT adoption · Risk · Digital government

1 Introduction 1.1 Background Governments are operating in an increasingly complex environment and are faced with limited resources to fulfil their mandates. Information and Communication Technologies (ICT’s) are key enablers for government service delivery. Cloud computing is one of the ICT solutions that the public sector may consider to increase demand for more efficient and effective service delivery from governments [1]. Cloud computing involves the delivery of technological resources on-demand through the Internet. As defined by scholars, cloud computing is about accessing Information and Communication Technology (ICT) services through the network without ownership of the infrastructure, software, or any specific platform [2, 3]. It is noted that the recent experiences with the Covid-19 pandemic highlighted that government entities across the globe were forced to reconsider how they deliver services, and many resorted to cloud technologies to enable them to deliver services in a time of crisis [4]. Cloud computing has been in existence for some time now, but there are countries such as Saudi Arabia that have not yet adopted this technology [5]. In South Africa (SA), one of the “Batho Pele” (People First) principles © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 599–611, 2022. https://doi.org/10.1007/978-3-030-96299-9_57

600

B. Mkhatshwa and T. Mawela

stated in the White Paper for Transforming Public Service Delivery issued by the then Minister Skweyiya from the Department of Public Service and Administration (DPSA) is the principle of value for money, which implies that public services should be rendered prudently and efficiently [6]. Adopting cloud computing may enable the SA government to accomplish the Batho Pele principle of value for money by saving on costs currently spent on IT software, infrastructure, and other developed systems. It is anticipated that the SA public service may yield benefits in adopting cloud computing and transforming the IT environment by changing how it operates and its current configuration. The South African (SA) State Information Technology Agency (SITA) recently launched the cloud infrastructure for the SA government in 2018. Most SA government departments have not fully adopted cloud computing as an instrument for modernising and transforming IT. The purpose of this study was to understand the perceptions of SA government representatives towards the risks associated with adopting cloud services in their environments. The research question that guided the study is: What are the perceptions of the public sector towards cloud adoption risks? The paper is organised as follows: next, an overview of the informing literature is provided followed by an overview of the research methods. Then the findings from the study are presented and discussed. The last section reflects on the limitations and highlights opportunities for future research.

2 Informing Literature 2.1 Cloud Computing Overview Cloud computing is a method utilised to facilitate expediency, and access networks through an on-demand shared supply of infrastructure, software, networks, and storage services that are configured for a quick supply and delivered with minimal control effort or vendor interaction [7]. Research suggests that cloud computing is an area that is anticipated to transform the IT function [8]. It is a customer or supply model where IT capabilities are accessed by paying for services based on the usage model. There are various Cloud Computing service models, namely Infrastructure as a Service, Platform as a Service and Software as a Service. Infrastructure as a Service (IaaS) is a service model where the cloud computing vendors provide IT infrastructure services to organisations which are employed for operations such as servers, hardware, network components and storage instead of consuming the on-premises services [9]. The Platform as a Service (PaaS) is when a client consumes the procured functional programs hosted on the cloud infrastructure [10]. PaaS is the infrastructure that is consumed as a service in addition to the servers, operating systems and applications hosted on the server, such as web servers [9]. Software as a Service (SaaS) is the cloud computing service where the software applications are consumed as a service through rental over the Internet. The SaaS is also regarded as a web-based and on-demand software [9]. The literature indicates a number of essential characteristics of cloud computing as follows [11, 12]: On-Demand Self-Service - The consumers can automatically offer computing abilities and resources on their own when required without any need for human interference. Broad Network Access - Access and abilities are applicable through the network over

Perceptions of Cloud Computing Risks in the Public Sector

601

standard devices, laptops and also PDAs or cellphones, etc. Resource Pooling - The resources such as the processing power, virtual machines, storage capacity, memory, network bandwidth, etc., are joined together to serve several clients using a multi-tenant model. The physical and virtual resources are dynamically allocated and reallocated. Rapid Elasticity - It has a dependency on resources, demand and capabilities that can be rapidly and automatically set up and scaled at any magnitude and any specific period. Measured Service - The client’s usage of the seller’s services and resources are automatically controlled, monitored, and reported contributing to the highest degree of transparency for the service provider and the customer. There are also various deployment models related to cloud computing: Private Cloud - The cloud infrastructure is established and functioning exclusively for a specific enterprise [9]. It may be managed by a third party or the company and may exist off-premises or on-premises. Public Cloud - The cloud infrastructure is made available to the general public or a large industry group and is owned by an organisation selling cloud services [10]. Hybrid Cloud - A hybrid cloud is a combination of the private and public cloud models. It is an environment where some internal and external cloud suppliers are employed. Community Cloud - The cloud infrastructure is shared by numerous entities and supports a particular community that is sharing the same concerns (e.g., compliance concerns, security requirements and mission policy) [8]. It may be managed by a group of businesses or a specific third party. 2.2 Cloud Computing Opportunities and Benefits Cloud computing is said to hold manifold benefits for organisations. It is indicated that cloud computing solutions have the ability to speed up business strategies for government departments and public sector agencies [13]. Government departments, agencies, local government and various organisations, including for-profit and non-profit, are anticipated to have more rapid, enhanced and economical IT services when they adopt cloud computing. Furthermore, entities may even go beyond what is anticipated by reaping cost scalability, significant agility, enhancement of business processes, greater corporate customer satisfaction and infrastructure scalability benefits from cloud adoption. [14]. The public sector’s key value attained from cloud computing services is cost reduction through improved system effectiveness [15]. The public sector can decrease its capital expenditure on IT equipment, internal servers, networking tools, storage resources, and software by moving to a “pay as you use” based cloud model [16]. 2.3 Cloud Computing Challenges and Risks The uptake of cloud computing is faced with several challenges and perceived risks. According to scholars [17], organisational readiness is a critical aspect that may hinder the adoption of cloud computing. Therefore, it facilitates specific essential basic services to be cloud-ready, active and optimally running. There are seven identified issues that may prevent the adoption of cloud computing. These are performance, convenience, compliance, integration, private cloud, security, and costs [18]. The cloud service providers are deemed to have access to data. They may unintentionally or intentionally release or utilise it for illegitimate purposes, which causes serious confidentiality and privacy

602

B. Mkhatshwa and T. Mawela

apprehension [5]. The ICT infrastructure expansion in South Africa is deemed as a critical aspect that hinders the progression of cloud adoption [19]. The issue of data sovereignty also hinders the adoption of cloud computing in public service as this is deemed critical for data protection and privacy of information [20]. Compatibility is required to determine whether the private or public cloud is compatible with the technology architecture used by the public bodies [21]. Furthermore, the degree to which cloud solutions are compatible with the applicable systems and the statutory regulation is critical when deciding to migrate to the cloud environment [22]. This suggests that the lack of compatibility between the application and information systems architecture of government systems and the cloud applications or platforms may be a challenge for cloud adoption. According to prior research [23], lack of budget and high costs for cloud computing services may be a barrier for cloud computing adoption in the public service.

3 Research Methods 3.1 Research Design The research was exploratory, qualitative and based on the interpretive philosophical stance. The SA government was chosen as a case for this study. The CIO’s, IT managers and IT technical support staff from various government institutions were identified through purposive sampling as respondents for the study. Data was collected through semi-structured interviews and a qualitative survey questionnaire. The study had a total of 32 respondents from various government departments. The qualitative data was analysed using the inductive thematic analysis approach [24] supported by the Atlas. ti tool. 3.2 Theoretical Framework The researchers leaned on the “Technology – Organisation – Environment” (TOE) Framework to assist in,analysing and understanding the cloud adoption risks perceived by the respondents in the study. The TOE highlighted that there are three main aspects that inform an organisation’s technology innovation adoption decisions. These aspects are the technological, organisational, and environmental perspectives. These elements were hypothesised as affecting technological innovation [25]. The TOE framework is used to identify technological, organisational, and environmental issues as they are related to the perceived risks of cloud computing adoption in the context of the SA government.

4 Findings and Discussion The following section discusses the respondents’ views as they pertain to the perceived risks of cloud computing.

Perceptions of Cloud Computing Risks in the Public Sector

603

4.1 Technological Factors Security. Security was identified as one of the critical areas that were flagged as a risk for cloud computing adoption. After analysing and interpreting the data, security was noted as a risk that must be managed when adopting cloud computing. Within the security theme, the following issues were noted: data security and security breaches, data sovereignty, data residency, data protection and malicious use of data. Participant 4 indicated that “the nature of work is that they deal with sensitive information so securing that information is very critical”, while this was substantiated by Participant 9, who indicated that “loss of information, access to information by unauthorised per-sonnel is a risk”. Additionally, Participant 9 was quoted as saying, “We need security around the information so that it cannot be hacked, misused, deleted or altered in some form or another”. This was confirmed by Participant 2, who said: “remember government information is sensitive if someone hosts that information then they need to make sure that security measures are put in place because if that information leaks and falls into the wrong hands, how are you going to answer for that”. The issue of data security was confirmed by the literature as one of the major risks for cloud computing adoption [26]. The security concern around data sovereignty came as through one of the critical risks for cloud adoption as Participant 8 even asked a question: “sovereignty, in this case, would be its data and security around that data, therefore to what level are we creating organic South African data hosting services?”. Another comment came from Participant 8, who indicated that “my big issues with cloud are what happens if tomorrow you have a funding issue and you can’t pay for your cloud services and they basically stop your service, and now you will be held to ransom because your information is not in your environment anymore. So, that is my fear, the sovereignty of information”. Researchers have previously raised the issue of data sovereignty as complicated and posing a risk in relation to cloud computing [20]. Network Infrastructure. The network was flagged as a risk that should be mitigated when adopting cloud computing. This risk was deemed critical for the adoption of cloud computing in public service. The issues of bandwidth, broadband, internet, and network connection were mentioned in relation to the network infrastructure risk. Participant 9 indicated that “You’re going to the rural areas, and there’s no connectivity. How does my admin officer down in the deep rural area get access to his information if he does not have connectivity?”. The same question was also asked by Participant 8, who said, “if you are going to implement a cloud service and the IT infrastructure is not right, and the bandwidth is not right, it is a slow death”. Participant 7 also stated that the “biggest challenge in the province is we do not have fast data or connectivity lines to the SITA network between departments”. While Participant 23 confirmed that “cloud computing relies much on internet connectivity and if it is not stable, service delivery may be negatively impacted”. Most of the issues were raised by participants from rural areas who emphasised that the problems raised confirmed that “the lack of robust broadband infrastructure” is a risk for cloud computing adoption. The risk concerning the network, which involves broadband, bandwidth, and internet connections, were identified as issues that must be mitigated when adopting the cloud.

604

B. Mkhatshwa and T. Mawela

Redundancy. Redundancy was also identified as a risk for cloud computing adoption. The risk concerns the redundancy of the systems or data hosted on the cloud. Participant 20 indicated that “if something goes down, many services will also go down so make sure that redundancy is well thought through”. The assertions were confirmed by Participant 2, who indicated that “data loss and not having access to do data recoveries is a risk because it is not our environment”. The redundancy was confirmed by the literature as one of the risks that should be mitigated when considering adopting the cloud. High availability of systems and data is an important challenge in cloud solutions and planning [27].

4.2 Organisational Factors Skills. The shortage of skills, understanding and knowledge about cloud computing was identified as a risk. Participant 13 indicated that: “the critical issue at this point is the human resources and understanding in terms of how we get the necessary skillsets when we move certain services into the cloud”. Participant 9 indicated that: “the first risk is that we have very few skilled people. We have very few skilled active personnel. We don’t have enough skills to make this thing work across the province”. Whereas Participant 28 mentioned that “no skilled officials, no training etc., and due to these, especially the lack of training, it brings along job insecurities due to a lack of understanding of what will happen to one’s role with the implementation of the technology”. The perspectives expressed by the respondents shows that the lack of skills, understanding and knowledge is a risk that should be mitigated. It has been demonstrated that the management, executive and other personnel should have knowledge of ICTs concepts and emerging innovations, as stressed by Participant 8, who indicated that “you can’t be a financial person and not have a basic understanding of digital literacy like what is cloud computing, you can’t be a Head of Department and not understand the basics of artificial intelligence, etc., so technology or cloud computing or any other thing should not be seen in isolation of other professions”. Cloud computing is noted in the literature as one of the foremost skills desired by employers [28], and this includes organisations such as the government. Contract Management. As noted from the findings, contract management was flagged as one of the critical risks when cloud computing is adopted in the public sector. According to Participant 30, “the reputable service providers must be used with security measures in place.” At the same time, Participant 29 raised the issue of vendor lock-in, stating that the government should “scrutinise the terms and understand the conditions for scaling up and down as well as exiting from each cloud provider”. This suggests putting the necessary controls around managing the cloud service providers that are providing cloud services to the government. This was affirmed by Participant 12 who indicated that “window locking and not being able to terminate the contract for cloud services with the service provider is a risk”. This suggests the importance of contract management of cloud computing services and the clarity required on the agreement between the cloud service provider and government institution receiving the cloud services.

Perceptions of Cloud Computing Risks in the Public Sector

605

4.3 Environmental Factors Compliance. Compliance was identified as a risk for the adoption of cloud computing in the public sector. The issues of compliance were raised by Participant 25, who indicated that “compliance-related issues, including 3rd party compliance-related issues and data breach”, are risks that are to be mitigated when adopting cloud computing. This risk has been clearly articulated in the literature [29]. Compliance with regulation has been identified as one of the significant risks that deserves attention before consuming cloud services [30]. Participant 25 further indicated that “for compliance, third parties should sign the non-disclosure agreements as well as sign and acknowledge the internal security policy”. This affirms that the compliance risk cannot be avoided but is a matter that should be taken seriously when managing the cloud providers to ensure that they comply with all the necessary measures that are put in place. Participant 7 confirmed this finding by indicating that “In terms of assurance of data security and privacy of data, it can be mitigated by compliance to the Protection of Personal Information Act (POPI Act) so that security is dealt with”. The issue of data hosted within the SA borders was also flagged as an integral part of compliance risk. This risk was supported in the literature [29] concerning compliance with the regulation as one of the key risk factors that should be taken into account before consuming cloud services when hosted through one of the cloud vendors. Transborder Information Flow. The analysis of the data also highlighted risks concerning the transborder information flow. The transborder information flow forms an integral part of the risks that were perceived as risk factors influencing cloud computing in the South African public sector’s perspective. Participant 12 stated that “you may think that the data is stored in SA, only to find that it is stored outside of the country”. This perspective suggests that sometimes organisations may not know as consumers that the data is not stored within the South African borders. This is supported by Participant 11, who indicated the risk of hosting and the need to check where that data is held. Participant 28 indicated that “the only risk that had been identified was data residency”. This same participant suggested that data residency is crucial. The transborder information flow outside the borders of South Africa poses a severe risk to cloud adoption in SA. This is because the information stored outside the SA borders may lead to data loss, as stated by Participant 11. It has also been revealed that companies may end up being, hesitant to be accountable for breaking the law that protects consumer privacy on behalf of the cloud solution vendor [1]. According to researchers [31], the control, transfer, storage, or dispensation of personal information outside of South Africa hinders cloud solutions that are highly dependent on redundant storage and exchang-ing data outside of South Africa. It is also indicated that the usage of cloud compu-ting services leads to regulatory and policy challenges [32]. Furthermore, cloud computing has been deemed as a solution that increases security and privacy concerns, particularly when it contains various cloud platforms around several countries with different legal systems that are not in accord [32]. The extant literature highlights that the data or information that is stored outside of South Africa’s borders must abide by the prescripts stated within the POPI Act [32, 33].

606

B. Mkhatshwa and T. Mawela

4.4 Perspectives of Cloud Computing Risks from the Literature The risks raised by the respondents highlight and echo what other scholars have noted regarding cloud computing. Table 1 presents several risks identified in the literature about cloud computing: Table 1. Overview of risks for cloud computing TOE factors

Risks

Description

References

Technological factors

Security

The identified security risks associated with cloud computing adoption are around data privacy, data protection, data sovereignty, data residency and security breach

[14, 34–39]

Network The lack of network infrastructures such as [34–36, 39] infrastructure broadband, poor bandwidth and lack of internet connection, especially in rural and remote areas, has been consistently identified as a risk that should be effectively managed for cloud adoption Redundancy

Organisational Contract factors management

Skills

Lack of assurance from the service provider [27] that the data hosted on the cloud environment can be recovered if it can be lost or if the service provider ends up bankrupt. High availability is an important challenge in cloud solutions and planning Poor management of contractual terms and agreements. At the end of the cloud computing services contract, data recovery or migration issues were identified and proven to be risks for cloud computing adoption. Vendor locking into an existing contract is also a contractual risk that is required to be managed effectively

[34–36]

Lack of skills, understanding, relevant [34–36, 40] experience and knowledge from internal resources for managing and maintaining cloud computing platforms has been identified and consistently substantiated as a major risk for adoption of cloud computing (continued)

Perceptions of Cloud Computing Risks in the Public Sector

607

Table 1. (continued) TOE factors

Risks

Environmental Compliance factors

Transborder information flow

Description

References

The lack of compliance with domestic regulations concerning the protection of personal data, data residency, data sovereignty and applicable procurement processes is another risk. Internal security policies and procedures are being disregarded in the cloud environment

[20, 36, 39, 41]

Transborder information flow has been [31, 32, 42–44] deemed as a critical risk to be managed. The transborder information flow is one of the risks that are to be effectively mitigated prior to the adoption and deployment of any cloud computing adoption initiative. Transborder information flow may lead to an organisation not complying with legal prescripts such as the POPI Act, which protects the customers’ personal information

4.5 Recommendations The respondents expressed their reservations related to the adoption of cloud computing. However, they also offered several recommendations on potential mitigation approaches for the identified risks. Technological: A key recommendation is that government entities should put in place cloud policies and standards that govern the deployment and management of cloud solutions. Additionally, there is a need to focus on upgrading the network infrastructure to support the migration to cloud computing, particularly in rural areas where the connectivity is limited and constrained. One view also expressed considering a government-owned cloud technology model, thus reducing dependence on external service providers. Organisational: The aspect of financial resources for the migration to the cloud becomes pertinent. A change management plan, continuous training and upskilling of government IT staff as well as raising awareness with general staff regarding cloud computing solutions, is another proposal from the respondents. This can also be coupled with the recruitment of specialist cloud computing human resources. Government procurement departments should ensure that cloud services are procured from reputable services providers. In addition, there should be well-defined service level agreements with associated penalties and a shared responsibility model that caters for instances of data breaches or security incidents to drive the accountability of service providers. Furthermore, the service level agreements should cater for issues relating to the exit strategy as an approach for dealing with the window locking situation that should be

608

B. Mkhatshwa and T. Mawela

avoided at the end of the contract. A stringent risk management approach such as the development of the risk management strategy and the risk register was also perceived as the appropriate control for mitigating the cloud computing adoption-related risks. Environmental: With regard to the protection of data and compliance with laws, policies, and regulations, this may involve the requirement that third parties and service providers sign and commit to non-disclosure agreements. Respondents also highlighted the enforcement of adherence to legal prescripts such as the POPI Act through robust awareness campaigns and insisting on controls that will protect the personal information of the South African consumers and citizens. Also, a government-owned private cloud may mitigate the risk of transborder information flows.

5 Conclusion, Limitations and Future Research 5.1 Concluding Remarks The study considered the sluggish rate of cloud computing adoption that has been noted in the public sector. The study sought to understand the perceptions regarding the key risks that government officials believe influence the adoption and uptake of cloud computing services. The respondents highlighted that the main risks they perceive to be associated with a shift to cloud computing centred around potential security breaches, a lack of supporting network infrastructure and concerns on redundancy. Additionally, poor skills and contract management were noted. This was coupled with concerns on the ability to comply with relevant legislation and risks with transborder information flows. The respondents also offered their views on potential approaches to address the risks they identified in adopting cloud computing. 5.2 Contribution, Limitations and Future Research Data was predominantly collected at the national and provincial levels of govern-ment, and it included 32 respondents. Future studies may focus on collecting data from the local government (municipality) level and increase the number of respondents. Future research can also include interviews with cloud service providers, which may also yield insights regarding the challenges that government entities face with cloud computing adoption from a supplier perspective. The study followed a qualitative and interpretive approach, other research strategies may be considered for future investigations. Acknowledgements. This work is based on the research supported in part by the National Research Foundation of South Africa (Grant Numbers 127495).

Perceptions of Cloud Computing Risks in the Public Sector

609

References 1. Gillwald, A., Moyo, M.: Modernising the public sector through the cloud. Res. ICT Africa 1, 1–53 (2017) 2. Almarabeh, T., Majdalawi, Y., Mohammad, H.: Cloud computing of e-government. Commun. Netw. 8, 1–8 (2016). https://doi.org/10.4236/cn.2016.81001 3. Wyld, D.C.: The Cloudy future of government it: cloud computing and the public sector around the world. Int. J. Web Semant. Technol. 1, 1–20 (2010) 4. Carlson, T.: Cloud technology is transforming public services. This is how (2020). https:// www.weforum.org/agenda/2020/08/cloud-technology-is-transforming-public-services-thisis-how/ 5. Al-Ruithe, M., Benkhelifa, E., Hameed, K.: Key issues for embracing cloud computing to adopt a digital transformation: a study of Saudi public sector. Proc. Comput. Sci. 130, 1037– 1043 (2018) 6. Skweyiya, Z.: Batho Pele - “People First”. In: Administration, Department of Public Service and Administration (DPSA) (Ed.). Pretoria (1997) 7. Hashim, H.S., Hassan, Z.B., Hashim, A.S.: Factors influencing the adoption of cloud computing: a comprehensive review. Int. J. Educ. Res. 3, 295–306 (2015) 8. Seke, M.M.: Higher education and the adoption of cloud computing technology in Africa. Int. J. Commun. 4, 1 (2015) 9. Hussein, N.H., Khalid, A.: A survey of cloud computing security challenges and solutions. Int. J. Comput. Sci. Inf. Secur. 14, 52 (2016) 10. Hashemi, S., Monfaredi, K., Masdari, M.: Using cloud computing for e-government: challenges and benefits. Int. J. Comput. Inf. Syst. Control Eng. 7, 596–603 (2013) 11. Makoza, F.: Cloud computing adoption in higher education institutions of Malawi: an exploratory study. Int. J. Comput. ICT Res. 9, 37– 54 (2015) 12. Amponsah, R., Panford, J., Acquah, J.: Factors affecting cloud computing adoption in a developing country Ghana: using the extended unified theory of acceptance and use of technology (Utaut2) model. Int. Res. J. Eng. Technol. 3, 59–76 (2016) 13. Muhammad, A.R.: Towards cloud adoption in Africa: the case of Nigeria. Int. J. Sci. Eng. Res. 6, 657–664 (2015) 14. Nghihalwa, E.N., Shava, F.B.A.: Secure cloud adoption framework (SCAF) for the Namibian government information technology departments. In: 2018 Second World Conference on Smart Trends in Systems, Security and Sustainability (Worlds4), IEEE, pp. 246–253 (2018) 15. Bhisikar, A.: G-Cloud: new paradigm shift for online public services. Int. J. Comput. Appl. 22, 24–29 (2011) 16. Liang, Y., Qi, G., Wei, K., Chen, J.: Exploring the determinant and influence mechanism of e-government cloud adoption in government agencies in China. Gov. Inf. Q. 34, 481–495 (2017) 17. Tweneboah-Koduah, S., Endicott-Popovsky, B., Tsetse, A.: Barriers to government cloud adoption. Int. J. Manag. Inf. Technol. 6, 1–16 (2014) 18. Sabi, H.M., Uzoka, F.-M.E., Langmia, K., Njeh, F.N., Tsuma, C.K.: A cross-country model of contextual factors impacting cloud computing adoption at universities in Sub-Saharan Africa. Inf. Syst. Front. 20, 1381–1404 (2018) 19. Adendorff, R., Smuts, H.: Critical success factors for cloud computing adoption in South Africa. a software development lifecycle view. In: SAICSIT’10 Proceedings of the 2010 Annual Research Conference of the South African Institute of Computer Scientists and Information Technologists, vol. 2010, pp. 304–313. ACM (2019) 20. Irion, K.: Government cloud computing and national data sovereignty. Policy Internet 4(3–4), 40–71 (2012)

610

B. Mkhatshwa and T. Mawela

21. Gangwar, H., Date, H., Ramaswamy, R.: Understanding the determinants of cloud computing adoption using an integrated Tam-Toe model. J. Enterp. Inf. Manag. 28(1), 107–130 (2015) 22. Alhammadi, A., Stanier, C., Eardley, A.: The determinants of cloud computing adoption in Saudi Arabia. Second International Conference on Computer Science and Engineering (CSEN 2015), 55–67 (2015) 23. Alkhwaldi, A., Kamala, M.A., Qahwaji, R.S.: Analysis of Cloud-based e-government services acceptance in Jordan: challenges and barriers. J. Internet Technol. Secur. Trans. 7(2), 556–568 (2018). https://doi.org/10.20533/jitst.2046.3723.2018.0069 24. Braun, V., Clarke, V.: Thematic analysis. APA handbook of research methods in psychology, Vol 2: Research Designs: Quantitative, Qualitative, Neuropsychological, and Biological. American Psychological Association, Washington, DC, US (2012) 25. Baker, J.: The technology–organization–environment framework. In: Dwivedi, Y., Wade, M., Schneberger, S. (eds.) Information Systems Theory: Explaining and Predicting Our Digital Society, pp. 231–246. Springer, New York (2012). https://doi.org/10.1007/978-1-44196108-2 26. Chang, V., Muthu, R.: Towards achieving data security with the cloud computing adoption framework. IEEE Trans. Serv. Comput. 9(1), 138–151 (2015) 27. Sousa, E., Lins, F., Tavares, E., Maciel, P.: Cloud infrastructure planning considering different redundancy mechanisms. Computing 99(9), 841–864 (2017). https://doi.org/10.1007/s00607016-0533-6 28. Nwokeji, J.C., et al.: Panel: incorporating cloud computing competences into computing curriculum: challenges & prospects. In: 2020 IEEE Frontiers in Education Conference (FIE), pp. 1–3. IEEE, (2020) 29. Jones, S., Irani, Z., Sivarajah, U., Love, P.E.D.: Risks and rewards of cloud computing in the UK public sector: a reflection on three organisational case studies. Inf. Syst. Front. 21(2), 359–382 (2017). https://doi.org/10.1007/s10796-017-9756-0 30. Phaphoom, N., Xiaofeng, W., Sarah, S., Sven, H., Pekka, A.: A survey study on major technical barriers affecting the decision to adopt cloud services. J. Syst. Softw. 103, 167–181 (2015) 31. Didi-Quvane, B., Smuts, H., Matthee, M.: Critical success factors for dynamic enterprise risk management in responsive organisations: a factor analysis approach. Conference on eBusiness, e-Services and e-Society, Springer, 704–717 (2019)https://doi.org/10.1007/978-3030-29374-1_57 32. Mohlameane, M., Ruxwana, N.: Exploring the impact of cloud computing on existing South African regulatory frameworks. South Afric. J. Inf. Manage. 22, 1–9 (2020) 33. Sun, J., Han, B., Ekwaro-Osire, S., Zhang, H.-C.: Design for environment: methodologies, tools, and implementation. J. Integr. Des. Process. Sci. 7, 59–75 (2003) 34. Dutta, A., Peng, G.C.A., Choudhary, A.: Risks in enterprise cloud computing: the perspective of IT experts. J. Comput. Inf. Syst. 53, 39–48 (2013) 35. Elena, G., Johnson, C.W.: Factors influencing risk acceptance of Cloud Computing services in the UK Government. arXiv:1509.06533 (2015) 36. Bannerman, P.L.: Cloud computing adoption risks: state of play. Governance 3, 2.0 (2010) 37. Samarati, P., Di Vimercati, S.D.C., Murugesan, S., Bojanova, I.: Cloud security: issues and concerns. Encyclopedia on Cloud Computing, pp. 1–14 (2016) 38. Alassafi, M.O., Alharthi, A., Walters, R.J., Wills, G.B.: A framework for critical security factors that influence the decision of cloud adoption by Saudi government agencies. Telematics Inform. 34, 996–1010 (2017) 39. Singh, A., Malhotra, M.: Security concerns at various levels of cloud computing paradigm: a review. Int. J. Comput. Netw. Appl. 2, 41–45 (2015) 40. Salem, M.M., Hwang, G.-H.: Critical factors influencing adoption of cloud computing for government organizations in Yemen. J. Distrib. Sci. 14, 37–47 (2016)

Perceptions of Cloud Computing Risks in the Public Sector

611

41. Chou, D.C.: Cloud computing risk and audit issues. Comput. Stand. Interf. 42, 137–142 (2015) 42. Bhadra, S.: Cloud computing in the risk society: its issues and hazards. Int. J. Adv. Res. Comput. Eng. Technol. (IJARCET), 8, 446–451 (2019) 43. Jangara, B.T., Bezuidenhout, H.: Addressing emerging risks in transborder cloud computing and the protection of personal information: the role of internal auditors. South. Afric. J. Account. Audit. Res. 17, 11–24 (2015) 44. Meltzer, J.P., Lovelock, P.: Regulating for a digital economy: understanding the importance of cross-border data flows in Asia. Global Economy and Development Working Paper, vol. 113, pp. 1–51 (2018)

Mitigating Security Problems in Fog Computing System Shruti1 and Shalli Rani2(B) 1 Goswami Ganesh Dutta Sanatan Dharma College, Chandigarh, India

[email protected]

2 Chitkara University Institute of Engineering and Technology, Chitkara University, Rajpura,

Punjab, India [email protected]

Abstract. Fog computing model was designed to provide computing at network’s edge among Internet of Things (IoT) devices. Its main purpose was to overcome the problems faced in cloud computing and provide services more effectively and efficiently. Being an extension of cloud computing, many security and privacy challenges that were faced by cloud computing are inherited by fog computing also. This paper presents different security and privacy challenges and work done by different researchers to overcome those challenges. Different techniques are represented with various metrices which they try to solve. We also proposed a method using ciphertext and shared key method to solve the problem of authentication and provide secure data sharing among the nodes. This method is supposed to be better than others in term of providing confidentiality, access control, authenticity and also protect from multiple splitting of private key. Keywords: Fog computing · Authenticity · Security · Attribute-based encryption

1 Introduction Increase in the number of devices and growth in the data all over the world was difficult for cloud computing to handle efficiently. Congestion issue being one of the most faced problem. To overcome this problem, a decentralized computing architecture known as “Fog computing” emerged in 2012. It was developed by Cisco to bring data processing close to data source. Fog computing is succeeding cloud computing as it provides greater computation and data storage capacity towards the edge of the network. It also reduces the latency, delay in response and bandwidth consumption for applications specially the real time ones like healthcare, traffic control, video streaming and gaming, providing high quality of service. Whereas, real time applications in cloud computing faces high latency. Fog computing reduces latency by performing processing, different functions, controlling procedure and storage close to the end user i.e., at the edge of the network. A large number of wireless heterogeneous (autonomous) devices are connected to each other in the fog. Fog can be both virtualized and non-virtualized computing model that offers © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 612–622, 2022. https://doi.org/10.1007/978-3-030-96299-9_58

Mitigating Security Problems in Fog Computing System

613

storage and networking between cloud servers and IoT devices. Time bound applications in fog are computed without involving any third-party using fog processing devices [1]. Therefore, the term Fog computing can be defined as “Fog computing is defined as a decentralized computing infrastructure located between end user devices/data source and cloud/data center that provides data. Resources are placed at such location that they increase the performance of the overall process”. The architecture of fog computing can consist of different levels depending on the system requirement. The most common one is as discussed [2]: The bottommost layer consists of IoT devices, sensors, actuators and mobile equipment. The middle one known as fog layer contains fog nodes having computing resources like servers, routers, switches and computers. The top layer is cloud layer which provides processing and storage [3]. Fog computing have many remarkable properties that makes it different from cloud computing [4] like the distance between end user and cloud increases the latency in the overall processing, it affects the time sensitive applications more. So, to overcome it, a layer known as fog layer is inserted between end user and cloud. Most of the processing takes place at this layer thus lowering the latency and providing better experience. Scalability is also an important property as providing bandwidth and computing tendency to large number of IoT devices is challenging. Increasing bandwidth in cloud computing scenario leads to increase in cost. In fog computing, fog nodes satisfy bandwidth requirement. Many businesses, entertainment and government sector services require geographical data for information and security of the users, fog computing provide location-based services using fog nodes. In fog environment, smart devices are connected to each other therefore security and privacy are most significant when services are provided by the service providers. “Trust” plays a crucial role in the concept of security and privacy. The security and privacy techniques used in cloud computing cannot be used in fog computing as there is difference in the overall architecture and functions of both [5]. The research to solve security issues and techniques is still in process. Fog computing being adopted widely with time invites security challenges day by day, so it is required to solve the issue using efficient techniques. The paper is organized as follow. Section 1 is the introduction about fog computing, why we moved from cloud to fog computing, its architecture and its different properties. In Sect. 2, security and privacy are discussed that how different factors affects both of these. Security technologies to secure fog computing are discussed in Sect. 3 and different security techniques in Sect. 3. Section 4 is about related work in fog computing security. Section 5 discusses the method proposed by the authors in the paper and Sect. 6 is the conclusion.

2 Security and Privacy Challenges in Fog Computing The security and privacy requirements of fog computing are applied keeping in mind the following challenges while executing any process. Trust: Trust in IoT computing can be explained as whatever the services are claimed by the provider are available to the user. It is believed that the requested task is fulfilled with security and privacy [6]. To provide reliable and secured services to the users in

614

Shruti and S. Rani

the fog environment, there should be some level of trust between the devices. Trust can be achieved by applying authentication, authorization and privacy to each device whether it is in LAN, distributed system, IoT or even in cloud environment. Trust in fog environment has two responsibilities to handle [5]: Firstly, the fog nodes providing services to end devices should authenticate that the service request made is genuine. Secondly, the end devices that requests or send the data should be capable enough to authenticate the intentions of the node for security purpose [7]. Authentication: Authentication is used to verify user’s identity by matching the credentials entered by the user with a database. It makes sure that only genuine users can access the fog nodes and use the services. It is the main security issue at different levels of fog node. As fog environment consists of various devices and third-party intervention is there so the chances of breach increases. Access Control. Access control is also an effective tool to ensure security. It restricts access to the system or resources. It can also be defined as a process that allows users to retrieve information from system or resources provided user has been authenticated before the access is granted. A policy-based resource access control technique is used in fog computing that helps to securely collaborate and provide interoperability between various devices connected. Deploying access control, make sure that user’s privacy is preserved and the trust between fog nodes, cloud services and end users is maintained. Intrusion Detection: In cloud architecture, intrusion detection is applied to prevent attacks like flooding attack, insider attack, port scanning and others by analyzing user login information, access control and log files. These can be deployed at both host machine and network side. In fog computing, there are some challenges faced during implementation of intrusion detection due to geographically distributed nodes, mobility and large computing environment. Privacy: Privacy leakage is one of the main concerns on internet today. Fog computing have privacy preserving algorithms running between fog layer and cloud. It is the responsibility of the fog node to collect data from sensors and end devices therefore they have to look after device privacy and communication privacy. Privacy assurance preserves data, user, usage, locations, devices and network information from unauthorized access [8–10]. Having limited resources, makes difficult for IoT devices to implement these techniques but on the other hand sensitive data need to be protected and the IoT users are concerned about it [11]. Availability: Depending on the position of fog computing gateway, if the communications are blocked by any reason, then the availability of IoT resources plays an important role. The resources should be available for the communication to continue smoothly [12] (Table 1). Attackers can violate the privacy by accessing sensitive user data. Different countermeasure can be taken into consideration to overcome these challenges like encryption of the data to be transmitted, decoy method, cryptographic technique and many such measures are proposed by different authors. Some of them we are going to discuss in related work section and a security method is also proposed by the author in proposed method section.

Mitigating Security Problems in Fog Computing System

615

Table 1. Challenges with their degree of risk on network Feature

Degree of risk on network

Authentication

Medium to significant

Access control

Medium

Privacy

Significant

Availability

Significant

Trust

Significant

3 Security Technologies In this section, security technologies of IoT to be used for securing fog computing environment are discussed [13]: A. Security Technologies for IoT Network: Fog environment communicates with different devices like fog nodes, cloud and user end devices; therefore, the communication needs to be secure among them. A node act as a single object in fog environment that does not represent the flow of the information. Therefore, a wrong information can cause problem. Although, there are some existing solutions that provides security but still there is a need for new methods and techniques. B. Security Technologies for Fog Node: An environment should be multi operating system so that fog nodes can work efficiently. In fog environment, most of the information is stored in a fog node so a wrong information will create a problem. Monitoring a node in real time requires a dynamic analysis technique but it results in performance overhead. Therefore, an operating system with low computational power cannot be used. C. Security Technologies for IoT Node: Fog nodes provide services based on the information gathered from end user devices, as these devices are close to the fog nodes any wrong information and out of order nodes will affect the overall process. End devices can be compromised by the attackers by predicting their behavior and can alter their working.

4 Related Work Different security techniques were proposed, some of them are discussed here. In [14], a fog security service method is proposed by the authors. It uses cryptographic schemes like identity-based encryption and identity-based signature to provide end on security in fog layer. FSS offers authenticity, confidentiality and non-repudiation. This scheme is executed and evaluated on OPNET simulator. The performance of FSS is found to be better than APaaS (Authentication Proxy as a Service) and legacy method. An efficient and service aware mutual authentication protocol known as SAMAFog is proposed for the fog network in [15]. Here, a user can register and use the services anonymously

616

Shruti and S. Rani Table 2. Different methods proposed by authors in Fog computing

Author

Year Proposed method

Result

Nadeem Abbas, 2019 FSS (fog security service) Muhammad Asim et al. [14]

FSS response time in different traffic loads was better than APaaS service and the legacy method

Arij Ben Amor, Mohamed 2019 SAMAFog method for fog Abid and Aref Meddeb [15] network is proposed

Low communication and computation cost, less execution time, robust, lightweight and efficient security

Chien-Ming Chen, Yanyu Huang et al. [16]

2020 An authenticated key exchange scheme for fog computing

Better performance and security. It has fewer computations and low latency and secure as compare to [24, 25]

M. A. Manazir Ahsan, Ihsan Ali et al. [17]

2019 A secure cloud storage scheme based on fog computing is proposed

A comparison between the proposed scheme and [26] shows that for data processing XOR-combination is faster than Reed-Solomon code whereas the communication cost of proposed scheme is higher than that of method proposed in [26]

Hassan Noura, Ola Salman et al. [18]

2019 A data protection scheme using dynamic key dependent approach was proposed

It encrypts, authenticate and fragment input data. Robust and maintains balance between security and system performance

Yinhao Jiang, Willy Susilo et al. [19]

2018 CP-ABE It protects from (ciphertext-policy) with KP key-delegation abuse (key-policy)-ABE is proposed (continued)

Mitigating Security Problems in Fog Computing System

617

Table 2. (continued) Author

Year Proposed method

Result

Peng Zhang, Zehong Chen et al. [20]

2016 An access control (CP-ABE) scheme in fog computing to support outsourcing capability and attribute update

It is secure under the decisional bilinear Diffie–Hellman assumption. The time of encryption and decryption are small and constant The efficiency analysis shows that cost of key generation, encryption, decryption and attribute update is less

S. Tu et al. [27]

2018 A physical layer security (PLS) to tackle impersonation attack is proposed A reinforcement learning technique i.e., Q-learning is also proposed to attain an optimal value of the test threshold in impersonation attack

The impersonation detection in proposed scheme is robust against dynamic environment

S. Alharbi, P. Rodriguez et al. [28]

2018 A fog computing-based security (FOCUS) system was proposed to protect IoT against malware cyber-attack

It protects against attacks with less response time and small bandwidth consumption

Alrawais A., Alhothaily A. et al. [29]

2017 Ciphertext policy attribute-based encryption (CP-ABE) based key exchange protocol is proposed for secure communications in fog environment

Confidentiality, authentication, verifiability, and access control is achieved by combining CP-ABE and digital signature techniques

618

Shruti and S. Rani

using virtual IDs. This method is robust against many types of attacks and is lightweight. It provides service aware authentication as compared to [21–23]. BAN-Logic method and AVISPA (Automated Validation of Internet Security Protocols and Applications) tool can be used to validate SAMAFog. Chien-Ming Chen et al. in [16], proposed an authenticated key exchange (AKE) method for fog computing. It secures the messages over an insecure network. AKE uses elliptic curve cryptography, bitwise exclusive OR and one-way hashing. Its security is analyzed using cryptographic tool ProVerif. On comparing AKE method with [24, 25], we find that its performance and security is better than other two. Another scheme known as fog-centric secure cloud storage [17] is proposed that protects data from unauthorized access, modification and destruction. It proposes different methods like XOR-combination, block management and hash function. XOR combines together with block management to preserve privacy and prevent data loss whereas hash operation detects any data modification. The security analysis of the above scheme provides privacy, data recovery and modification detection. On comparing this scheme with [26], we find that our scheme is efficient in terms of time and memory usage. In 2019 [18], a cryptographic method to secure data in fog computing is proposed. Here, dynamic key-dependent approach is combined with GF(2w ) to achieve data protection and availability. Data collected is dispersed into n encrypted fragments using dynamic key. Data recovery is done by considering any k out of n fragments along with corresponding dynamic key. This complicates the attackers as he has to compromise at least k fog nodes to obtain original data. On the other side, stored data is protected by redundant data from (n-k) fog nodes failure or unavailability (Table 2). Jiang et al. [19], introduced CP-ABE scheme that prevents key-delegation abuse issue. This scheme is further extended to traceable CP-ABE, which can trace “traitors”, who have leaked the key. In 2016, first access control scheme [20] for fog computing was proposed that supports outsourcing and attribute update. Encryption and decryption are outsourced to fog nodes, reducing the computational overhead. An efficient updating system to address attribute update is proposed that concentrates only on updates of ciphertext related to corresponding update attribute. The security of this scheme is assured by decisional bilinear Diffie-Hellman method. Thus, the scheme is efficient and save computation cost and time for encryption and decryption and update process. In [27], a Physical layer security (PLS) technique was proposed to deal with impersonation attack. In this the properties of physical layer are taken into consideration to cope up with the attack. The authentication of PLS in dynamic environment is done using Qlearning scheme. It is found that the accuracy of the receiver to detect the attack using Q-learning is 20%–60% better than fixed detection rate. To protect IoT devices from malware cyber-attack a system known as FOCUS was proposed [28]. It provides double protection or we can say two-layer security protection. FOCUS is fast in terms of response and efficient as it is implemented in fog environment close to the end user. It uses VPN to secure IoT devices communication and a challenge-response authentication is applied to protect from DDoS attacks. As compare to cloud computing, the response time of FOCUS is less and bandwidth consumption is small. Comparison of some of the techniques discussed in the related work is done to give more detailed insight into the work and services provided (Table 3).

Mitigating Security Problems in Fog Computing System

619

Table 3. Comparison of some proposed methods discussed in related work Paper

Method used Authentication

Traitor detection

[19]

CP-ABE (PKI)

One-way

[20]

CP-ABE, encryption using symmetric key

[29]

CP-ABE (PKI), encryption using symmetric key

Integrity provided

Security key updates

Delegation abuse

ID hashing Yes

No

Yes

One-way

No detection

Yes

Yes

No

Mutual

No detection

Yes

Yes

No

5 Proposed Method In the method proposed, we have used encrypted key exchange protocol based on CPABE to improve the security of the fog environment. A set of attributes is associated with every fog node. A ciphertext is assigned to these fog nodes along with an access structure well-defined over attribute set. The access structure can be designed using AND-gate. To obtain a shared key, a fog node can decrypt the ciphertext if it has specified attributes in access structure. Combination of CP-ABE technique and digital signature is used. Based on corresponding attribute set for each fog node, a private key is issued to each one. An encryption algorithm is implemented on fog nodes to generate encrypted symmetric key. Then this encrypted key is broadcasted to group of fog nodes. On receiving this key, fog nodes using their private key runs decryption algorithm to obtain the symmetric key. Thus, this way security is maintained in fog environment. The private key of the user contains all components for all attributes. A secret sharing scheme is applied to all the attributes and a bilinear map of keys and cipher components is forced for every attribute. By this, key is neither able to split nor combined with any other private key. This further reduces the chances of splitting the private key and generating the new one, protecting from being attacked by malicious user (Fig. 1).

620

Shruti and S. Rani

Fig. 1. Layout of proposed method

6 Conclusion Using the above proposed method, we will be able to achieve security goals like authenticity, access control and confidentiality. Compared to the basic certificate-based methods of security this method is more efficient and give high security. Above method will also protect the fog environment from the malicious user by avoiding the splitting of private key to make a new private key.

References 1. Naha, R.K., et al.: Fog computing: survey of trends, architectures, requirements, and research directions. IEEE Access 6, 47980–48009 (2018). https://doi.org/10.1109/ACCESS.2018.286 6491 2. Alraddady, S.I.S., Li, A., Soh, B., AlZain, M.: Dependability in fog computing: challenges and solutions. La Trobe J. Contrib. (2021). https://doi.org/10.26181/60bf1fc0bad1f 3. OFC: Openfog reference architecture for fog computing, OpenFog Consortium, Fremont, USA (2017) 4. Guan, Y., Shao, J., Wei, G., Xie, M.: Data security and privacy in fog computing. IEEE Netw. 32(5), 106–111 (2018). https://doi.org/10.1109/MNET.2018.1700250 5. Mukherjee, M., et al.: Security and privacy in fog computing: challenges. IEEE Access 5, 19293–19304 (2017). https://doi.org/10.1109/ACCESS.2017.2749422 6. Sicari, S., Rizzardi, A., Grieco, L.A., Coen-Porisini, A.: Security, privacy and trust in internet of things: the road ahead. Comput. Netw. 76, 146–164 (2015) 7. Patwary, A.: Towards secure fog computing: a survey on trust management, privacy, authentication threats and access control. Electronics 10, 1171 (2021). https://doi.org/10.3390/ele ctronics10101171 8. Aghasian, E., Garg, S., Montgomery, J.: User’s Privacy in Recommendation Systems Applying Online Social Network Data, A Survey and Taxonomy, arXiv 2018, arXiv:1806. 07629 9. Fu, A., Song, J., Li, S., Zhang, G., Zhang, Y.: A privacy-preserving group authentication protocol for machine-type communication in LTE/LTE-A networks. Secur. Commun. Netw. 9, 2002–2014 (2016)

Mitigating Security Problems in Fog Computing System

621

10. Aghasian, E., Garg, S., Montgomery, J.: An automated model to score the privacy of unstructured information—social media case. Comput Secur 92, 101778 (2020) 11. Koo, D., Shin, Y., Yun, J., Hur, J.: A hybrid deduplication for secure and efficient data outsourcing in fog computing. In: Proceedings of the 2016 IEEE International Conference on Cloud Computing Technology and Science (CloudCom), Luxembourg, pp. 285–293, 12–15 December 2016 12. Butun, I., Sari, A., Österberg, P.: Security implications of fog computing on the internet of things. In: 2019 IEEE International Conference on Consumer Electronics (ICCE), pp. 1–6 (2019). https://doi.org/10.1109/ICCE.2019.8661909 13. Lee, K., Kim, D., Ha, D., Rajput, U., Oh, H.: On security and privacy issues of fog computing supported internet of things environment. In: 6th International Conference on the Network of the Future (NOF), pp. 1–3 (2015). https://doi.org/10.1109/NOF.2015.7333287 14. Abbas, N., Asim, M., Tariq, N., Baker, T., Abbas, S.: A mechanism for securing IoT-enabled applications at the fog layer. J. Sens. Actuator Netw. 8, 16 (2019). https://doi.org/10.3390/jsa n8010016 15. Amor, A.B., Abid, M., Meddeb, A.: SAMAFog: service-aware mutual authentication fogbased protocol. In: 15th International Wireless Communications & Mobile Computing Conference (IWCMC), pp. 1049–1054 (2019). https://doi.org/10.1109/IWCMC.2019.876 6575 16. Chen, C.-M., Huang, Y., Wang, F., Kumari, S., Wu, M.-E.: A secure authenticated and key exchange scheme for fog computing. Enterp. Inf. Syst. 1–16 (2020). https://doi.org/10.1080/ 17517575.2020.1712746 17. Ahsan, M.A.M., Ali, I., Imran, M., Idris, M.Y.I., Khan, S., Khan, A.: A fog-centric secure cloud storage scheme. IEEE Trans. Sustain. Comput. https://doi.org/10.1109/TSUSC.2019. 2914954 18. Noura, H., Salman, O., Chehab, A., Couturier, R.: Preserving data security in distributed fog computing. Ad Hoc Netw. 94 (2019). https://doi.org/10.1016/j.adhoc.2019.101937, 19. Jiang, Y., Susilo, W., Mu, Y., Guo, F.: Ciphertext-policy attribute-based encryption against key-delegation abuse in fog computing. Future Gener. Comput. Syst. 78 (2017). https://doi. org/10.1016/j.future.2017.01.026 20. Zhang, P., Chen, Z., Liu, J., Liang, K., Liu, H: An efficient access control scheme with outsourcing capability and attribute update for fog computing. Future Gener. Comput. Syst. 78 (2016). https://doi.org/10.1016/j.future.2016.12.015 21. Ibrahim, M.H.: Octopus: an edge-fog mutual authentication scheme. IJ Netw. Secur. 18(6), 1089–1101 (2016) 22. Amor, A.B., Abid, M., Meddeb, A.: A privacy-preserving authentication scheme in an edgefog environment. In: IEEE/ACS 14th International Conference on Computer Systems and Applications (AICCSA), pp. 1225–1231. IEEE (2017) 23. Jia, X., He, D., Kumar, N., Choo, K.-K.: Authenticated key agreement scheme for fog-driven IoT healthcare system. Wirel. Netw. 25(8), 4737–4750 (2018). https://doi.org/10.1007/s11 276-018-1759-3 24. Wazid, M., Das, A.K., Kumar, N., Vasilakos, A.V.: Design of secure key management and user authentication scheme for fog computing services. Futur. Gener. Comput. Syst. 91, 475–492 (2019). https://doi.org/10.1016/j.future.2018.09.017 25. Jia, X., Debiao, H., Kumar, N., Choo, K.-K.R.: Authenticated key agreement scheme for fogdriven IoT healthcare system. Wirel. Netw. 25, 4737–4750 (2019). https://doi.org/10.1007/ s11276-018-1759-3 26. Wang, T., Zhou, J., Chen, X., Wang, G., Liu, A., Liu, Y.: A three-layer privacy preserving cloud storage scheme based on computational intelligence in fog computing. IEEE Trans. Emerg. Top. Comput. Intell. 2(1), 3–12 (2018)

622

Shruti and S. Rani

27. Tu, S., et al.: Security in fog computing: a novel technique to tackle an impersonation attack. IEEE Access 6, 74993–75001 (2018). https://doi.org/10.1109/ACCESS.2018.2884672 28. Alharbi, S., Rodriguez, P., Maharaja, R., Iyer, P., Bose, N., Ye, Z.: FOCUS: a fog computingbased security system for the Internet of Things. In: 15th IEEE Annual Consumer Communications & Networking Conference (CCNC), pp. 1–5. https://doi.org/10.1109/CCNC.2018. 8319238, 2018 29. Alrawais, A., Alhothaily, A., Hu, C., Xing, X., Cheng, X.: An attribute-based encryption scheme to secure fog communications. IEEE Access 5, 9131–9138 (2017). https://doi.org/ 10.1109/ACCESS.2017.2705076

A Detailed Review of Organizational Behavior of College Employees V. I. Roy(B) and K. A. Janardhanan Noorul Islam Centre for Higher Education, Thuckalay, Tamil Nadu, India [email protected]

Abstract. Education is the cornerstone of the intellectual power of a nation that forms the muscle profile of a country in the world community. The development of a country is determined by the quality of its education and it in turn depends upon the excellence of its teachers. The sharpening of the younger generation depends upon the teacher’s quality. That is the success of any institution. The value of an educator cannot rise beyond the level of any system of education and it is not an exaggeration to say it. In any educational program, the teacher is the most important element. Teachers play a vital role in the educational system. No matter how much investment is made to improve the physical and educational facilities, the education cannot be improved unless there are enough eligible teachers who can manage the educational course well so that the students can achieve the desired educational progress. Human skills are of great importance in any organization for the quality of teachers and the support of the work environment. This research utilizes the previously accumulated knowledge as a result of continuous human effort. The researcher analyzed the related literature, which introduces the arena of Organizational Behavior of oneself with current knowledge. Keywords: Education · Organizational behavior · Management

1 Overview All kinds of human, physical, financial and technological resources must be utilized economically and efficiently in the modern world. Without the optimum use of these resources no organization can achieve its objectives. In optimizing the use of these resources in a given organization like college, an organization plays a stimulus role known as ‘Management’. A person responsible for supervising and motivating employees are known as ‘managers’ (Trustee/principal in college), and the field of management is referred to the way in which they handle knowledge, skills, techniques and practices. To attain maximum productivity at the least cost, managers set an effective organization between human and non-human resources by performing their functions. The force that combines numerous resources is the management (maybe of college). The organization brings together the resources and organizes them to achieve the goal. Learning about college structural behavior is very fascinating. It is a sculpture on the portion of the college Principal to recognize, designate, forecast, and transform the © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 623–632, 2022. https://doi.org/10.1007/978-3-030-96299-9_59

624

V. I. Roy and K. A. Janardhanan

behavior of a person under him. A lot of education in the arena of organizational behavior has been undertaken in addition to enormous literature, which is useful to college principals while managing the college employees. To examine employee’s behavior, various models and research instruments are there and they are part of the college organization. The field of organizational behavior has contributed to various disciplines such as medical sciences, social psychology, anthropology, politics, sociology, economics, psychology, and social psychology. The objective is that it may help the college Principal to explore the work from the subordinates and the work so performed. If necessary, maybe appropriately amended and applied for college objectives successfully. In such formalized and non-formalized environments, the place of college OB is acknowledged and explicated ahead. After reading this the queries emerged in minds: Why do teachers, on-teachers, and students perform the mode they do? What causes different people to respond in a different way to similar conditions? Why are certain academic institutions more popular than others, even though they perform to be managed in a similar routine? All of these questions – and more – are the substance of what college organizational behavior (OB) is entirely about, is delineated in this chapter and also at appropriate places of this thesis with reference to academic college. Here Sect. 1 describes the introduction of the present work, Sect. 2 explains the details of the review of the organizational behaviors and the organizational effectiveness, the conclusion of the work is explained in Sect. 3.

2 Literature Review The researcher enables to explain the restrictions of this arena by review of the allied literature. To avoid unproductive and useless problem areas, the researcher studies the related literature. By selecting those areas the researcher finds the result as very positive findings and to the knowledge in a meaningful way would be likely to add to his endeavors. The researcher finds related literature analysis to avoid unintentional repetition. Literature evaluation of related papers for further research in their studies is listed reveals the recommendations of previous researchers. This analysis defines, encapsulates, estimates, and elucidates the literature which was prepared by the process of reading books, collecting, selecting, journals, reports, abstracts, and other reference materials. Related research identifies a review of the current research project set within a theoretical and conceptual background. 2.1 Review Based on Human Behavior Management In our day-to-day life, the behavior of people in organizations like college and the process of organizational activity is based on college and its Management. Due to the diversity and complexity of human behavior, the need for people management in any organization stands up beyond the lack of resources to fulfill human desires. The understanding of how to develop the skills of management for better living depends heavily upon the fulfillment of our quest which is applied to satisfactory use of the people in the organization. In 1776, Adam Smith argued for a novel organizational quote according to human nature [1] as the division of labor. Later, the German Sociologist Max Weber came up

A Detailed Review of Organizational Behavior of College Employees

625

with the idea of charismatic leadership and began the idea of rational organizations to control the individuals under it. In the beginning, the study of OB can dash its roots back to Max Weber and prior organizational studies. In the 1890’s the advent of scientific management is generally considered to have begun as an academic discipline, representation of Taylorism is the movement at its peak. Fredrick Winslow Taylor uses systematic goal setting, it was introduced and it motivates employees and personnel rewards were given in the organization that could be reflected as the beginning of the OB chastisement. How human/personnel factors and psychology affect OB and are analyzed with this epoch-making study that concentrates on organizational studies. The Hawthorne Effect is the study of the shift of focus of OB. The movement of Human Relations concentrated on personnel actualization, teams and the motivation of goals of the person within the organization is the segment of OB. The development of OB as the contribution made by the studies of eminent scholars such as Chester Barnard [2], Henri Fayol [3], Mary Parker Follett [4], Frederick Herzberg [5], Abraham Maslow [6], David McCellan [7] and Victor Vroom [8]. The psychology of social psychology has strongly influenced this area, which provides immense research for academic study in the 1960s and 1970s. The contribution to the study of organizational behavior was bounded rationality, contingency theory, and resource dependence, an explosion of theorizing, institution theory, bounded rationality, population ecology theories, informal organization, and resource dependence. Fredrick W. Taylor. (1856–1915) [9] in the early 1900s, started studies in time-andmotion as scientific management at the Midvale Steel Company and considered him as the father of scientific management. As an industrial engineer, he had inadequacies in work-related occupations, and he thought that by studying scientifically the certain movements that make up the overall work objective, a more rational and effective way of acting out the work could be identified. He noticed that different workers performing similar jobs in different ways while he was working as foreman in the steel industry. He was of the opinion that not every man could do his work optimally, and he sought to discover the “one best way” to do the work powerfully. His opinion became right and, in some cases, “Taylorism” results in increased productivity by about 400%. Productivity is improved over the existing level in all cases. College employees are primarily motivated by economic rewards and if offered the opportunity give a sense of direction to better their economic positions. Therefore, theory of Taylor suits to college as: • The toil scientifically appraised the optimal method to determine work in a given OB • Employees can be made more efficient by giving them hints on how to ensure their job. • Employees would be agreeable to follow these prescriptions if they make payments based on the “Differential Scale of Compensation”. In addition to arguing the scientific method, should be used to develop a task to do in the best way. The four principles of Taylor based on scientific management are tailored for use of the college OB as under:

626

V. I. Roy and K. A. Janardhanan

(i)

Learn each part of the job scientifically and improve the best way to execute college teaching and non-teaching duties. (ii) Carefully select Employees (teachers/Non-teachers) and the task performance and train them by using the developed scientific method. (iii) Collaborate totally with Employees (teachers/Non-teachers) and make sure that the right method is used by them. (iv) Effort and accountability should be divided so as to plan the college works by using scientific principles so as to make the college Employees (teachers/Non-teachers) responsible for executing their prescribed work accordingly. Many may criticize the use of Taylor’s model in college as mentioned above (i to iv) since it is dehumanizing the workplace and treating Employees (teachers/Nonteachers) like machines, but his overall modification for use of college employees will be welcomed when the teachers and non-teachers feelings are respected by providing pecuniary and non-pecuniary rewards to them for accepting the work under scientific management. Taylor as well as Frank, Lillian Gilberth and Henry L. Gantt were the scientific managers in the movement which made particularly notable contributions. The Gilbreth. [10] Other notable figures in scientific management include Frank Gilbert (1868–1924) and Lillian Moller Gilbert (1878–1972) who donated to the subject of OB. He noticed the inefficiencies in OB as Frank was always involved in training young bricklayers that were handed down from experienced workers. Motion study was proposed by him to overwhelm this situation to streamline the bricklaying process. For different types of jobs he designed scaffolding and for mortar, consistently devised precise directions. On the basis of these, the motions were reduced and Frank take 18 ½ and 4 min to involve in bricklaying. Using these Workers increased the number of bricks with no increase in physical exertion that laid per day from 1000 to 2700. Lillian Moller is the wife of Frank, she completed her doctorate in psychology along with doing projects with him. To disregard ways of reducing tasks, fatigue motions is unnecessary and to explore their interests that expanded was aimed to continue their studies. The isolation of 17 basic motions is involved in the part of their work; each known as fatherly (“Gilbreth” spelled backward, with the “t” and “h” reversed). The brigs include movements used to learn selection, position, and hold – motions. In many diligences, Gilbreth used the therblig concept to study tasks. Gilbreth used to study jobs on the concept of therblig and motion picture technology that was pioneered. In a book, the thesis of Doctoral of Lillian’s was published. The findings of psychology to the workplace are applying on one of the early works of the Psychology of Management. At the insistence of the publisher, the author L.M. Gilbreth was lilted to disguise the fact that the book was written by a woman. Work is divided into essential parts or elements on the basis of analysis. For efficient work with synthesis, only those elements that are necessary to include in the task are reconstituted. She also had a particular interest in the human implications of scientific management, arguing that developing scientific management help people to reach their maximum potential. Lillian Gilbreth was the first woman who ranks as a major contributor to the development of management to gain prominence in science. Gilbreth’s motion picture technology in studying jobs may not be applicable to the college environment since the academic culture respect human feelings. However,

A Detailed Review of Organizational Behavior of College Employees

627

one can modify Gilbreth’s model and use the same in the college libraries, laboratory, personality development class, Gymnasium, conduct of examination, security check, practical training, and so on. Henry L. Gantt (1861–1919). [11] Henry was solitary of the closest associates of Taylor, made several contributions, and develop into an independent consultant. The best recognized and still in use Gantt chart is a graphic aid to planning, scheduling, and control. For reaching the standard in the allotted time he planned to give workers extra payments like an incentive system and when workers reached standard they will be awarded bonuses. To avoid difficulties for the workers he used to encourage supervisors to Coach them. The importance of the operating functions which was, accepted by Taylor, Frank, Lillian Gilberth, and Henry Gantt, the scientific managers were not the first or only group. In 1832, British mathematician Charles Babbage with some astounding managerial insights and earlier found the advantages of division of Labour which had been carefully pointed out by Adam Smith in his book ‘Economy of Machinery and Manufacture’ that A describes transference of skill.

2.2 Review Based on the Personnel/Human Relations Movement In (1930s–1950s), the Human Relations Movement began on the way to current OB development which was the second major step. Management practices employee cooperation and ethics that can be categorized as human relationships. To “treat people as human beings (instead of machines in the productive process) the human relation approach, acknowledge their needs to fit and to feel important by listening to and regarding their complaints where possible and by involving them in certain decisions concerning working conditions and other matters, then the morale would surely improve and workers would cooperate with management in achieving good production” is stated by Raymond Mills. Elton Mayo’s popular Human Relations Movement and his famous studies of Hawthorne at the Western Electric Company’s Hawthorne Plant were conducted; our management thinking today the foundation of much of in many ways it remained [12]. Hawthorne studies before they officially started in 1923 and 1924, in a Philadelphia textile mill the department of the mule-spinning a research team headed by Elton Mayo to investigate the causes of very high turnover. The worker’s team set up a rest-pause after interviewing and consulting, which resulted in more positive worker attitudes and morale and reduced turnover. The Personnel/human Relations is the core part of college OB, when the same is maneuvered under the perspicacity of experts like Fredrick W. Taylor, The Gilbreth Henry L. Gantt, Adam Smith, Charles Babbage, Frank and Lillian. 2.3 Review Based on Experts Perceptions The organizer of the movement wants to study crowds scientifically. The leadership essence was written by the Greek philosopher Plato [13]. The topic addressed by Aristotle was persuasive communication. Niccolò Machiavelli, an Italian philosopher in the 16th century, the foundation was laid for contemporary work on organizational power and

628

V. I. Roy and K. A. Janardhanan

politics. Adam Smith in 1776 promoted the division of labor on the basis of a new form of organizational structure. Later, German sociologist Max Weber [14] initiated a discussion of charismatic leadership and wrote about rational organizations. In a little while after, Frederick Winslow Taylor [15], motivates employees by the use of goal setting and rewards that systematically was hosted. In the United States Harvard professor EltonMayo an Australian and his colleagues at Western Electric’s Hawthorne plant conducted productivity studies in the 1920s. At the peak of this movement, Max Weber traces its roots. The advantages of scientific management with organizational studies are generally considered to have begun as an academic discipline that was presented by Taylor in the 1890s. Promoters of scientific management explained the increased productivity studies with precise sets of instructions and time-motion of organization. Different compensation systems of studies were done. The focus of organizational studies after the First World War was the transformation forced by the identification of the Hawthorne Effect. In that lifted how the analysis of human factors and psychology will affect organizations. Within an organization each person focused on the Human Relations Movement’s actualization team and the motivation of the goals. In United States, Elton Mayo the Harvard Professor and his colleagues in Australia in 1920s lead a productivity study at Hawthorne plant of Western Electric Company. Quantitative research that emphasises in academic study about the social psychology that was strongly influenced in the year 1960s and 1970s. The study of Qualitative methods about anthropology, psychology and sociology became more acceptable. Karl Weick was a leading scholar. Herbert Alexander Simon and James G. March done an influential work and the so-called “Carnegie School” of organizational behavior. “Assemblages of interacting of the organizations and within the framework working towards a common goal or set of goals of structured relationships is inter related by human and non-human resources. With all aspects organizations are anxious about the behavior of individuals, how to influence the organizations and how organizations turn influence on individuals” is the view of Herbert Alexander Simon. Interdisciplinary field in organizational behavior is independent of many behavioral sciences such as anthropology, psychology, sociology, and many others. The pressing problems of management’s are the concepts of behavioral sciences and the administrative theory and practice of unique mission of organizational behavior which is to be applied. Frederick Herzberg, David McClelland, Henri Fayol and Abraham Maslow revealed their views through writings. A number of available strategies are there in approaching the problems of organizational behavior. The study of organization and management took a closed historical view. 2.4 Review Based on the Performance Evaluation and Organizational Effectiveness The existence of clear and well communicated goals is one of the essential components of a successful appraisal system by Clawson, J. G. (2009). The evaluation has little meaning without any recommended standard. Implementing an appraisal system is an important part of participation that states a great deal of evidence about it. Setting goals on people who have participated and produce guidelines which were established when they feel the changes appear less resistant inevitably in such systems. Goal accomplishment is generally associated with organizational effectiveness. To examine effectiveness, this

A Detailed Review of Organizational Behavior of College Employees

629

is an accurate method. Although when we rely completely on the goals then problems arise. Actually, in most cases, the accomplishment of one goal to another is too complex in the case of an organization. Numerous subsystems and elements are considered as the systems model for a more realistic approach. These books attempt to look at the outputs of the selected organization. Effective performance is the first and organizational effectiveness is the second specific manner which is defined. 2.5 Review Based on the Power Relations in Organizations Bakker, D. W. et al. (2005), the book initiates the review of sources of power and tells that the power is influenced from various sources. Charisma to structural or positional influences are the personal things that form the sources of range. The question of leadership style arise from the usage of power. The power of one on another is accompanied with dependency. The former is more easily influenced by the latter when one person is prejudiced by another for something he or she values. Coalition formation is discussed as the organizational goals and the political complex interrelationship. The natural systems view regardless of whether one accepts or the goals of the organization is viewed by individuals, one must explain how the desires of various interest groups and/or individuals are translated into objectives of organizations. An important issue becomes the theory of political coalition. Characteristic behavioral tone will be added to the processes evident in organizations by political processes, viewing power and goal formation. Recognition of organizational behavior is important as human beings are similar to political animals. 2.6 Review Based on the Environment, Organizations and Behavior The complex linkage between individual behavior, environmental externalities and organization has been explained in the book of Coleman, D. (1995). The technological demands of their tasks can be associated with the social relationships in work groups that explains socio technical systems. The Enthusiastic socio technical view about proponents of the enthusiasm, exists considerable controversies about natural effect of environmental factors on organizational structure and behaviour. 2.7 Review Based on the Organization and the Individual Carnegie, D., (1979), defines a group behaviour between transition and organizational design influences. In addition, the analysis established some specific goals. The importance of structure on organizational behavior has been not to overestimate or underestimate the attempt. The author made a survey on individual’s interaction views with the organization’s objectives which varied. Charles Perrot has elucidated that both the bureaucracy of Weber’s theory and contemporary structure are developing. To personality factors rather than to positions and roles the importance attached to the point of agreement in both formulations. Among the private companies in Malaysia, the effect of practice was investigated by Acton Thomas and Golden Willie Training: (2002) in human resource (HR) management. Using the survey technique he collected the data from 153 private companies of

630

V. I. Roy and K. A. Janardhanan

Malaysia on six variables. The effect of six HR variables came to a result that by using regression techniques on the basis of business performance improves the HR planning, performance appraisal, training and development, team work, compensation, employee security. They found that with the business performance four HRM practices such as: performance appraisal, HR planning, training and development, team work are correlated. Compensation and incentive for employees and the performance is a very important factor for the business organization that they ignore. The implicit personnel theory effect is investigated by Heslin Peter, Latham Gary and Walle Don (2005) in Canada, that is a combination of complex nuclear, fossil-fuel, and hydroelectric generating stations which operates in the public corporation on the basis of performance appraisal systems. Using questionnaire technique the data was collected and that was the data from 82 managers on two variables. Behavioral observation scale and implicit person theory are the two factors found by using regression technique. A motivational variable is described in implicit person theory which predicts change in employee behavior that the managers acknowledge. The relationship between management and organizational behavior is investigated by Laurie J. Mullins. The data was collected from 350 employees on two variables using questionnaire technique. The individual’s strength and weaknesses on organizational performance are the impact of variables found out by the regression technique. He found strength and weaknesses of individuals by performance appraisal and how such strength may be utilized and how to overcome weakness. From Unilever Company the data was collected from 80 employees on 9 variables using the questionnaire technique. The 9 variables effect of work was found as content, payment, promotion, recognition, working conditions, benefits, personal, leader/supervisor role, and general on the employee motivation and satisfaction level. Relationships between reward and recognition; motivation and satisfaction respectively were found by them that were statistically significant. The use of outcomes by revisiting current reward and recognition programs can be utilized by business units and can focus on addressing the needs of diverse groups of people within the business unit. The organization is the formal structure and organizing is the process to formulate the structure. Normally the organizing process consists of seven steps: specifying objectives, enumerating of activities, classifying and grouping of activities, assigning of responsibility, delegation of authority, establishing relationships, preparing of Organization chart and manuals. Labour division on scalar functional processes and specialization formal structure and span of control are key elements of classical Organization theory. The Neo-classical theory of Organization is improved over the classical theory. The theory recognizes the significant impact of the human factors on the structure of the Organization. Flat structure, decentralization, informal Organization, and some other aspects are key elements of neoclassical Organization theory. Modern theories suggest sophisticated and scientific ways to explain and design complex Organization structures. It contains (1) System approach to Organization design, and (2) Contingency approach to Organization design”. Job reengineering approach, job enlargement approach, job rotation approach, job characteristics approach, job enrichment approach, socio-technical job design and QWL approach, HPWPs, are main approaches for job designing.

A Detailed Review of Organizational Behavior of College Employees

631

2.8 Review Based on the Health-Related Quality of Life Among College Teachers The influences of work-related factors like role overload, job rank, working hours, role boundary and number of SCI papers, chronic disease on college teachers’ QOL, and demographic variables like age, gender, education, and marital status is investigated by Ge et al. A negative meaningful relation was observed among weight and activity level as well as teacher’s life quality and physical activity level. The teachers’ QOL became more impaired when they experienced more stress from organizational climate, staff relations factors and work overload were shown in the early findings. In general, the perspective of individuals has been confined to the influencing factors of teachers’ QOL, whereas the organizational level of factors was rarely taken into concern in previous research. Conceptual independence has four positive constructs, but there is a common, underlying link that runs between them and ties them together, that is, to accomplish tasks and goals. To a motivational propensity, each of the facets contributes the shared mechanism. Those who are resilient to adversity will be high in self-efficiency and those who are confident on a specific task will be having more hope and will bounce back quickly after temporary hopelessness. Lead to positive organizational behavior by a psychological state which is stated as PsyCap, reported as a positive resource for fighting occupational stress and turnover, depression, and burnout. Conducted studies on health-related quality of life (HRQOL) of college teachers to explore the effect of organizational behavior variables. According to the above-reviewed literature, the QOL of teachers is enhanced by variables which are the potential resources. This study helps to investigate relationships between the teachers’ HRQOL and organizational behavior variables such as PsyCap, group identification, organizational justice, POS, and psychological empowerment. For improving HRQOL of teachers, whether they are positive resources, that could provide how colleges can improve their teacher’s health by practical ideas and thus improves organizational effectiveness.

3 Conclusions Several things have been acknowledged through this literature survey on OB. In organizations, human behavior of different origins can be traced as the first interest. Perhaps even to Biblical times, the interest goes back to that we have a prominent variety of contributions. Moreover, contributions are made by moral philosophy and politics along with engineering, economics, sociology, and psychology. The management of an organization that relates the practical matters of systematic administration as a product of the twentieth century is properly viewed. In the 1920s and 1930s, the mechanistic assessments of Taylor, Weber, and Fayol derived that Mayo and others fulfilled humanistic interests. The existing study is limited to an only specific organization. In recent years no scholar has made a special issue for investigating Organizational Behavior Management in engineering colleges. To overcome this, the research helps to the fulfillment of social and legal responsibilities and to make an effective contribution to the execution of the academic aims of the college. The particular focus of the thesis is to ensure that how far the college management serves the OB related employee’s objectives; how far the performance of the personnel function is judged; on the extent to which these objectives

632

V. I. Roy and K. A. Janardhanan

are met and on the efficiency with which advises, services and guidance are provided by the college management.

References 1. Abad, A.: Management and Organization Development. Rachna Prakashan, New Delhi (1972) 2. Arnold, H.J., Feidman, D.C.: Organizational Behavior. McGraw Hill International, New York (2017) 3. White, A., Phillip, B.: Organizational Behavior. Prentice Hall, Englewood Cliff (1965) 4. Keith, D., Scott William, G.: Human Relations and Organizational Behavior: Readings and Comments. McGraw Hill, New York (1969) 5. Lufthansa, F.: Organizational Behavior, 7th edn. McGraw-Hill, New York (1995) 6. New Storm, J.W., Davis, K.: Organizational - Human Behavior at Work, 9th edn. McGraw Hill, New York (1989) 7. Whyte, W.F.: Organizational Behavior. Irwin, Dorsey Homewood III (1969) 8. Woodward, J. (ed.): Industrial Organizations: Behavior and Control. Oxford University Press, Oxford (1970) 9. Aquinas, P.G.: Organizational Behavior. Excel Books, New Delhi (2008) 10. Fayol, H.: General and Industrial Administration. Sir Isaac Pitman & Sons Ltd., London (1949) 11. Robbins, S.P., Coulte, M.: Management, 10th edn. Prentice Hall, Upper Saddle River (2003) 12. Schermerhorn, J.: Management, 8th edn. Wiley, Hoboken (2004) 13. Tannenbaum, R., Schmidt, W.H.: How to choose research methods. Harv. Bus. Rev. 51, 162–180 (2007) 14. Haimann, T.: Professional Management. Eurasia Publishing House, New Delhi (1976) 15. Davis, K.: Research on Human Behavior of Work. Tata McGraw-Hill, New Delhi (1975)

Non-invasive Flexible Electromagnetic Sensor for Potassium Level Monitoring in Sweat Gianvito Mevoli1(B) , Claudio Maria Lamacchia2 , and Luciano Mescia1 1 2

Politecnico di Bari, via E. Orabona 4, 70125 Bari, Italy {gianvito.mevoli,luciano.mescia}@poliba.it IAMAtek srl, via Nicholas Green 13/A, 70127 Bari, Italy [email protected]

Abstract. In this paper, a novel wearable and non-invasive potassium sensors based on a flexible electromagnetic filter is presented. The filter topology includes a tapered feed line loaded four supershaped double split ring resonators with gap parallel to the line. The Kapton dielectric substrates was taken into account for its flexibility, chemical and high temperature resistance, high threshold for wear and abrasion. The embedded resonators have a reduced footprint and they was designed in order to exhibit enhanced sensitivity close to the operating frequency of 3 GHz. Moreover, a parametric analysis was carried out with the aim to identify the shape and geometrical parameters ensuring the optimal sensing capabilities. A prototype of the designed flexible sensor was realized and preliminary laboratory tests was carried out by considering a collagen sensing layer. The obtained results highlight a lowering of the resonance peak as the potassium concentration increases.

1

Introduction

During the last decade, wireless technologies have assumed importance in the modern world. In particular, their rapid expansion has covered various engineering fields ranging from mobile communications to sensors, home automation, aerospace and defense. Moreover, they have also promoted the development of the mobile health (m-health), a new model of health care based on both electronic health and smartphone technology [1]. In particular, the use of mobile communication devices and technologies makes possible the real-time and continuous monitoring of the patient health through a network constituted by sensors that are able to detect anomalies of the biological activities and health status in a non-invasive way. Several biological fluids can be studied to monitor the patient health. In particular, for the monitoring of the blood glucose levels the related sensors are generally invasive. Currently, the most used method in this context is the “finger-pricking” method based on testing strips and a glucometer. It is a non-continuous monitoring method over time and it causes pain to the patient. Another fluid of interest is the interstitial fluid that offer information about metabolic dysfunctions, functional organ failure and efficacy of c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 633–642, 2022. https://doi.org/10.1007/978-3-030-96299-9_60

634

G. Mevoli et al.

the administered drugs. Recently, microneedles arrays are used allowing optimal penetration for the sampling of the interstitial fluid, minimizing any damage to the capillaries and nerve endings present in the dermis, avoiding the contamination of sweat. Saliva is another complex fluid containing many analytes that diffuse through the blood and some research groups are investigating the potential use of mouthguards as a platform for the continuous, minimally invasive sensing of metabolites in saliva. On the other hand, sweat is the most accessible body fluid in nature and plays a biological role of primary importance for thermoregulation. Although the detection of sweat for diagnostic purposes is very promising, some challenging still persist such as the limited knowledge of the other fluids, the sampling process, the production of sweat is variable, and the epidermis contamination. 1.1

Background

The new taxonomy of the m-health reflects four sensors categories: continuous monitoring sensors; diagnosis sensors; prognosis sensors and medical assistance sensors. Each of these categories can have its own subcategory, as illustrated in Fig. 1. The sub-classification of these main categories is not exhaustive since the development of sensing for m-health applications is still ongoing and constantly updated. However, common characteristic is their ability to communicate and transmit information from the human body to another system or device that is able to implement a particular action in response to a given stimulus. One of the main issue to take into account when designing sensors concerns their daily use on the human body, especially under less favorable conditions than the highly controlled ones, typical of a clinical infrastructure. To address this problem, the concept of a multisensor has recently been developed. It consists in the integration of sensors of different nature within the same device, allowing the measurement of endogenous (glucose, blood perfusion, sweating, movement, etc.) and exogenous (temperature, humidity, etc.) factors [3–5]. In fact, in many wearable systems prototypes, the sensor and signal processing subsystems are integrated into a single wearable object, such as T-shirts, gloves, bracelets and patches. The main properties of these devices are their small size, resistance to humidity and shocks, flexibility, lightness, low power dissipation, wireless connectivity and a long life cycle [6,7]. To fabricate these smart sensors, with certain operative characteristics, various studies have been performed using hybrid structures [8], hybrid materials [9,10], multidimensional carbon nanofibers and nanomaterialbased electrodes [11]. The wearable sensor is the most sophisticated form of “deformable” sensors as it involves the development and implementation of multiple technologies ranging from materials science to electronic and mechanical engineering. However, the network of intelligent health monitoring sensors can be very useful to observe some changes in the patient health so that the specialist can perform a timely diagnosis or alerting him in the right timing during a critical condition. In the long term, the patient will be able to benefit from this

Non-invasive Flexible Electromagnetic Sensor

635

Fig. 1. Taxonomy of m-health sensors [2].

constant and continuous monitoring, since persistent observation allows optimal treatment about chronic diseases as well as to control the rehabilitation period after a trauma or after an intervention surgical. Hence, wearable devices have increased the efficiency of health monitoring [12]. The wearable technologies currently available on the market are the following: • Electrocardiogram (ECG): it measures the electrical activity of the heart (wearable patch or chest pads); • Electroencephalogram (EEG): it measures the electrical activity of the brain (wearable scalp); • Heartbeat: it measures the heart cycle rate (Pulse oximeter, wearable shirt); • Heart tone: it records the heart sound (Phonocardiograph or wearable shirt); • Body temperature: it measures the body temperature (Wearable patch); • Blood pressure: it measure the blood pressure (Blood monitoring cuffs); • Blood glucose: it measures the glucose level (wearable patches or glucose monitoring devices); • Respiration rate: it measures the respiratory rate in the unit of time (piezoelectric or piezoresistive wearable sensors); As part of the basic and industrial research activities relating to the study and design of a new generation of wearable biosensors, electromagnetic based approaches have been recently proposed as an alternative solution to provide continuous and non-invasive monitoring of the blood constituents [13], vital signs [14] and sweat electrolytes [15]. Taking into account that body fluids with different electrolytic concentrations can result in different losses and hence absorption

636

G. Mevoli et al.

of electromagnetic waves, in this paper we report a flexible microwave-based sensor for non-invasive potassium levels monitoring in sweat. The sensor is composed of an ultra-light weight conformal filter printed on a Kapton substrate and operating near 3 GHz. The proposed filter is based on a set of double supershaped split ring resonators loaded tapered feed line. Moreover, the sensor structure can be placed directly on the skin and has an overall height of 135 µm.

2

Proposed Design

Biological tissues are complex dielectric structures whose dielectric properties result from the interaction of the electromagnetic energy with the tissue constituents at the cell and molecular level. From the macroscopic point of view, such interaction is given by the different dynamic processes involving reorientation of bipolar molecules, interfacial polarization, ionic diffusion (with respect to ions of different signs of charges), conductivity of surface cell structures, and motion of the molecules. In fact, the cell membrane is semipermeable to ions and some of these are able to cross it without difficulty, while some will be rejected. Furthermore, the intra and extra-cellular membranes are composed of electrolytes characterized by resistive properties. At low frequency, the DC current flows unaltered and due to the high capacity of the cell membranes it does not affect the electrical behavior of the cell. On the other hand, at higher frequency the currents penetrate through the cell by polarization, charging and discharging the cell membranes. Therefore, by varying the frequency it is possible to identify four distinct regions: α for low frequencies, β for radio frequencies, γ and δ for microwaves. The dispersion α is associated with the polarization connected to the double layer of charge and with the ion conduction effects of the electrolytes. The dispersion β is mainly due to the Maxwell–Wagner effect. Cell suspensions such as blood, typically, exhibit significant β dispersion in the radio frequency range between 100 kHz and 10 MHz. Furthermore, the orientation of free water molecules causes dispersion γ; the water bound to proteins and the internal motion of these proteins generates an additional process which is precisely the dispersion δ observable in the frequency range between the dispersion β and γ. Dielectric dispersion is the corresponding frequency dependence of permittivity which typically display multi-relaxation characteristics which can be modeled by using the following Cole–Cole based relationship [16,17] ε(ω) = ε∞ +

N  k=1

Δεk 1 − (jωτk )

1−αk

−j

σ ωε0

(1)

where τk , Δεk and αk denote the relaxation time, strength and distribution parameter relevant to the k-th relaxation process, respectively, N is the number of dielectric relaxation processes, ω is the angular frequency, σ is the static conductivity, ε0 is the permittivity of free space, ε∞ is the relative permittivity at infinite frequency. However, a proper selection of N , ε∞ , Δεk , αk , τk , σ

Non-invasive Flexible Electromagnetic Sensor

637

allows an accurate modeling of the measured permittivity spectrum of a wide variety of biological media and mixtures. On the basis of these general characteristics, we conducted a preliminary analysis aimed at evaluating the dielectric properties of saline mixture loaded with potassium. To this aim, the experimental data reported in [18] was considered. The obtained results were used to identify the operating frequency around 3 GHz ensuring the most significant change of the dielectric properties due to the variation of the biomarker concentration. Furthermore, following a number of technological considerations, the DSRR (Double Split Ring Resonator) technology was chosen to realize the sensing element. After a detailed study on the filter topology and DSRR geometries, an in-depth analysis about the dielectric materials to be used as substrates was carried out. In particular, specific requirements as flexibility, lightness, biocompatibility and small size was considered. Due to the its excellent physical and electrical properties along a wide temperature spectrum as well as the demonstrated biocompatibility and non-cytotoxicity, the selected flexible material was the Kapton. Figure 2 shows the final layout resulting from various optimization steps concerning the geometric structure of the rings, the distance of the rings from the feeding line, the gaps of each ring and the type of feeding. Moreover, the coplanar printed circuit board technology was implemented to facilitate the mounting of the connector on the device. The side conductors are separated from the central track by a small gap, which has a constant width along the length of the line. The size of the central conductive trace, the gap between conductors together with the thickness and electrical permittivity of the dielectric substrate are crucial to define the effective dielectric constant, the characteristic impedance and the losses. Moreover, to make possible the integration of the feeding connector on the top layer with a reduced reflection coefficient, a tapered feeding line has been designed, the detail of which is shown in the right part of Fig. 2. Finally, in order to improve the quality factor of the resonator and its frequency response, the rings geometry was optimized and defined by referring to the two-dimensional version of the superformula [19]. Once the sensor structure was defined, its sensing characteristics was investigated by performing a parametric analysis aimed to quantify how the variation of the physical and geometrical parameters as well as the potassium concentration affect the device performance. In particular, different substrate (FR4, PET, KAPTON) and superstrate (PDMS, collagen) materials as well as the rings geometry, the number of rings, the feeding line geometry, potassium concentration, rings gap size was investigated. Figure 3 shows some of the most significant simulation results regarding the frequency response of the sensor versus different potassium concentration. The obtained S11 scattering parameter clearly highlight an impedance matching degradation close to the 3 GHz as the concentration increases. Furthermore, the corresponding S21 parameter reaches its maximum value of about −5 dB for a pure saline solution.

638

G. Mevoli et al.

Fig. 2. Supershaped DSRR in coplanar configuration. (left) top layer, (right) detail of the feeding line including the connector.

Fig. 3. Scattering parameters S11 (left) and S21 (right) versus the frequency for different potassium concentration.

3

Sensor Prototypes and Experiments

The design and optimization of the microwave sensor has resulted in the definition of the physical and geometric parameters characterizing the executive project used for the prototyping. In Table 1 are summarized the corresponding numerical values. For the prototype execution, the CAD project used to perform the electromagnetic simulations was reworked in such a way as to generate particular files (gerber files) that are easier to use for the production of flexible printed circuits. In this context, special post-processing measures have been implemented aimed at i) fine control of tolerances; ii) removal and or modification of

Non-invasive Flexible Electromagnetic Sensor

639

Table 1. Geometric parameters characterizing the electromagnetic sensing system based on supershaped DSSR. Parameter Value (mm) Description Ws

25

Ground/substrate width

ls

40

Ground/substrate length

h

0.1

Substrate thickness

t

0.035

Metal thickness (copper)

g

0.3

Gap ring width

w

0.6

Feed line width

sc

0.2

Distance between ring and power line

sr

0.85

Distance between external and interior rings

Wr

0.3

Metal ring width

a1

8.45

y-axis side length

a2

8.45

x-axis side length

geometric details that could create problems during the production phase; iii) exact definition of the position of the via holes and pads for the connectors soldering; iv) define the references for the perfect alignment of the circuits in the sheet as well as the packing of different circuits for the optimal covering of the standard sheet. Figure 4 illustrates the final prototype. Once the microwave sensor samples were acquired, a number of functional tests were carried out for the experimental verification of performance. First of all, an accurate analysis of samples was performed in order to verify the required tolerances satisfaction. To this aim, an optical microscope was used. The result was perfect compliance with the requirements. Flexibility tests were then performed by subjecting the circuit to different mechanical stresses. Also in this case the result was in

Fig. 4. Electromagnetic sensor prototype.

640

G. Mevoli et al.

conformity with what was requested, indeed the circuit also resisted to quite critical stresses. Finally, preliminary tests were carried out to verify the degree of adhesion and anchoring of the power connectors as well as the welds to anchor the copper microstrips. The outcome of the tests was in accordance with what was expected and defined in the prototyping phase. The electromagnetic characterization was carried out using a vector network analyzer, the electronic calibration kit, the phase compensation cables connecting the VNA and patch cables allowing the transition from the U.FL-R-SMT connector mounted on the circuit board and the SMA connector of the VNA cables. Furthermore, in order to guarantee the repeatability of the results and to reduce all the possible errors related to the deformation and uncontrolled bending of the cables and the sensor, the circuit was anchored, by means of adhesive tape, on a flat work bench. In order to analyze the performance of the circuit in a wide frequency band, the VNA calibration was done in the frequency range from 1 GHz to 5.5 GHz. In accordance with what was defined in the design phase, the designed sensor exhibited a narrow band-pass filter behavior. However, despite the good impedance matching (S11 ≈ −15dB), the modulus of the transmission coefficient has a value of about −9 dB) which is lower than that obtained by electromagnetic simulation. This discrepancy is probably caused by the imperfect planarity of the prototype and by the effect of the work surface on which it is directly supported. In fact, since the circuit is in coplanar technology and considering the reduced thickness of the kapton substrate, the electromagnetic field is likely to interact also with the wood of which the table is made. Finally, the other resonances at higher and lower frequencies of the operating frequency band defined in the design phase are mainly due to the connecting cables between the sensor and the VNA cables. In order to test the sensor performance to the variation of the potassium concentration, a functionalized superstrate layers were applied. In fact it was expected that the application this layer should reduce the transmission coefficient. Thanks to its proprieties as lightness, flexibility, adherence, ease of use and biocompatibility, a specific collagen superstrate was designed. In particular, a series of laboratory tests were performed comparing the configuration with and without functionalized layer. Figure 5 illustrates the sensor prototype

Fig. 5. Prorotype of the electromagnetic sensor with a functionalized layer of collagen.

Non-invasive Flexible Electromagnetic Sensor

641

equipped with functionalized layer of collagen placed on the top of the ring resonator circuit. Some preliminary experiments was carried out by considering a sample without collagen layer, a sample with collagen layer immersed in pure water, a sample with collagen layer immersed in an aqueous solution containing a concentration of 1000 ppm of potassium ions. Some of the most significant results are shown in Fig. 6. It can be observed that the presence of potassium ions produces a reduction of the transmission coefficient in the frequency range from 3 GHz to 3.5 GHz, a trend that is in accordance with the numerical results.

Fig. 6. Scattering parameters S11 (left) and S21 (right) of the sensor with and without the collagen layer.

4

Conclusion

In this paper, a novel wearable electromagnetic sensor for monitoring the human health is illustrated. The designed sensor is able to monitor the potassium level in sweat in a continuous and non-invasive way. Once the design, optimization and realization of the sensor were made, the device was tested both with and without a collagen layer. The obtained results show an optimal functional performance at the operating frequency in terms of reflection and transmission coefficients. Acknowledgment. This work has been partially supported within the Bando INNONETWORK-sostegno alle attivit` a di R&S per lo sviluppo di nuove tecnologie sostenibili, di nuovi prodotti e servizi plan: T-CARE- Teleassistenza e monitoraggio innovativi dei parametri vitali a domicilio con biosensori indossabili, Project code: B37H17005210007.

References 1. Istepanian, R.S.H., Woodward, B.: m-HEALTH: Fundamentals and Applications. Wiley, Hoboken (2017) 2. Hoffmann, K.P., Solzbacher, F.: Recording and processing of biosignals. In: Kramme, R., Hoffmann, K.P., Pozos, R.S. (eds.) Springer Handbook of Medical Technology, pp. 923–946. Springer Handbooks. Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-540-74658-4 46

642

G. Mevoli et al.

3. Abell´ an-Llobregat, A., et al.: A stretchable and screen-printed electrochemical sensor for glucose determination in human perspiration. Biosens. Bioelectron. 91, 885–891 (2017) 4. Corrie, S.J., et al.: Blood, sweat, and tears: developing clinically relevant protein biosensors for integrated body fluid analysis. Analyst 140, 4350–4356 (2015) 5. Liu, C., et al.: A glucose oxidase-coupled DNAzyme sensor for glucose detection in tears and saliva. Biosens. Bioelectron. 70, 455–461 (2015) 6. Yilmaz, T., Munoz, M., Foster, R.N., Hao, Y.: Wearable wireless sensors for healthcare applications. In: International Workshop on Antenna Technology (iWAT), pp. 376–379 (2013) 7. Coviello, G., Avitabile, G., Florio, A.: The importance of data synchronization in multiboard acquisition Systems. In: 2020 IEEE 20th Mediterranean Electrotechnical Conference (MELECON), pp. 293–297 (2020) 8. An, B.W., et al.: Stretchable and transparent electrodes using hybrid structures of graphene-metal nanotrough networks with high performances and ultimate uniformity. Nano Lett. 14, 6322–6328 (2014) 9. Ji, S., et al.: High dielectric performances of flexible and transparent cellulose hybrid films controlled by multidimensional metal nanostructures. Adv. Mater. 29, 1700538 (2017) 10. Jung, E.D., et al.: Highly efficient flexible optoelectronic devices using metal nanowire-conducting polymer composite transparent electrode. Electron. Mater. Lett. 11, 906–914 (2015) 11. Kim, K., et al.: Nanomaterial-based stretchable and transparent electrodes. J. Inf. Disp. 17, 131–141 (2016) 12. Panteloupoulos, A., Bourbakis, N.F.: A survey on wearable sensor-based systems for health monitoring and prognosis. IEEE Trans. Syst. Man Cybern. Part C Appl. Rev. 40, 1–12 (2010) 13. Hanna, J., et al.: A slot antenna for non-invasive detection of blood constituents concentration. In: 2019 IEEE International Symposium on Antennas and Propagation and USNC-URSI Radio Science Meeting, Atlanta, GA, USA, pp. 1003–1004 (2019) 14. Rahman, M., et al.: Resonator based switching technique between Ultra Wide Band (UWB) and single/dual continuously tunable-notch behaviors in UWB radar for wireless vital signs monitoring. Sensors 18, 3330 (2018) 15. Eldamak, A.R., Fear, E.C.: Conformal and disposable antenna-based sensor for non-invasive sweat monitoring. Sensors 18, 4088 (2018) 16. Caratelli, D., et al.: Fractional-calculus-based FDTD algorithm for ultrawideband electromagnetic characterization of arbitrary dispersive dielectric materials. IEEE Trans. Antennas Propag. 64, 3533–3544 (2016) 17. Mescia, L., Bia, P., Caratelli, D.: Fractional-calculus-based electromagnetic tool to study pulse propagation in arbitrary dispersive dielectrics. Physica Status Solidi (A) 216, 1800557 (2019) 18. Jensen, P.D., et al.: Cole-cole parameter characterization of urea and potassium for improving dialysis treatment assessment. IEEE Antennas Wirel. Propag. Lett. 11, 1598–1601 (2012) 19. Mescia, L., et al.: Modeling of electroporation induced by pulsed electric fields in irregularly shaped cells. IEEE Trans. Biomed. Eng. 65, 414–423 (2018)

A Study on Sequential Transactions Using Smart Card Based Cloud Voting System Roneeta Purkayastha(B) and Abhishek Roy Department of Computer Science and Engineering, Adamas University, Kolkata, India

Abstract. People’s mandate is an integral part of any good governance for smooth and unbiased administration of state and its policy making. The unbiased representation of people’s mandate is the stepping stone of this process. However, due to several real world anomalies including the threat perspective and manipulation of voters by the non-state actors, nowadays it is difficult to collect and maintain the actual mandate of the people through the conventional approach of voting. To resolve this issue, advanced technologies like Cloud Computing, Edge Computing, etc. may be applied to upgrade the conventional voting system into the technological form, which mainly relies on virtual medium for transmission of sensitive information between the voter and the state sponsored agents like Election Commission, Presiding Officer, Polling Party, etc. The application of advanced technologies can reduce the threat perspective and manipulation of voters by nonstate actors to a significant level mainly due to electronic message communication between the state sponsored agents and voter, however, this electronic message communication should be monitored properly from its initial state so as to prevent the illicit attempts of adversaries over the electronic transaction. Since this electronic voting is conducted through a virtual medium like the Internet, there are chances of latency problems while reflecting the exact record in a real time situation. To resolve these issues and provide an integrated voting system to the voter, authors have proposed a Cloud Voting System (CVS) using a multipurpose smart card interface namely Multipurpose Electronic Card (MEC), which also helps to uniquely identify the voter during the voting process. As multiple stakeholders have to engage among themselves during the real world implementation of the proposed Cloud Voting System (CVS) using Multipurpose Electronic Card (MEC), it is essential to maintain the proper sequence of operations among its stakeholders. To achieve this objective, in this paper authors have explained the proper sequence of operations between the various stakeholders of Cloud Voting System (CVS) using the Sequence Diagram of Object Oriented Modeling (OOM). Keywords: Cloud voting · Actors · Sequence diagram

1 Introduction For smooth and unbiased state administration and policy making, people’s mandate forms an integral component of any good governance. It becomes difficult to collect and maintain people’s mandate through conventional voting approach due to several real © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 643–653, 2022. https://doi.org/10.1007/978-3-030-96299-9_61

644

R. Purkayastha and A. Roy

world anomalies. The application of advanced technologies such as Cloud Computing and Edge Computing can reduce the threat perspective and manipulation of voters by non-state actors to a significant level due to electronic message communication between the state sponsored agents and voter, however, this electronic message communication should be monitored properly from its initial state so as to prevent the illicit attempts of adversaries over the electronic transaction. Since this electronic voting is conducted through a virtual medium like the Internet, there are chances of latency problems while reflecting the exact record in a real time situation. To resolve these issues and provide an integrated voting system to the voter, authors have proposed a Cloud Voting System (CVS) [8] using a multipurpose smart card interface namely Multipurpose Electronic Card (MEC) [2], which also helps to uniquely identify the voter during the voting process. The identification process of voters can give rise to either of two outcomes. In case of successful verification, voters are allowed to cast their vote. In case of unsuccessful verification, the voter is not allowed to cast their vote and the voting process for that specific voter is aborted through system time out. To enhance the security feature of the voting process, biometric verification of voters should be included within the electronic voting process for secure pairing of voters with a terminal device used for casting the vote. Thus a multi level security can be achieved to ensure the Privacy, Integrity, Non-Repudiation and Authentication (PINA) of the electronic voting system. As this electronic voting system includes the direct and indirect engagement of multiple stakeholders like voters, state sponsored agents like Election Commission, Presiding Officer, Polling Party, etc., it becomes essential to maintain the proper sequence of operations between them, else the entire process will function out of sequence and will abort abnormally. As the mandate of the voter should be reflected in an unaltered manner, the identity of the voter should be verified strictly. The Multipurpose Electronic Card (MEC) will act as an electronic interface between the voter and the state and state sponsored agents for casting of vote. Moreover, this smart card will also help to uniquely identify the voter using an unique identification number whose counterpart should be available with the state well in advance right from the registration of voter in the voter list. Furthermore, this smart card will facilitate the voter to transmit its biometric parameters to the state to pass through the multiple levels of security protocol of our proposed Cloud Voting System (CVS). Hence, the voter can cast voting staying within its comfort and safe zone thereby avoiding the threats of non-state actors during the voting process. Moreover, the application of this Cloud Voting System (CVS) [8] will help the administration to arrange, monitor and manage the entire voting process in a centralized manner thereby engaging less human resources and limited resources for successful completion of the entire voting process at a faster pace even for the distant locations. As the entire process will continue in a phase wise manner where the output of the previous phase will be accessed as the input of the next phase of operations, it becomes essential to maintain the proper sequence of operations between the stakeholders. To achieve this objective, in this paper authors have explained the proper sequence of operations between the various stakeholders of Cloud Voting System (CVS) using the Sequence Diagram of Object Oriented Modeling (OOM).

A Study on Sequential Transactions Using Smart Card

645

Section 2 states the origin of work. Section 3 describes our Cloud voting system (CVS). Section 4 explains the sequential transactions of our proposed Cloud Voting System (CVS) using Sequence Diagrams of Object Oriented Modeling (OOM). Section 5 states the conclusion of our research work and also explores its future scope.

2 Origin of Work Electronic Governance (E-Governance) [1, 2, 4] enables the delivery process of electronic services to be faster, cost-effective and transparent. It reduces the cost with respect to every context such as human resource, logistics support, infrastructure, etc. The EGovernance system is based on a citizen-centric multi-faceted smart card namely Multipurpose Electronic Card (MEC) [1, 2, 4] for the one-to-one interaction between Citizen and Government. In this system, the Citizen initiates the transaction with the help of Multipurpose Electronic Card (MEC). The Citizen attempts to login to the E-Governance interface. The citizen then undergoes the verification process which can have either of the two outcomes. If the citizen is a valid one, the SERVICE REQUEST is forwarded to the E-Service Server. The E-Service Server forwards it to the appropriate service-specific server after checking the transaction parameters. If the citizen is invalid, the transaction is aborted and the system is timed-out. In this E-Governance system, various types of electronic services provided to the Citizen are banking services, employment services, health care services, educational services etc. with the help of Multipurpose Electronic Card (MEC). As an extension to this E-Governance system, the authors have proposed the Electronic Voting system [6]. It has a biometric authentication based system which passes through two stage security encompassing Multipurpose Electronic Card-based unique identification number verification and biometric (finger-print) authentication. The EGovernance interface receives the SERVICE REQUEST sent by Citizen and it passes through two-stage security authentication which may lead towards either of following two situations: 1. Situation – 1 (i.e. Success): In case of successful verification, SERVICE REQUEST of citizen proceeds further for completion of desired electronic transaction. 2. Situation – 2 (i.e. Failure): In case of unsuccessful verification, SERVICE REQUEST of citizen is aborted. In case of successful verification, the SERVICE REQUEST is forwarded to Electronic Service (E-Service) Server. Electronic Service server is the repository of multivariate electronic services which are provided to the citizen on case to case basis. The electronic service requested by citizen is matched with the Electronic Service server and the electronic transaction proceeds further to complete the desired operation and send SERVICE RESPONSE to the citizen.

646

R. Purkayastha and A. Roy

To deliver more services to citizen in a secured manner, this Electronic Governance system is enhanced to Cloud Governance system [3, 5, 7]. This cloud based electronic service delivery model will deliver increased number of services like cloud voting, cloud banking [3], cloud healthcare [7], cloud transportation [10], etc. to the citizen through an integrated environment using Multipurpose Electronic Card (MEC). To access these cloud based electronic services, citizen has to pass through two step biometric authentication system which may lead towards any of the following two situations: 1. Situation – 1 (i.e. Success): In case of successful validation, SERVICE REQUEST proceeds towards completion of desired operation. 2. Situation – 2 (i.e. Failure): In case of unsuccessful validation, SERVICE REQUEST is aborted. The private cloud attached to this Cloud Service (i.e. C-Service) server stores and manages the exact type of service requested by citizen like cloud voting, cloud banking [3], cloud education [1], cloud healthcare [7], etc. In case of cloud banking, the SERVICE REQUEST is forwarded to Bank server, which routes the request to specific bank server through hybrid cloud. In this paper, authors have proposed a Cloud Voting System (CVS) using a single window interface namely Multipurpose Electronic Card (MEC), which facilitates the voter (i.e. citizen) to cast his/her vote from their comfort and safe zone thereby avoiding any unforeseen incident which may be caused by any non-state actors.

3 Proposed Cloud Voting System The proposed Cloud Voting System (CVS) [8, 9] is shown using Block diagram [9], which is elaborated using Sequence Diagrams in the subsequent section. As shown in Fig. 1 [9] the steps of proposed Cloud Voting System (CVS) are mentioned below: Step 1: Voter attempts to login to the Cloud Voting System (CVS) with the help of smart card namely, Multipurpose Electronic Card (MEC) and biometric authentication (i.e. fingerprint scanner) paired with any terminal device like mobile phone or personal laptop through cloud medium. Step 2: The SERVICE REQUEST of the voter is forwarded to Government and Government sponsored agent for verification of voter. Step 3: The SERVICE REQUEST is then sent to Election Commission through private cloud. Since the private cloud is capable of storing sensitive information, it is protected from malicious security attacks using Standard Operating Procedure (SOP). Step 4: This verification of voter may lead towards any one of the following two situations: Situation 1: In case of valid voter, SERVICE REQUEST is forwarded to the next step of voting process. Situation 2: In case of invalid voter, SERVICE REQUEST is aborted. Step 5: In case of successful verification of voter, Cloud Edge server is used to record the output of verification process.

A Study on Sequential Transactions Using Smart Card

647

Step 6: The details of contesting political party candidate, political party affiliation etc. are displayed to the valid voter for casting of vote. Step 7: After completion of voting process, an acknowledgement slip is provided to the voter. Furthermore, the numbers of polled votes are stored in the Data Center. Step 8: Cloud Edge Server updates its record for each voter. The intermediate results are stored in the Data Center periodically through READ and WRITE operations. Step 9: The final result is retrieved by the Election Commission from the secured Data Center and the accurate voting result is published within reasonable time. Step 10: Cloud Voting System (CVS) process terminates successfully after display of polling result. Our proposed Cloud Voting System (CVS) deliver voting result within a reasonable time. It aims to mitigate the issue of latency by leveraging Cloud computing and Edge devices for Edge Computing. Multiple stakeholders plays a vital role during cloud based voting system which is conducted in a step by step manner. In this case, the output of previous step is considered as essential input for the next step, failing which the entire process will function out of sequence and will generate erroneous result. It is essential to explain the sequence of transactions which have been achieved using Sequence Diagrams in the subsequent section of this paper.

Fig. 1. Block diagram for proposed cloud voting system

3.1 Implications of Proposed Cloud Voting System In this sub-section, the implication of our proposed Cloud Voting System is discussed, assuming the following parameters. Let, Election Commission may be noted as EC, Presiding Officer may be noted as PO, Date of Election may be noted as DoE, Constituency or Seat name may be noted ConName, Constituency or Seat identification number may be noted as ConID, Contesting Political Parties for a specific constituency identified by ConID be noted as PP1, PP2, PP3 (assuming three different political parties can contest per constituency or seat, which may be enhanced in future.)

648

R. Purkayastha and A. Roy

Let, Political Party (PP) supported Candidates (PC) be noted as following: Candidate PC1 for Political Party PP1, Candidate PC2 for Political Party PP2, Candidate PC3 for PP3, Contesting Independent Candidate (IC) for a specific constituency or seat may be noted as ConID_IC1 and ConID_IC2 (assuming two independent candidates can contest per constituency or seat, which may be enhanced in future). In case the voter is not interested to support none of the above candidates for a specific constituency or seat, then it may be noted as NOTA_ConID. Voter List may be noted as VL_ConID, which will contain the details of various voters having unique VL_ConID_VoterID. As per our proposed Cloud Voting System (CVS), the smart card interface used by voter for voting purpose, namely Multipurpose Electronic Card (MEC) will provide unique identification number of voter using password and biometric parameter, which should match with VL_ConID_VoterID. The proposed Cloud Voting System is assumed to utilize these parameters for the successful execution of the voting process. All the necessary parameters are assigned by the Election Commission and transmitted to Presiding Officer (PO) via Public Kiosk. The Presiding Officer (PO), initiates the voting process for the voter. The voter sends its unique identification number, which is further verified by the Presiding Officer (PO). Depending on the result of verification, further course of action is taken. If the voter is valid, biometric parameter is requested by Presiding Officer (PO) and this is verified accordingly. For the valid voter, the choices of Candidate PC1 for Political Party PP1, Candidate PC2 for Political Party PP2, Candidate PC3 for PP3 and Independent Candidates ConID_IC1 and ConID_IC2 are displayed and voter chooses his/her candidate based on wisdom and good judgment. The Voter list, VL_ConID, is updated after casting of votes by each valid voter, whether it is a NOTA vote or vote for a Political Candidate or Independent Candidate. The Voting report is generated for checking the status of the voting process. Both polled vote list and NOTA vote list updates are maintained for the voting process.

4 Sequential Transactions of Cloud Voting System A Sequence Diagram is one of the important components of Unified Modeling Language (UML) to depict the flow of interaction among multiple objects of the application. The following Sequence Diagrams will explain our proposed Cloud Voting System (CVS) till the generation of report, which will be discussed in future. 1. Initialization of Voting Parameters: Figure 2 shows the Sequence Diagram of Election Commission, Kiosk and Presiding Officer. The constituency, polling party, presiding officer, contesting political parties, contesting candidates are assigned and all information are transmitted to Presiding Officer. Presiding Officer acknowledges the information received from Election Commission, which is notified to the Election Commission.

A Study on Sequential Transactions Using Smart Card

649

Fig. 2. Sequence diagram of initialization of voting parameters

2. Verification of Voter: Figure 3 shows the Sequence Diagram of Presiding Officer, Kiosk and Voter. This voting process is initiated by Presiding Officer for a specific constituency. All information regarding public display of contesting political parties and candidates are issued to the voter. The voter acknowledges to the Presiding Officer using public kiosk. Now the presiding officer initiates voter identification and the voter sends its unique identification number and detail to the Presiding Officer. This verification can lead towards any of the two situations, i.e. Successful validation of voter or Unsuccessful validation of voter. In case of successful validation, the voting process is allowed to proceed further, whereas in case of Unsuccessful validation, the voting process is aborted, and voter is notified using system timeout. After that, the voter is asked to enter the password. If the verification process is a success, the request for sending biometric parameter is forwarded to the voter, otherwise the process is aborted. The voter sends its biometric parameter to the Presiding Officer. The Presiding Officer, on verification of the parameter sends the “Success” or “Failure” message to the voter after which the voting process continues towards next step only in case of successful validation. 3. Casting of Vote: Figure 4 shows the Sequence Diagram of Presiding Officer, Kiosk and Voter. After the successful validation of the voter, the list of contesting candidates and political parties for a specific constituency are displayed to the voter. The voter casts its vote as per personal choice. An acknowledgement slip is generated to record completion of voting. 4. Voter list Updation: Figure 5 shows the Sequence Diagram of Presiding Officer, Kiosk and Voter list. The voter list is updated with reference to several parameters including the pre-poll and post-poll data, which is explained through Fig. 5. The updated Voter list is sent to Presiding Officer for continuation of voting process till the voter list is exhausted. 5. Voting Report: Figure 6 shows the Sequence Diagram of Presiding Officer, Kiosk and Election Commission. At this stage a clear picture about the present status of voting process is generated for the citizen. If it is successful voting, the proposed voting system is prepared to accept the mandate of next voter. Whereas in case of

650

R. Purkayastha and A. Roy

Fig. 3. Sequence diagram of verification of voter

A Study on Sequential Transactions Using Smart Card

651

Fig. 4. Sequence diagram of casting of vote

Fig. 5. Sequence diagram of voter list updation

any anomaly, the proposed voting system is prepared to conduct re-poll thereby recording the reason for re-poll in the cloud voting server.

652

R. Purkayastha and A. Roy

Fig. 6. Sequence diagram of voting report

5 Conclusion Our proposed Cloud Voting System (CVS) facilitates the voters to cast their vote from their comfort and safe zone without the hassles of standing in long queues, especially during any pandemic situation like Coronavirus (COVID19). Furthermore, the voter can avoid any unforeseen incident caused by any non-state actor to threat and manipulate the voter. As our proposed Cloud Voting System (CVS) is based on two-level authentication using Multi-purpose Electronic Card (MEC) based unique identification number and biometric parameter of voter paired with terminal electronic device, it is safe and secure throughout the electronic transaction. As it is essential to understand the sequence of transactions among various entities of our proposed voting system, in this paper we have shown the Sequence Diagrams of entities and primary actors. As the future scope of this research work, the deployment of the proposed Cloud Voting System (CVS) will be explained using the Unified Modeling Language (UML) Diagrams.

References 1. Paul, S., Bandyopadhyay, K., Roy, A.: A study on integrated cloud education system. In: Sengodan, T., Murugappan, M., Misra, S. (eds.) Advances in Electrical and Computer Technologies. LNEE, vol. 711, pp. 289–295. Springer, Singapore (2021). https://doi.org/10.1007/ 978-981-15-9019-1_26 2. Roy, A.: Object-oriented modeling of multifaceted service delivery system using connected governance. In: Jena, A.K., Das, H., Mohapatra, D.P. (eds.) ICDCIT 2019. SBPR, pp. 1–25. Springer, Singapore (2020). https://doi.org/10.1007/978-981-15-2455-4_1 3. Roy, A.: Smart delivery of multifaceted services through connected governance. In: 2019 3rd International Conference on Computing Methodologies and Communication (ICCMC), pp. 476–482. IEEE, March 2019

A Study on Sequential Transactions Using Smart Card

653

4. Biswas, S., Roy, A.: An intrusion detection system based secured electronic service delivery model. In: 2019 3rd International Conference on Electronics, Communication and Aerospace Technology (ICECA), pp. 1316–1321. IEEE, June 2019 5. Ghosh, A., Das, T., Majumder, S., Roy, A.: Authentication of user in connected governance model. In: Batra, U., Roy, N.R., Panda, B. (eds.) REDSET 2019. CCIS, vol. 1230, pp. 110–122. Springer, Singapore (2020). https://doi.org/10.1007/978-981-15-5830-6_10 6. Khatun, R., Bandopadhyay, T., Roy, A.: Data modeling for E-voting system using smart card based E-governance system. Int. J. Inf. Eng. Electron. Bus. 9(2), 45 (2017) 7. Mohapatra, S., Paul, K., Roy, A.: Object-oriented modeling of cloud healthcare system through connected environment. In: Mandal, J.K., Mukhopadhyay, S., Roy, A. (eds.) Applications of Internet of Things. LNNS, vol. 137, pp. 151–164. Springer, Singapore (2021). https://doi.org/ 10.1007/978-981-15-6198-6_14 8. Purkayastha, R., Roy, A.: An Integrated Environment for Cloud Voting System using Edge Computing (2021). Available at SSRN 3769149 9. Purkayastha, R., Roy, A.: Object oriented modelling of cloud voting system. In: 2021 Asian Conference on Innovation in Technology (ASIANCON), pp. 1–7. IEEE, August, 2021 10. Mobin, G., Roy, A.: A literature review on cloud based smart transport system. In: 2021 5th International Conference on Trends in Electronics and Informatics (ICOEI), pp. 1245–1250. IEEE, June 2021

Detecting Spinal Abnormalities Using Multilayer Perceptron Algorithm Arju Manara Begum, M. Rubaiyat Hossain Mondal, Prajoy Podder, and Subrato Bharati(B) Institute of Information and Communication Technology, Bangladesh University of Engineering and Technology, Dhaka 1205, Bangladesh {0416312030,rubaiyat97}@iict.buet.ac.bd, [email protected], [email protected]

Abstract. Integration of Internet of Healthcare Things (IoHT) and Machine Learning (ML) can be used successfully in healthcare systems to increase the accuracy of computer-aided diagnosis. This paper emphases on the application of IoHT and ML in detecting spinal abnormalities, which can be integrated with IoHT. The novelty of this work is in the use of multilayer perceptron (MLP) to a spinal dataset to obtain high accuracy in spinal abnormality detection. The dataset has 310 samples and is freely available on Kaggle repository. We use Pearson correlation coefficient (PCC), ReliefF and Gain ratio (GR) filter-based feature selection methods to select the top 10, 8, 6 and 5 features according to relevance or weight of features in preprocessing stage. In classification stage, we use logistic regression (LR), support vector machine (SVM), and Bagging algorithm in addition to MLP. The experimental results indicate that the PCC feature selection technique and MLP classification algorithms give the most promising results. A maximum classification accuracy of 88.0645% is obtained when MLP is used after selecting the top 5 spinal features by PCC feature selection method. This accuracy obtained by MLP and PCC is higher than 86.96% reported in the literature of spinal disease. Keywords: IoHT · Spinal abnormalities · Feature selection · PCC · GR · ReliefF

1 Introduction In this paper, we have demonstrated a very common disease in Bangladesh, lower back pain (LBP), in terms of IoHT and machine learning (ML). Changes in a person’s lifestyle, accidents, job and workplace restrictions, and psychological issues all have an impact on lower back pain. Lower back discomfort can be affected by a number of different things, including irritation, strain, ligament sprain, or damage to the discs, bones, or joints in the human body as a whole. Back injuries cause degeneration in numerous body components due to persistent inflammation of the lumbar spine (lower back), spinal discs (between the vertebrae), muscles (spinal cord), and ligaments. Many causes of lower back pain can be identified, but just a few of them include spondylolisthesis, disc herniation, and lumbar spinal stenosis. People working in agriculture, heavy industry, factories, textile © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 654–664, 2022. https://doi.org/10.1007/978-3-030-96299-9_62

Detecting Spinal Abnormalities Using Multilayer

655

workers, cesarean mothers, older people suffering from osteoporosis, drivers, and riders working in different riding services, i.e., Uber, Food Panda etc., are most vulnerable to LBP. Early detection of LBP is very important as chronic LBP can result in disability. It is extremely difficult for people suffering from severe back pain to travel from village to village for treatment or to a hospital in a distant city. It would be very helpful for them if they got their treatment in local public hospitals. Each public hospital in Bangladesh needs to be equipped with wearable suits with sensors and high-speed internet for this kind of service. In our proposed system, first, a radiographic image of the spine of a patient will be collected from the sensors of a wearable suit. The system will automatically send it to an expert using the Internet to find out the physical data of the spine. In the next stage, this data will be used to predict spinal abnormalities by a ML model (MLM), and a report will be sent to a doctor. Finally, the doctor will prescribe medicine and send it to the patient. In the ML process, we use three feature subset selection methods, which are the Correlation Coefficient method, ReliefF, and Gain Ratio, to decrease the dimensionality of the LBP dataset, and different classification algorithms are used to find the maximum accuracy for LBP patients. So, the main focus of our research is to find a ML model that will give a more accurate output in terms of spinal abnormalities. In our experiment, we used the spinal dataset from Kaggle to train and test MLM. The result of our experiment shows a maximum accuracy of 88.065%, which is better than the accuracy values of 83.87%, 85.32%, and 86.96% reported in the literature of [1, 2], and [3], respectively. This work is in the consideration of the MLP algorithm in obtaining a reliable accuracy value in spinal abnormality detection. Hence, the key contributions of this work are summarized as follows. • Selecting more suitable features from the LBP dataset retrieved from the Kaggle repository by applying Pearson correlation coefficient (PCC), ReleifF and gain ratio (GR) methods. • Using ML algorithms of logistic regression (LR), support vector machine (SVM), multilayer perceptron (MLP) and bagging SVM approaches to identify critical aspects of LBP. • Comparing the MLP with LR, SVM and bagging SVM for several feature selection approaches including PCC, ReleifF and GR in terms of accuracy.

2 Related Works IoHT and Machine Learning can be combined to build an expert system that will provide a more reliable and fast diagnosis of diseases for doctors, and patients will get their services from anywhere, at any time, at a low cost [1–13, 17]. For example, in [4], the authors proposed an IoT based healthcare system for heart disease prediction and diagnosis. They used UCI heart disease dataset to train the ML-based prediction system, and then the trained classifier is used to test incoming patients for identify whether they suffer from heart disease or not. Patient data is collected by sensors in IoT networks. Finally, the user will be presented with the test results. For classification, they used LR, J48, MLP and SVM. Big data technologies were used by the authors of [5] to propose a

656

A. M. Begum et al.

novel architecture. Health data collected by IoT devices is absorbed into spark streaming via Kafka and then streamed to the user. After that, they used Spark to transform the traditional decision tree (DT) approach into a scalable, distributed, parallel, and rapid DT, which they applied to stream data from diverse illness sources to forecast health status. The system estimates a patient’s health state based on a variety of factors and then sends an alarm message to healthcare professionals, saving the details in a distributed database that may be used for health data analytics and real-time reporting. A variety of factors, including pelvic incidence, pelvic tilt, sacral slope, and others, are linked to lumbar spine lower back discomfort. Early identification and treatment of patients with spinal abnormalities has become more dependent on identifying lower back pain in those people. Low back pain affects a large percentage of the population and is a major cause of disability, lowering productivity and negatively impacting one’s quality of life. According to the definition, it’s a non-specific ailment that refers to complaints of pain and discomfort in or near the lumbosacral spine that can be caused by many conditions like inflammatory or degenerative, or even cancerous. Several research works [1–3] have been executed on Spine dataset published in Kaggle to detect the signal abnormalities. The authors in [1] applied five base classifiers i.e., Naïve Bayes (NB), MLP, Bayes Net, Random Forest (RF), and DT. Five base classifiers were combined for voting to detect lower back pain. In this paper, they did not use any preprocessing methods on the dataset. Classification accuracy of 81.9%, 83.87% and 83.87% are achieved for RF, NB and MLP, respectively. Authors in [2] implemented a ML model to identify spinal abnormalities on Spine dataset. They used PCA for extracting the most significant features, features with p-value greater than 0.05 are extracted from the dataset. Finally, RF and KNN algorithms are used for classification. The accuracy rates obtained after applying RF and KNN are 85.32% and 79.5% in identifying the abnormality. Authors in [3] split the dataset into train and test datasets. They have used PCA and Chi-square feature selection methods to extract the most significant features. For classification they have used SVM, LR, bagging SVM and bagging LR models. For training dataset, they have obtained accuracies 86.30%, 85.47%, 86.72% and 85.06% after applying them respectively. Accuracy rates for testing dataset after applying bagging SVM, LR, bagging LR and SVM are the same being 86.96%.

3 Lower Back Pain Dataset In the LBP dataset, there are 310 patients with a total of 12 different features. This data has been divided into two groups: normal records and abnormal records [14]. There are 100 normal patients and 210 abnormal patients [14]. Non-numerical categorical features were converted to numerical values as part of the data labeling process. The abnormal and normal values are recoded as 0 and 1. Using data preparation techniques, imperfect raw data can be transformed into a usable and efficient format. Data Preprocessing has two stages, as illustrated below: (a) Data Transformation: It is necessary to convert all of the nominal data into numerical data. (b) Missing Data Handle: In this case, the null values are replaced with the calculated attribute’s mean value.

Detecting Spinal Abnormalities Using Multilayer

657

4 Methodology Our research methodology is depicted in Fig. 1. Spine dataset of Kaggle [14] contains continuous features, so before applying the feature subset selection method GR, we have to apply a discretization process to the features. But the GR method of WEKA supports continuous features, and applying discretization is not necessary. For the other two feature selection methods (PCC and ReliefF) discretization is not necessary. After selecting the top 10, 8, 6, and 5 features by applying filter methods, we apply four classification algorithms (LR, SVM, MLP and Bagging) to calculate the accuracy of spinal abnormality prediction. We conduct WEKA to implement our model. The crossvalidation process of WEKA is used during classification.

Fig. 1. Work flow diagram

Moreover, feature selection is a pre-processing step of Machine Learning to select relevant and non-redundant features to reduce the dimensionality of a dataset. Removing redundant and irrelevant features helps ML algorithms to improve the performance in terms of accuracy and time to build the model. Feature selection techniques can be classified into feature ranking techniques and feature subset selection techniques depending on either individual evaluation or subset evaluation [6]. Using the feature ranking method, each feature is evaluated individually for a predetermined threshold value based on various selection metrics i.e., gain ratio, symmetric uncertainty, and information gain. The highest-ranked features are then selected as significant features. If you utilize the subset evaluation method, also known as the feature

658

A. M. Begum et al.

subset selection method, a search strategy like “best first” is used to create the number of possible combinations of the subsets, and then metrics like “correlation” and “consistency” are used to evaluate each feature subset. Feature section methods can also be categorized into three different methods depending on how the supervised learning algorithm is applied in the feature selection process. These are the filter, wrapper, and embedded methods [7]. Using the wrapper method, the produced feature subsets may be validated using any of the search strategies, including the supervised learning algorithm. Instead of using the full learning algorithm, the embedded method just employs a portion of the learning algorithm in the selection process, which results in higher accuracy [8, 9]. Filter techniques of feature selection are independent of the evaluation model used in ML algorithm, which use different measures like correlation, statistical dependencies of features to select relevant and non-redundant features. Filter methods can be further classified as univariate and multivariate approaches depending on the type of feature dependencies consider for the feature selection method. In univariate methods only features to class dependencies are considered which leads to selecting only relevant features [10]. Examples of univariate filter methods include Information Gain, Gain Ratio, Symmetric Uncertainty, Relief, chi-square, etc. In multivariate methods feature to feature dependencies are also considered which leads to selecting relevant and nonredundant features [10]. 4.1 Correlation Based Feature Selection Methods Correlation between features can be measured based on linear and nonlinear relations between two features. For a linear correlation, classical correlation coefficients and nonlinear correlation statistical measures such as information theory are used [15]. One of the most common classical correlation coefficients is Pearson correlation coefficient (PCC). For a pair of features (y, z), the correlation coefficient r is given by the equation Py,z = 

cov(y, z) (var(y) ∗ var(z))

(1)

Where, var() is the variance of a feature and cov(y, z) denotes the covariance between x and y. Pearson correlation coefficient when used for multiple sample is commonly represented as Py,z . Suppose there are n samples in two random variable y and z, PCC is defined as [16, 17] n i=1 (yi − y)(zi − z)  ryz =  (2) n 2 n 2 (y − y) (z − z) i=1 i i=1 i r’s range is from −1 to 1, inclusive. If y and z are wholly uncorrelated, then r has a value of 1 or −1; otherwise, r is zero. The approach of measuring the correlation between two nonlinear features is based on the information-theoretical concept of entropy.

Detecting Spinal Abnormalities Using Multilayer

659

Entropy is defined as information of uncertainty, randomness, or surprise in a dataset. The entropy of a variable Y is expressed as  H (Y ) = − P(yi )log 2 P(yi ) (3) i

Entropy zero means dataset is extremely predicable. It has only one class. Entropy high means dataset is unpredictable of many mixed classes.   H (Y |Z) = P(z i ) P(yi |zi )log 2 (P(yi |zi )) (4) i

i

Where, P(yi |zi ) is denoted as the posterior probabilities of Y given the values of Z. The quantity by which the entropy of Y reductions reflects further information about Y given by Z and is known information gain and is provided by IG = H (Y ) − H (Y |Z) = H (Z) − H (Z|Y )

(5)

Gain ratio (GR) is often defined as [18], GR =

IG H(Y)

(6)

After normalization, the value of GR always has a range of between 0 and 1. Value 1 of GR means the knowledge of X completely predicts Y and value 0 of GR means X and Y are not related [18]. 4.2 ReleifF Relief is a feature selection algorithm motivated by instance-based learning [19, 20]. It was introduced by Kira and Rendell [21, 22]. Assumed threshold of relevancy, sample size, and a training data, ReliefF identifies those features which are statistically significant to the class with a threshold value of 0 to 1. It selects a sample from a training set randomly first, then finds the Nearest Hit from the same class and the Nearest Miss from the other class and updates the weight of each feature. It uses Euclid’s distance, or Manhattan Distance, for selecting Near Miss and Near Hit. After repeating the above actions multiple times, the weights of all features will be updated. Then it calculates the average weight of features (‘relevance level’) and finally selects those features whose average weight is above a provided threshold level. ReliefF is the updated version of the original Relief to mitigate this problem, while ‘F’ indicates the sixth algorithm variation [23].

5 Experimental Results All experiments were performed in WEKA platform as we mentioned early [24, 26]. We used default settings of WEKA for all three feature selection methods and classification algorithms except MLP. For MLP we used different folds (5, 10, 15, 20) and learning rate (0.3, 0.4) to get better accuracy.

660

A. M. Begum et al.

5.1 Feature Selection Results The ranking orders of the three feature selection techniques are given in Table 1. The order of the features in this table is based on their relevance or merit. From the table, it can be said that degree spondylolisthesis and pelvic incidence are the two top-ranked features in all three FS methods. Moreover, pelvic incidence, degree spondylolisthesis, pelvic tilt, pelvic radius and lumbar lordosis angle are ranked between 1 and 5, but not in the same order for PCC, ReliefF and GR. Table 1. Features set after applying feature selection methods Rank

PCC

ReliefF

GR

1

Degree spondylolisthesis

Degree spondylolisthesis

Degree spondylolisthesis

2

Pelvic incidence

Pelvic incidence

Pelvic incidence

3

Pelvic tilt

Pelvic radius

Lumbar lordosis angle

4

Lumbar lordships angle

Pelvic tilt

Pelvic tilt

5

Pelvic radius

Llumbar lordosis angle

Sacral slope

6

Sacral slope

Cervical tilt

Pelvic radius

7

Cervical tilt

Scoliosis slope

Thoracic slope

8

Scoliosis slope

Direct tilt

Cervical tilt

9

Pelvic slope

Sacrum angle

Sacrum angle

10

Thoracic slope

Thoracic slope

Pelvic slope

11

Direct tilt

Scoliosis slope

Direct tilt

12

Sacrum angle

Pelvic slope

Scoliosis slope

5.2 Classification Results For evaluation of classification algorithms, LR, SVM (linear), MLP and Bagging [25] are applied to the Spine dataset. When we applied these ML algorithms to a full set of dataset, the accuracy rates were 84.1935%, 84.1935%, 80.6452% (folds = 15 and LR = 0.3) and 84.5161% for LR, SVM, MLP and Bagging respectively. We selected ten, eight, six, and five features according to relevance after applying PCC, ReliefF and GR to different phases of our experiment and applying ML algorithms to compare accuracy rates. The results are depicted in Tables 2, 3, 4, 5. In Table 2, we see that bagging shows a highest accuracy rate of 85.4839% for the top ten features of PCC. In Table 3, we see that bagging shows a highest accuracy rate of 85.8065% for the top eight features of PCC. In Table 4, we see that MLP is showing a highest accuracy rate of 86.4516% for 15 folds and a 0.3 learning rate for the top six features of PCC and GR. And interestingly, MLP is also showing the lowest accuracy rate at 83.871% for 5 folds and a learning rate of 0.3. In Table 5, we see that MLP is showing a highest accuracy rate of 88.0645% for 20 folds and a learning rate of 0.4 for the top five features of PCC and ReliefF.

Detecting Spinal Abnormalities Using Multilayer

661

Table 2. Accuracy rates for top 10 features of PCC, ReliefF and GR FSM

LR

SVM

MLP

Bagging (SVM)

PCC

83.5484%

84.1935%

81.9355% (fold = 5)

85.4839%

ReliefF

82.9032%

84.1935%

80.6452% (fold = 15)

84.5161%

GR

84.1935%

83.2258%

78.7097% (fold = 15)

83.5484%

Table 3. Accuracy rates for top 8 features of PCC, ReliefF and GR FSM

LR

SVM

MLP

Bagging (SVM)

PCC

85.1613%

84.8387%

84.1935% (fold = 15)

85.8065%

ReliefF

84.8387%

84.5161%

85.1613% (fold = 10)

85.4839%

GR

83.8710%

85.1613%

83.2258% (fold = 20)

85.1613%

Table 4. Accuracy rates for top 6 features of PCC, ReliefF and GR FSM

LR

SVM

MLP

Bagging (SVM)

PCC

85.4839%

85.1613%

86.4516% (Fold = 15)

85.8065%

ReliefF

85.8065%

85.1613%

83.871% (fold = 5)

86.1290%

GR

85.4839%

85.1613%

86.4516% (Fold = 15)

85.8065%

Table 5. Accuracy rate for top 5 features of PCC, ReliefF and GR FSM

LR

SVM

MLP

Bagging (SVM)

PCC

85.4839% 84.8387% 88.0645% (fold = 20, learning rate = 0.4) 85.8065%

ReliefF 85.4839% 84.8387% 88.0645% (fold = 20, learning rate = 0.4) 85.8065% GR

82.9032% 83.2258% 83.5484% (fold = 15, learning rate = 0.4) 84.5161%

Figure 2 depicts the comparison of ML algorithms with other classification algorithms for the top ten, eight, six, and five features. Figure 2(a) depicts the accuracy curves of ML algorithms for the top 10 features in a chart. Here we see that bagging has an upper bound of accuracy and MLP has a lower bound of accuracy. Figure 2(b) depicts the accuracy curves of ML algorithms for the top 8 features in a chart. Here we see that Bagging has an upper bound of accuracy and MLP has a lower bound of accuracy. Figure 2(c) depicts the accuracy curves of ML algorithms for the top six features in a chart. Here we see that the bagging has an MLP upper bound and lower bound of accuracy. Figure 2(d) depicts the accuracy curves of ML algorithms for the top five features. Here we see that the MLP has an upper bound and LR has a lower bound of accuracy.

662

A. M. Begum et al.

Fig. 2. ML algorithms for the top (a) 10 features, (b) 8 features, (c) 6 features, (d) 5 features

6 Conclusion The main goal of this paper is to provide an expert system that will assist doctors in diagnosing and identifying spinal abnormalities more accurately from anywhere at any time. We used the spine dataset from Kaggle in our experiment, and using feature selection methods PCC, ReliefF, and GR, we found that PCC is the most reliable method to be used in the prepossessing stage when 10, 8, 6, and 5 top-ranked features are used. PCC and ReliefF have selected the same five top-ranked features in a different order, and anyone can be selected for filtering features. For classification, we have used LR, SVM, MLP, and Bagging with SVM. MLP with 20 folds and a learning rate of 0.4 has shown 88.0645% accuracy. So, experts can use MLP along with PCC or ReliefF to identify spinal abnormalities of patients. The PCC, ReliefF and GR methods have been deployed in this experiment, which considers only feature to class relations between features and selects relevant feature subsets only. To select relevant and non-redundant feature subsets, feature to feature relationships have to be considered. Multivariate methods can be used to select nonredundant features. An ensemble of FS methods can be used to get a diverse and stable subset of features. In our future work, we will build an ensemble model of feature selection methods for a relevant and non-redundant feature subset with diversity and stability, which eventually will help experts to identify spinal abnormalities more accurately. The computation time of classifiers are not explored in this work. Additionally, the prediction accuracy of various less frequently used classifiers for spinal disorders has not been evaluated. Future research may collect data from a bigger sample of individuals without and with spinal issues. Novel classification approaches may be improved to increase the detection and prediction rates of spinal abnormalities from LBP.

Detecting Spinal Abnormalities Using Multilayer

663

References 1. Bhatt, M., Dahiya, V., Singh, A.K.: A comparative analysis of classification methods for diagnosis of Lower Back Pain. Oriental J. Comput. Sci. Technol. 11(2), 135–139 (2018) 2. Abdullah, A.A., Yaakob, A., Ibrahim, Z.: Prediction of spinal abnormalities using machine learning techniques. In: 2018 International Conference on Computational Approach in Smart Systems Design and Applications (ICASSDA), pp. 1–6. IEEE (2018) 3. Raihan-Al-Masud, M., Mondal, M.R.H.: Data-driven diagnosis of spinal abnormalities using feature selection and machine learning algorithms. PLoS ONE 15(2), e0228422 (2020). https://doi.org/10.1371/journal.pone.0228422 4. Ganesan, M., Sivakumar, N.: IoT based heart disease prediction and diagnosis model for healthcare using machine learning models. In: 2019 IEEE International Conference on System, Computation, Automation and Networking (ICSCAN), pp. 1–5. IEEE (2019) 5. Ed-daoudy, A., Maalmi, K.: A new Internet of Things architecture for real-time prediction of various diseases using machine learning on big data environment. J. Big Data 6(1), 1–25 (2019) 6. Urbanowicza, R.J., Meekerb, M., La Cavaa, W., Olsona, R.S., Moorea, J.H.: Relief-based feature selection: introduction and review. J. Biomed. Inf. https://doi.org/10.1016/j.jbi.2018. 07.014 7. Guyon, I., Elisseeff, A.: An introduction to variable and feature selection. J. Mach. Learn. Res. 3(Mar), 1157–1182 (2003) 8. Kohavi, R., John, G.H.: Wrappers for feature subset selection. Artific. Intell. 97(1–2), 273–324 (1997). https://doi.org/10.1016/S0004-3702(97)00043-X 9. Breiman, L., Friedman, J.H., Olshen, R.A., Stone, C.J.: Classification and Regression Trees. Wadsworth and Brooks (1984) 10. Saeys, Y., Inza, I., Larranaga, P.: A Review of Feature Selection Techniques in Bioinformatics. Oxford University Press, Bioinformatics (2007) 11. Bharati, S., Podder, P., Mondal, M.R.H.: Hybrid deep learning for detecting lung diseases from X-ray images. Inf. Med. Unlock. 20, 100391 (2020) 12. Mondal, M.R.H., Bharati, S., Podder, P.: CO-IRv2: optimized InceptionResNetV2 for COVID-19 detection from chest CT images. PloS one 16(10), e0259179 (2021) 13. Bharati, S., Mondal, M.R.H.: 12 Applications and challenges of AI-driven IoHT for combating pandemics: a review. Comput. Intell. Manag. Pand. (2021). https://doi.org/10.1515/978311 0712254-012 14. https://www.kaggle.com/sammy123/lower-back-pain-symptoms-dataset. Accessed on 15 Nov 2021 15. Jianga, S., Wang, L.: Efficient feature selection based on correlation measure between continuous and discrete features. Inf. Process. Lett. 116(2), 203–215 (2016) 16. Guha, R., Ghosh, K.K., Bhowmik, S., Sarkar, R.: Mutually Informed Correlation Coefficient (MICC)-a new filter based feature selection method. In: 2020 IEEE Calcutta Conference (CALCON), pp. 54–58. IEEE (2020) 17. Mondal, M.R.H., Bharati, S., Podder, P.: Diagnosis of COVID-19 using machine learning and deep learning: a review. Curr. Med. Imaging (2021). https://doi.org/10.2174/157340561766 6210713113439 18. Novakovi´c, J.: Toward optimal feature selection using ranking methods and classification algorithms. Yugoslav J. Oper. Res. 21(1) (2016) 19. Aha, D.W., Kibler, D., Albert, M.K.: Instance-based learning algorithms. Mach. Learn. 6(1), 37–66 (1991) 20. Callan, J.P., Fawcett, T., Rissland, E.L.: Cabot: an adaptive approach to case-based search. IJCAI. 12, 803–808 (1991)

664

A. M. Begum et al.

21. Kira, K., Rendell, L.A.: The feature selection problem: traditional methods and a new algorithm. AAAI. 2, 129–134 (1992) 22. Kira, K., Rendell, L.A.: A practical approach to feature selection. In: Proceedings of the Ninth International Workshop on Machine Learning, pp. 249–256 (1992) 23. Kononenko, I.: Estimating attributes: analysis and extensions of relief. In: Bergadano, F., Raedt, L. (eds.) ECML 1994. LNCS, vol. 784, pp. 171–182. Springer, Heidelberg (1994). https://doi.org/10.1007/3-540-57868-4_57 24. Bharati, S., Rahman, M.A., Podder, P.: Breast cancer prediction applying different classification algorithm with comparative analysis using WEKA. In: 2018 4th International Conference on Electrical Engineering and Information and Communication Technology (iCEEiCT), pp. 581–584. IEEE (2018) 25. Mondal, M.R.H., Bharati, S., Podder, P., Podder, P.: Data analytics for novel coronavirus disease. Inf. Med. Unlock. 20, 100374 (2020) 26. https://www.cs.waikato.ac.nz/ml/weka/. Accessed on 15 Nov 2021

Wireless Sensor Networks Time Synchronization Algorithms and Protocols Message Complexity Comparison: The Small-Size Star-Topology Case Giuseppe Coviello , Gianfranco Avitabile , and Antonello Florio(B) Department of Electrical and Information Engineering, Polytechnic University of Bari, 70125 Bari, BA, Italy {giuseppe.coviello,gianfranco.avitabile,antonello.florio}@poliba.it https://etlclab.poliba.it

Abstract. The paper analyzes the issues related to the time synchronization in a sensor network, evidencing the related problems and introducing a brief revision of the algorithmic approaches. Among the possible solutions, we analyze a protocol-agnostic fractional low-power algorithm proposed by the authors. After a quick recall on the theory of operations and on the obtained results, we introduce a metric that allows to compare three of the most employed time synchronization algorithms with the one proposed from a power-efficiency point of view in the case of small networks with star topology.

Keywords: Time synchronization efficiency

1

· Wireless Sensor Networks · Power

The Issue of Time Synchronization in Sensor Networks

One of the most important constraints to address when dealing with sensor networks, is to guarantee each node of the network agrees on the same time value. When multiple nodes are involved, the probability of time mismatches becomes higher, and this may lead to critical issues. Usually, sensor nodes are devoted to collect data coming from environmental variables or biological signals and the monitored signals come from different sources. Hence, it is important to guarantee a good degree of time synchronization in order to obtain acceptable results in the post-acquisition phase, for instance during data merging operations. This problem becomes more prominent when looking at wirelessly connected nodes, called Wireless Sensor Networks (WSNs). In order to reduce the production costs of each unit, the clock source is generated by crystal oscillators that are stable enough to make microcontrollers and microprocessors operations well performed. One device devoted to extracting the c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 665–674, 2022. https://doi.org/10.1007/978-3-030-96299-9_63

666

G. Coviello et al.

time value from the timing circuitry is the Real-Time Clock (RTC) device. However, real clocks experience some non-idealities, due to temperature variations [19], aging or load capacitance mismatches [3,4]. These artifacts can be modeled [16,18] as the initial offset, that is the initial time difference between the device to synchronize and the reference time, the oscillation frequency offset between the device and the reference clock, and the rate at which it occurs, respectively the clock skew and the clock drift. In this paper, we explore and compare different solutions for achieving time synchronization in a small-size WSN. Being the power efficiency one of the critical constraints in WSNs as in many other research fields [8,14], we will introduce a metric to compare a subset of the proposed algorithms from a power-efficiency point of view. The paper is organized as follows: in Sect. 2 we offer a brief revision of the literature on WSN synchronization protocols and algorithms. Then in Sect. 3 we focus on the background and main results regarding a novel protocol-agnostic time synchronization algorithm the author of this paper proposed recently. In Sect. 4 we propose a possible metric to compare from a power-efficiency point of view the described approaches. Finally, the conclusions close this work.

2

Algorithms and Protocols for Time Synchronization

Many protocols manage time coordination and many approaches have been developed during years of research in the field. Nevertheless, almost every protocol has a single time reference node for the whole network. A possible taxonomy of time synchronization algorithms for WSNs is related to the physical extension of the network itself. In fact, if we deal with WSNs that are large, the propagation delay could not be considered as negligible. Also, the packet loss events are more frequent and the clock sources can introduce artifacts that are not similar in their behavior due to fact that the boards are placed in locations with different environmental characteristics. In very large WSN, there are many approaches to simplify and speed up the synchronization phase. One example is presented in [12] in which the network is divided into broadcast subdomains and the synchronization message dissemination is performed thanks to the creation of a spanning tree. Spanning trees are also employed in [13] and [9] as a way to manage multi-hop. Another approach inherited from Delay Tolerant Networks (DTN) is the Time Synchronization Algorithm for Mobile WSNs (TSAM) proposed in [17], in which the time-reference node physically moves, disseminating the synchronization message in the entire network physical perimeter. For those scenarios in which the links between the nodes are not stable, a resistive-network-inspired approach like the one proposed in [1] can be very helpful because of its flexibility. However, given the computational weight associated with each step, the effort is justified only for very large WSNs. For what concerns small-extension WSNs like Wireless Body Area Networks (WBANs), there are some algorithms that stand out when analyzing the scientific literature because of their large employment. One of those is the Timing

Wireless Sensor Networks Time Synchronization Algorithms

667

synchronization Protocol for Sensor Networks (TPSN) [9]. TPSN is a dynamic pair-wise synchronization approach in which synchronization-request senders are synchronized to receivers. It has two main phases. The first one is a discovery phase, initiated by the root node. At this stage, a broadcast message is sent by every node in order to establish a spanning-tree hierarchy based on level number exchange. Then, in the synchronization phase, each node belonging to level i sends a synchronization_pulse to the one on level i − 1 (assuming the root node has level 0), which answers with an ACK message. Given this architecture, the protocol is also suitable for large WSNs. Reference Broadcast Algorithm (RBS) [7] does not employ dedicated synchronization messages, and it ensures a time synchronization that is relative to the master’s timer. Sometimes the absolute network time does not have to be correct (e.g. with respect to the UTC), but every node must agree on a common time. Once the root-node has sent a reference_packet, the receivers log the local time of reception of the packet. Thereafter, each receiver sends to the others its observation and, once all the receivers have collected the whole network observations, they average them, in order to reach a consensus on the average time of reception. Unfortunately, RBS is not optimized for low-power contexts. However, in [10], authors provide a revised RBS-based approach for low-power energy-harvested WSN. This RBS flavor integrates a linearly incremental switchoff time for the radio section. The increment is managed by a routine that checks the synchronization error slope, in particular, the switch-off time is increased when the synchronization error decreases. Otherwise, if the error is increasing, the off-time is reduced referring to the slope of the interpolated error samples. Another largely employed algorithm is the Flooding Time Synchronization Protocol (FTSP) [13]. FTSP synchronizes the sender time with multiple receivers using message timestamped at the MAC layer. The approach is simple, i.e. the central node floods the entire network with a broadcast packet containing its local MAC time so that each receiver logs the reception time and adjusts its own to it. Along with those algorithms, the authors of this paper recently proposed a novel protocol-agnostic time synchronization algorithm for small-size WSN, in particular concept for WBANs with medical purposes where the sensors are used to continuously acquire patients’ data [2]. The proposed algorithm is called Fractional Low-Power Synchronization Algorithm (FLSA) [5]. We will recall its basics in the next section to understand how FLSA operates and introduce the results that lead us to list also this algorithm in the here mentioned comparison. In order to properly compare the classical algorithms previously discussed in detail, let us assume the case of a star-topology WSN with a chosen reliable center node, that is reachable from all the N peripheral nodes. In classical distributed systems theory, the complexity CA of an algorithm or protocol A is measured as the number of messages exchanged between a set of nodes to reach a given goal for the set itself [11]. In this section, we adopt this metric to compute and compare the complexity.

668

G. Coviello et al.

Using the already introduced metric, it is possible to compute the following complexities. In TPSN, the root node sends 1 time_sync packet, in broadcast. Then, each node sends a synchronization_pulse and the receiver must answer with the same amount of acknowledgments. Globally the complexity can be quantified as (1) CTPSN = 1 + N + N = 2 · N + 1 For what concerns RBS, it must send 1 broadcast reference_packet, then each receiver must send N packets with local time. So, the overall result gives: CRBS = N 2 + 1

(2)

Then, regarding FTSP, a single broadcast sending is performed, so that, in absence of the spanning-tree construction, CFTSP = 1

(3)

As it will be clear in Sect. 3, FLSA’s complexity can be evaluated using to the same metric, bringing to: (4) CFLSA = 1 = CFTSP Hence, the following relationship holds under the exposed hypothesis: CFTSP = CFLSA ≤ CTPSN ≤ CRBS

3

(5)

Fractional Low-Power Synchronization Algorithm

As already anticipated in Sect. 1, the System Timer (ST) is supplied by a RTC module, especially in low-power devices. The RTC modules integrate a set of registers: the counter register RC , the capture/compare register RCOMP , the prescaling register RP and a tick event generator. We assume that the master is sending a synchronization message every TM , the so-called synchronization slot duration. In the case of the simplest timer adjustment implementation, each slave can increase or decrease its ST value by only changing the value in the RC register, that is, the number of ticks, tRTC , determining the tST , namely the ST granularity. However, in this way it is not possible to obtain all the possible values for the ST increment/decrement ΔST being RC an integer value. If we aim to handle also fractional values, one of the classical techniques is the multi-modulus division [15] that consists of employing two or more counters/registers. FLSA is based on four whose interactions are depicted in Fig. 1. On its first stage, FLSA is initialized, once the desired TM and ST granularity tST are set. With this data, the system of equations to be solved to evaluate the initialization values is: ⎧ A B ⎪ ⎨RC · IA + RC · IB = NM   (6) TM ⎪ I + I = ⎩A B tST

Wireless Sensor Networks Time Synchronization Algorithms Fractional Timing Initialization

669

Slot Skipping Algorithm Received Messages Queue Management Algorithm

Slave Timer Correction Algorithm

Fig. 1. FLSA routines interaction diagram

with NM being the number of ticks determining the synchronization slot TM , RiC the i-counter register value and Ii the number of times that value is took in the average determining the fractional value. Note that i = A, B since FLSA adopts a dual-modulus fractional division for its implementation. When solving the system we obtain a unique solution that represents the initialization vector for the fractional timer. The second subalgorithm manages the reception queue of the messages and it is called Received Messages Queue Management (RMQM) Routine. The routine is in charge of prioritizing the synchronization message processing with respect to all other kind of messages. Also, RMQM has to evaluate the equivalent accumulated time delay with respect to the master’s timer, since the routine considers also the number of skipped slots on its delay evaluation. Once the two thresholds ΔTM,min and ΔTM,max are set to be the lower and upper bound to the master-slave time delay respectively, RMQM can consider the the slave to be synchronous when |ΔTM | ≤ ΔTM,min and the slot skipping routine is executed. Otherwise, if the other case occurs, the slave simply updates its ST with the received value, since the required number of synchronization rounds to be performed is too high to handle and the entire synchronization process too long to be waited. As already anticipated, when the slave is considered to be synchronous to the master, it is possible to save energy by switching off the radio section of the node, namely running the slot skipping routine in Fig. 2. This routine operates in terms of synchronization slots, computing a weighted moving average of the delays in order to establish if the node is capable of being synchronized on a longer term. The last routine manages the timer correction. The subroutine checks if a simple variation of the IA and IB counters is sufficient. Otherwise, it varies the B RA C and RC dividers values according to the increment or decrement required by ST. Therefore, depending on the sign of ΔNM , the various possibilities reported in the flow-chart in Fig. 3 may occur. The algorithm was tested during a 7-days long measurement campaign. The measurement setup values campaign were TM = 20 s, tRTC = 30.517 µs, tST = 100 µs, ΔTmin = 500 µs, ΔTmax = 5 s and Smin sync = 5 slots. The sampling

670

G. Coviello et al. START

Ssync←Ssync +1+ Ss

Update AVGwd

Ssync≥ Sminsync

No

END

Yes

● ● ●

Subtract AVGwd Skip Ss slots Ss←Ss +1

Fig. 2. Flowchart of the slot-skipping routine being Ssync (j) the number of slots in which the node j can be considered to be synchronous to the master, SS the number of skipped slots and AVGwd the weighted moving average value.

frequency of the Inertial Measurement Unit (IMUs) employed for the campaign was fIMU = 100 Hz. It was then possible to establish the following results [5]. From an accuracy point of view, the maximum number of skipped slots before requiring a new calibration is equal to 53. Thus, for 1474 slots (i.e., 29480 s with the experiment settings) the time mismatch was lower than the set ΔTmin = 500 µs. On the other hand, the worst case exhibits only 12 skipped slots. Hence, the minimum time interval in which the whole network was synchronized over a 7-days long experiment was 1600 s. From a power consumption point of view the experiments showed a power saving ranging from 85% to 96% for each slave of the network when compared to the simple periodic synchronization message broadcast dissemination.

4

A Metric for the Power-Efficiency Comparison

In this section, we introduce a comparison between FLSA and other synchronization algorithms in terms of complexity. In Sect. 1, we already introduced the number of exchanged messages during the synchronization procedure as a possible and common employed in distributed-systems theory complexity metric. In our comparison, we will take as an example the two algorithms most similar to FLSA, which are FTSP and RBS. In particular, we will consider also the RBS flavor for energy harvested networks introduced in Sect. 2 and here referred to as RH. Actually, we already introduced a message complexity evaluation in Sect. 2. However, as stated in [6], also the so called empirical parameters have an impact when characterizing the complexity of an algorithm. One of the most influential parameters is the synchronization period TM . However, for the majority of

Wireless Sensor Networks Time Synchronization Algorithms

671

START

No

IB>|ΔNM|

Yes

ΔNM>0

Yes

Yes

IA>ΔNM

No

No

IA←IA-ΔNM IB←IB+ΔNM

RC_A= RC_B=1

RC_A= RC_B=-1

END

Fig. 3. Timer correction routine flowchart

algorithms and protocols proposed in the literature, it is not simple to have a common metric taking into account for this parameters. Aiming to compare the performances also in terms of power efficiency, we introduce here a new metric. Definition 1. We define the radio section power efficiency of an algorithm j in the case i as (i,j)  Ton (7) ηji = 1 − (i,j) · 100% Tobs (i,j)

being Ton the time in which the algorithm i asks for the radio-section to be (i,j) switched on in the case j, and Tobs the observation time. (i,j)

(i,j)

For compactness, we will express Ton and Tobs in terms of number of synchronization slots. For what concerns FTSP and RBS in their simplest implementation, we know that they send a synchronization message at the beginning of every synchronization round. This implies that the efficiency computed according to the previously introduced metric is: (8) ηRBS = ηFTSP = 0% as expected, since those algorithms were not optimized for low-power consumption contexts. On the other hand, if we consider RH, it is possible to obtain better results. Let us assume that the off-time increments are discrete values multiple integers BC ) of the time slot. We can then compute the efficiency in the best case (BC, ηRH WC and the worst case (WC, ηRH ). Let us, further, assume that the total number of (RH,j) slots in which the radio stays switched-on during Tobs is B. Then, given the linear increasing switch-off time property: BC ηRH =1−

B+

B

B

i=1

i

≈1−

2 1+B = 3+B 3+B

(9)

672

G. Coviello et al. RH best case efficiency computation

100 95 90 85 80 75 70 65 60 55 50 45 40 35 30 0

5

10

15

20

25

30

35

40

45

50

55

60

Fig. 4. RH best case efficiency computation for B ∈ [1, 60]

An example of efficiency curve computed for RH is shown in Fig. 4. The worst case does not depend on the number of rounds, and it is such that: WC ηRH = ηRBS

(10)

Let us now compute the efficiency of FLSA in the BC and WC. For what concerns the BC, we know that we start from k−slots in which the radio portion is switched on, followed by k−slots in which it is switched off (we can refer to this phase as transient). After this, the radio stays on for 1 slot, followed by k +1 slots in which it is off. Then, again, after 1 slot being on, it turns off for k + 2 slots. So, if we suppose to compute the efficiency over B “on-rounds”, as in the case of RH, we obtain

and having

(F LSA,BC) = Ton,transient + Ton Ton

(11)

(F LSA,BC) Toss = Ton,transient + Toff,transient + Ton + Toff

(12)

Ton,transient = Toff,transient = k Ton = B − k

B−k Toff = i=1 (k + i) = (B − k)k +

so that BC ηFLSA =

(13) (B−k)(B−k+1) 2

B 2 + B − k2 + k B 2 + 3B − k 2 + k

(14)

For what concerns the WC, FLSA achieves the 0% value since it represents the case in which the radio is on at every synchronization slot. The main advantage of FLSA over the modified version of RBS consists of the fact an initial set of time slots, k, is set to be the time needed for the slave to synchronize itself with the master. As soon as the steady condition is reached, the slave incrementally increases its stop time starting from those k slots. As already reported, note that FLSA operates with slots while the modified version of RBS operates with continuous time intervals. Also, FLSA switch on time can be reduced, since the slave is not on for the whole round duration. The steady-state behavior of RH

Wireless Sensor Networks Time Synchronization Algorithms

673

FLSA best case efficiency computation

100 95 90 85 80 75 70 65

k=1 k=5 k = 10 k = 20 k = 30

60 55 50 0

10

20

30

40

50

60

Fig. 5. FLSA best case efficiency computation for B ∈ [k, 60] varying the value of k ∈ N \ {0}

and FLSA is similar, but RH calls for a linear interpolation among the error samples each time it has to re-define the stop-time. FLSA in the best case, incrementally increases the skipped slots (in which the radio stays off) starting from a given number of slots value, and in case of failure, it restarts counting from 0. Hence FLSA is more severe with respect to RH in those situations in which an error occurs, but the overall computational complexity and efficiency are not impacted but mitigated. Note that the impact of k becomes negligible when a long observation period with strong synchronization is considered (Fig. 5).

5

Conclusions

After having introduced the main issues related to the implementation of a WSN from a time synchronization point of view, we analyzed different protocols and algorithms employed in the literature. We put a particular emphasis on FLSA, a protocol-agnostic time synchronization algorithm suitable for small WSNs with star topology. With the introduced metric for radio section power-efficiency comparison, we proved FLSA to be more efficient when compared to FTSP and RBS since it implements a routine that allows to switch off the radio section as soon as the device is capable to keep the synchronization error below a given threshold. When compared to RH, FLSA appears to be more rigorous on the synchronization recovery phase when huge errors occur while maintaining a restrained computational complexity.

References 1. Al-Shaikhi, A., Masoud, A.: Efficient, single hop time synchronization protocol for randomly connected WSNs. IEEE Wirel. Commun. Lett. 6(2), 170–173 (2017) 2. Casalino, G., Castellano, G., Zaza, G.: On the use of FIS inside a telehealth system for cardiovascular risk monitoring. In: 2021 29th Mediterranean Conference on Control and Automation (MED), pp. 173–178 (2021) 3. Coviello, G., Avitabile, G.: Multiple synchronized inertial measurement unit sensor boards platform for activity monitoring. IEEE Sens. J. 20(15), 8771–8777 (2020)

674

G. Coviello et al.

4. Coviello, G., Avitabile, G., Florio, A.: A synchronized multi-unit wireless platform for long-term activity monitoring. Electronics 9(7), 1118 (2020) 5. Coviello, G., Avitabile, G., Florio, A., Talarico, C., Roveda, J.: A novel low-power time synchronization algorithm based on a fractional approach for wireless body area networks. IEEE Access, p. 1 (2021) 6. Djenouri, D., Bagaa, M.: Synchronization protocols and implementation issues in wireless sensor networks: A review. IEEE Syst. J. 10(2), 617–627 (2016) 7. Elson, J., Girod, L., Estrin, D.: Fine-grained network time synchronization using reference broadcasts. ACM SIGOPS Oper. Syst. Rev. 36(SI), 147–163 (2002) 8. Facchini, F., Boenzi, F., Digiesi, S., Mossa, G., Mummolo, G.: Greening activities in warehouses: A model for identifying sustainable strategies in material handling, pp. 0980–0988 (2015) 9. Ganeriwal, S., Kumar, R., Srivastava, M.B.: Timing-sync protocol for sensor networks. In: Proceedings of the 1st International Conference on Embedded Networked Sensor Systems, pp. 138–149 (2003) 10. Kawagoe, H., Sugano, M.: Implementation of Time Synchronization for Energy Harvesting Wireless Sensor Network. In: Proceedings of the 2017 VI International Conference on Network, Communication and Computing - ICNCC 2017, pp. 175– 178. ACM Press, New York, USA (2017) 11. Kshemkalyani, A.D., Singhal, M.: Distributed Computing: Principles, Algorithms, and Systems, 1st edn. Cambridge University Press, USA (2008) 12. He, L., Kuo, G.-S.: A novel time synchronization scheme in wireless sensor networks. In: 2006 IEEE 63rd Vehicular Technology Conference, vol. 2, pp. 568–572 (2006) ´ The flooding time synchronization 13. Mar´ oti, M., Kusy, B., Simon, G., L´edeczi, A.: protocol. In: Proceedings of the 2nd International Conference on Embedded Networked Sensor Systems, pp. 39–49 (2004) 14. Mescia, L., Massaro, A.: New trends in energy harvesting from earth long-wave infrared emission. Advances in Materials Science and Engineering (2014) 15. Miller, B., Conley, R.: A multiple modulator fractional divider. IEEE Trans. Instrum. Meas. 40(3), 578–583 (1991) 16. Sundararaman, B., Buy, U., Kshemkalyani, A.D.: Clock synchronization for wireless sensor networks: a survey. Ad Hoc Netw. 3(3), 281–323 (2005) 17. Wu, X., Wang, Y., Wang, F.: A consistent and low-overhead time synchronization method for wireless sensor networks. In: 2015 Fifth International Conference on Instrumentation and Measurement, Computer, Communication and Control (IMCCC), pp. 1917–1921 (2015) 18. Wu, Y., Chaudhari, Q., Serpedin, E.: Clock synchronization of wireless sensor networks. IEEE Signal Process. Mag. 28(1), 124–138 (2011) 19. Yang, Z., He, L., Cai, L., Pan, J.: Temperature-assisted clock synchronization and self-calibration for sensor networks. IEEE Trans. Wireless Commun. 13(6), 3419– 3429 (2014)

Trust Management Model in IoT: A Comprehensive Survey Muhammad Saeed1 , Muhammad Aftab1 , Rashid Amin1(B) , and Deepika Koundal2 1 University of Engineering and Technology Taxila, Taxila, Pakistan

[email protected] 2 University of Petroleum and Energy Studies, Dehradun, India

Abstract. The Internet of Things (IoT) is a network of interlinked objects, without the need for human involvement these interlinked objects that can gather and transmit data via a wireless network. IoT deals with sensors with the support of the Internet. During the transmission process, it is necessary how to secure the data. When devices are connected to the Internet, there are certain chances of attacks on these devices. We need some security mechanisms for secure communication. The only way to protect the data from attacks is trust management. This paper explains the five dimensions of trust design models, such as adaptability, availability, integrity, privacy, reliability, accuracy, and scalability. Different attacks on IoT devices are discussed, along with their proposed solutions by research community. The focus of this paper is a literature survey on trust management model in IoT. Furthermore, we present the unsolved problems and discuss their effects on IoT models. Keywords: IoT · Trust management · Sensors · Holi trust · Networking

1 Introduction The Internet is classified into five categories. These categories are The Internet of Documents, The Internet of Commerce, The Internet of Applications, and The Internet of things. Internet of Things (IoT) was introduced in 1999 after the implementation of wireless concepts in various networking fields [1]. The Internet of Things is a combination of two words “Internet” and “Things.” It means an interconnected worldwide network based on sensors. The Internet of Things (IoT) is a network that collects and exchanges data from physical devices. These physical devices are embedded with software’s, and some sensors, that enable them to gather and transmit data efficiently. IoT improves its accuracy and efficiency by using a computer system for sensing and controlling the functionality of the devices. Those devices that are used in the IoT system are called smart devices. Carnegie Mellon University introduced the concept of smart devices in 1982. IoT is about performing certain functions without any human intervention. Today IoT is used in almost every industry like in automobile industries, transportations, and healthcare. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 675–684, 2022. https://doi.org/10.1007/978-3-030-96299-9_64

676

M. Saeed et al.

The Internet of things (IoT) concepts can be used for trust management. Trust Management provides security services in many applications, i.e., P2P, Grid, and Adhoc networks. Also, Trust Management provides many penalties in these applications. In reliable data fusion and mining, Trust Management is used and plays important roles. It is used to improve the security and privacy of users. For security, integrity, reliability, and expectations on the dependability, the concept of Trust Management is very difficult. In this paper, we discuss how to achieve a good trust level in IoT. The traditional internet architecture is not capable of handling large numbers of users/devices and cannot provide security on the huge numbers of nodes/devices efficiently. Internet of things is a type of Internet. There are many devices connected and communicated with each other [2]. Humans, monitors, laptops, smart devices, and sensors are all examples of devices used in the Internet of Things [3, 4]. Every physical node/object in the Internet of Things is directly connected to the Internet and can communicate with other nodes [5]. Some recent applications of IoT are smart cities, smart grid, and most important is E-health [3, 4]. The Internet of Things’ major goal is to improve the accuracy and efficiency of smart devices while also digitizing people’s lives. During communication in the Internet of Things, nodes encounter security challenges such as trust and access control. Node’s trust management for multiple communication devices is one of the growing trends in IoT device communication [2]. IoT does not support the traditional network or device’s security mechanism, so due to a huge number of interconnected devices, the traditional network security mechanism [6] cannot be directly implemented on the IoT. In the IoT environment, all devices work on user data according to user rights and needs [6]. During the communication and sharing of information among nodes, the affected nodes send the wrong recommendation. Communication through effected nodes means security and privacy are compromised by anyone [2]. The above mention challenges are addressed through the trust management mechanism. In trust management nodes can maintain trust and share information with affected nodes [2]. Before starting the communication among nodes, every node checks the trust toward others; if their trust value meets the condition, then it starts the communication, otherwise, nodes are not able to communicate with each other [3]. The rest of this article is organized as follows. Section 2 presents upcoming Internet and the concept of the Internet of Things, in the Sect. 3 the properties of trust and the objectives of trust is discussed, Section 4 discusses the Trust classification tree while Sect. 5 discusses trust management. Section 6 presents trust-related attacks, while Sect. 7 concludes the paper.

2 Properties and Objective of Trust Multiple fields use the concept of Trust with multiple definitions. Trust describes the actions of the IoT objects and also computes the trust value of a device or node. Trust management covers all properties of Trust that are mention above. In 2012 new proposed model was introduced based on IoT applications that are Dynamic trust management protocol. For managing the Trust of the nodes, every node in dynamic trust management acts as an independent node. Trust assessment in dynamic trust management is eventdriven and depends on some trust parameters [2]. Following parameters are used for providing a trustworthy relationship in IoT:

Trust Management Model in IoT: A Comprehensive Survey

677

3 Trust Classification Tree The trust classification tree is used to find the matrices and approaches to trust computation. It is also used to represent the difference between the approaches used for trust computation. Figure 2 shows the five basic components of Trust. Trust composition, trust propagation, trust aggregation, trust updating, and trust formation are some of the trust components. These five components are very necessary for any trust management model. Let us discuss all these components in detail. 3.1 Trust Composition Trust composition gives us information about the nodes that are used for trust computation like the QoS (Quality of services) and social Trust. QoS presents the surety of IoT devices that must be able to give quality service against the service request. In general, the QoS presents the enactment and is stately by skill. Cooperativeness and job end proficiency. QoS measures the point to point packet forwarding ratio, consumption of energy, and the ratio of the packet delivery [6]. Social Trust starts from the mutual connection between the vender of the IoT nodes and is to compute the informality, goodness, security, and connectivity. The concept of Social Trust is based on the Social IoT system environment, in which IoT nodes are not allowed to guess the QoS trust [7]. 3.2 Trust Propagation Trust propagation is a way for spreading trust among peers. Distributed and centralized propagation systems are the two basic types of propagation mechanisms. IoT nodes. automatically propagate the Trust to other nodes in Distributed Trust Propagation. They do not need any centralized entity for the propagation of Trust. Every node has a table which stores the data forwarding information. The social IoT environment supports the distributed trust environment [8]. In centralized trust propagation, the system needs a centralized object that gives service to all IoT nodes Y.B. Saied is commonly used as a centralized object to store trust-related data and to allow the iota nodes to answer requests. 3.3 Trust Aggregation The process of estimating the relationship between the variables is called aggregation analysis. The relationship between the Trust and the collection of factors can be estimated using aggregation analysis. Weighted sum, belief theory, and fuzzy logic are some of the techniques used. 3.3.1. Weighted Sum: A prominent technique called a weighted sum is utilized for Trust aggregate evidence. For acquiring a higher weight with a higher reputation, many authors employed the weighted sum technique. Martinez-Zulia and Skarmeta [9] based their credibility on quality of service and social trust. Credibility relates to the recommendation offered through indirect aggregation. Weight

678

M. Saeed et al.

associated. Chen et al. [8] also used social Trust to aggregate indirect weights. For the trust property, the weighted sum is designed to examine direct and indirect aggregation trust. This trust property holds true for both indirect and direct trust aggregates. 3.3.2. Belief Theory: Evidence theory, or Dumpster-Shaper Theory, is another name for belief theory. Belief theory is used to give uncertainty reasoning. The Dumpster-Shaper Theory is founded on two concepts: obtaining a level of belief and Dumpster’s rule. They are obtaining degree work for one question, which is taken from subjective probabilities. Dumpster rule is used for combining these belief degrees when these degrees are independent. The number of answers shows the proposition of the degree of the belief (these answers related to the questions). 3.3.3. Fuzzy Logic: Fuzzy logic is a form of multi-value logic. It works on the principles of fuzzy measurement. Fuzzy logic deals with reasoning slightly than static and particular. Fuzzy logic compares with set the range value of the degree is between 0 and 1. When the truth value becomes completely false or true, the fuzzy logic extends to handle the partial truth. 3.4 Trust Update The trust update component updates the Trust across the nodes. According to trust update, Trust has two schemes that are event and time driven scheme. 3.4.1. Event-driven Scheme: when events or transactions occur, then the Trust is updated. This type of scheme is called an Event-driven scheme. Recommendations are sent if the request is received. The nodes depend on each other and request for the recommendation of the other nodes. Trust evidence is computed time by time and updated accordingly. If no evidence encounters within the time delay, then overtime is frequently applied. 3.4.2. Time-driven Scheme: A trust aggregation technique is used to update trust, and evidence is collected on a regular basis in the time-driven scheme. Selfobservation or recommendations are used as evidence. When there is no evidence, trust is frequently applied. 3.5 Trust Formation Trust formation means how to create overall Trust in many trust properties. Trust formation has two aspects: a single Trust and multi Trust. 3.5.1. Single Trust: In a single trust, the trust rules only use one trust property. Because of the relationship between the service request node and the service, the service quality in social IoT is pairwise. Provider affect the service quality. In simple words, the relationship between nodes is one to one. 3.5.2. Multi Trust: In the multi Trust, Trust is multidimensional; for trust formation, multiple trust properties are used like informality, goodness, privacy, and connectivity, etc. Several trust qualities, such as closeness, honesty, unselfishness, and competence, are utilized to access a MANET node’s total Trust. For trust

Trust Management Model in IoT: A Comprehensive Survey

679

formation, certain ways are used. Without combining trust, properties used them individually and defined the minimum threshold for each trust property. Notice that these trust properties depend on the application requirements. For example, if the honesty is important, so the threshold value should be high. And for competence, it is not critical so that the threshold can below. Use the weighted sum of the overall trust metrics. The application fundamentals can reflect the weight assigned for honesty. High weight is used because honesty is important. The malicious attacks, including bad-mouthing and ballot-stuffing attacks, defended more effectively due to the high weight of honesty. On the other hand, high weight is set for competence when honesty is not necessary for a friendly environment. Chen et al. [8] proposed a model for maximizing average user satisfaction experience.

Fig. 2. Trust computation tree

4 Trust Management in IoT In our daily life, IoT devices are used to make our life easy [13]. The main properties of the IoT objects are located ability, reachability, controllability, and addressability [14]. IoT applications can be used in many situations. These applications are composed of sensors and processing devices [15]. In IoT, all devices are interconnected to each other [16]. The main parameters of the design of the IoT architecture are communications, security, and processes [17]. Exiting IoT techniques use some factors for trust management that are given below. 4.1 Accuracy: Accuracy is used to check the technique for the reputation and the Trust of the IoT. In other words, to check whether the technique is useful for IoT security or not [9]. 4.2 Adaptability: Adaptability defines whether the techniques being used for trust management can handle the dynamic structure of the network or not has IoT nodes are added or deleted rapidly. 4.3 Availability: Availability is used to validate whether network services are given, including attacks [10]. 4.4 Integrity: During the transformation of data between nodes, the responsibility of integrity is to protect and ensure the integrity of the data [11].

680

M. Saeed et al.

4.5 Privacy: During the information collection process for Trust, privacy is used to care about the private data of the user [12] 4.6 Reliability: Reliability is used in IoT applications to ensure the proper functionality of the network. It is an important slice of the trust partners by commercial rate [13].

5 Trust Related Attacks For itself, each IoT node [14] might be a service provider (SP) or a service requester (SR). As a result, IoT nodes are beneficial to provide effective services. And IoT devices need the best S.P.’s for providing the best quality of services when the S.R. is available. A malicious S.P. node acts and provides bad services. And malicious nodes provide services for their benefits. In this survey, we are concerned with trust-related attacks. These trust-related attacks interrupt the trust system and provide bad services. Most common trust-related attacks include bad-mouthing and ballot stuffing. Self-interest based attacks include self-promoting and opportunistic services attacks. To evade detection by the malicious attack are called on-off attacks. As a result, malicious IoT nodes can carry out the trust-related attacks listed below. 5.1 Self-Promotion Attacks (SPA): A malicious node gets importance over other nodes for the special recommendation for itself. So malicious nodes provide bad services. 5.2 Bad Mouthing Attacks (BMA): For providing a bad recommendation, the malicious node can run the Trust of a trusted node. So the decreased chance of select the services of these trusted nodes. 5.3 Ballot-Stuffing Attacks (BSA): By providing a good recommendation for it self, the malicious node boosts the trust of other malicious nodes. So, there is a chance for these malicious nodes selected to provide services. It invokes collision recommendation attacks. 5.4 Opportunistic Service Attacks (OSA): Due to good recommendations, the malicious node can gain a high reputation opportunistically. Due to good recommendations, these malicious nodes collaborate with other malicious nodes to make the bad-mouthing and ballot staffing attack. 5.5 On-Off Attacks (OOA): Due to on-off attacks, malicious nodes perform bad services, i.e., on and off. On and off services means a malicious node performs bad services randomly. Bad- Mouthing and ballot-stuffing attacks included in these services. Bad-mouthing and ballot-stuffing attacks are collaborative attacks that mean that malicious nodes focus on a particular part and target this particular part to make it malicious (Table 1).

Trust Management Model in IoT: A Comprehensive Survey

681

Table 1. Trust related attacks

6 Discussion Most of the research that is done in the area of IoT focuses on the improvement of the understanding of the IoT applications and up to what extent we trust those applications. IoT’s main concern is the security and reliability of the devices and system. In IoT, Trust among nodes is very important, especially in a wireless sensor network. IoT transmission is started when nodes trust each other; for this purpose, trust management is proposed by many researchers in recent years. The main objective of trust management is to build Trust among nodes for smooth and reliable communication between them. If any node is affected in the wireless sensor network by a malicious attack, then the performance of all nodes will be affected [7]. In the Internet of Things, the most important thing is to send data between nodes according to data communication laws’ attributes. The Internet of Things (IoT) ensures secure data exchange between devices by establishing a high level of trust among them. Data merging also relies heavily on IoT trust management [8]. For the exchange of secure data among IoT nodes, many approaches are proposed for trust management. Kamran et al. [9] introduced the Holi trust management system for cross-domain trust management. Because the domain is divided into a few communities, each of which has a limited number of servers, this method was implemented to improve domain security. Each community has three servers: domain servers, community servers, and trust servers. These servers work in parallel. Responsibility of the Domain server is to manage all communities of the domain. The community server is responsible for the management of the security of nodes during low trust levels. This technique is event-driven in this quality of technique to enhance the efficiency of the system. This technique was used to build Trust among the node based on previous direct observations, experience, and recommendation parameter. This approach work against the effected nodes. Fang et al. [10] examine cyber security in the ICN, as well as attack activities and defense strategies. For acute attacks, the author used a quick and efficient trust management mechanism. The acute attack comes and goes. In a short amount of time, this system can detect and neutralize assaults. Conti et al. [3] proposed a technique to solve the discussed problem. They proposed a methodology called the Dirichlet-Distribution-Based Trust Management Scheme (DDTMS). This technique was based on three categories. First, the basis of Dirichlet distribution, the interaction process designed for the reputation distribution of the node. Secondly, calculate the trust value of the nodes due to

682

M. Saeed et al.

joint historical interactive information and custom parameters. Finally, the overall results were obtained on the third-party recommendation. And custom parameters. Finally, the overall results were obtained on the third-party recommendation. Amazon [5] developed a cloud platform Amazon Web Services (AWS) in IoT. Anyone can connect smart devices with Amazon easily. AWS provides some other services. These services include Amazon Dynamo (D.B.), Amazon S3, Amazon Machine Learning, etc. we utilize these services easily by AWS (Table 2). Table 2. Articles comparison

A. Meena et al. [12] worked on trustworthy decision making in IoT. For trust computation, they used trust matrices such as community trust, direct Trust, centrality, and cooperativeness. Rupayan Das et al. [13] presents a technique based on four ideas, i.e., Self-trust, green Trust, social Trust, and QoS trust. Self-trust are computed in node through the following methods. Internet of things data processing trust internet of things data privacy trust, and Internet of things data transmission trust, green trust concern with the characters of the network. Two main characteristics of green Trust are lifetime trust and response trust. Social trust responsibility to point out the performance of the device in the environment. The trust checks on the node through QoS trust. All of the papers listed in the discussion section were trust computations based on time-driven and event-driven trust assessment. However, the hybrid trust assessment technique was not mentioned in any of the articles. Second, all of the articles attempted to safeguard communication in this regard. All of the researchers were focused on device and domain level security, but none of them were concerned with trust tempering prevention. 6.1 Future Directions • Context awareness and the subjective attributes of the user (trustor) are not taken into account when evaluating trust. Because the trust rating result is not individualized, it

Trust Management Model in IoT: A Comprehensive Survey

• • •

•

683

is difficult to provide intelligent IoT services. Services that are “only here, only now, and only me” are still a work in progress. The research of SMC is still in its infancy. The majority of solutions are impractical in terms of processing complexity, communication costs, adaptability, generality, and integrity, making them difficult to implement. HCTI is practically ignored in present research, yet it is among the most important factors influencing IoT’s eventual success. The eventual success of IoT is determined by holistic trust management along with a positive user experience. Although DPT solutions based on trusted computing platforms have been developed, but they may be too large for wireless sensors. with limited capabilities to employ. Lightweight trust and security techniques for small items inside the Internet of Things must be developed, with a particular emphasis on mitigating DoS and DDoS attacks. Although the interaction of DTCT techniques with other trust management approaches has not been studied, it can be a future research area for achieving vertical trust management goals in a variety of scenarios.

7 Conclusion In this survey, we highlight the significance of the IoT. We discussed trust management and trust management properties and the relationship of these properties. We reviewed some previous techniques and matched the adaptability of the trust management to search the key issues and explain challenges. We also pointed out the importance of trust management in IoT. We designed a classification tree in this survey according to five dimensions. These dimensions included trust composition, trust propagation, trust aggregation, trust formation, and trust update. We also discussed accuracy, adaptability, availability, integrity, privacy, reliability, and scalability. These factors are used to compute trust value and then distribute it to other IoT devices that are connected to the network. We discussed trust-related attacks for explaining how malicious nodes can effect on IoT devices and make these devices malicious.

References 1. Sfar, A.R., et al.: A roadmap for security challenges in the internet of things. Digital Commun. Netw. 4(2), 118–137 (2018) 2. Awan, K.A., et al.: Robusttrust–a pro-privacy robust distributed trust management mechanism for the internet of things. IEEE Access 7, 62095–62106 (2019) 3. Conti, M., et al.: Internet of Things Security and Forensics: Challenges and Opportunities. Elsevier (2018) 4. Din, I.U., et al.: The internet of things: A review of enabled technologies and future challenges. IEEE Access 7, 7606–7640 (2018) 5. Ammar, M., Russello, G., Crispo, B.: Internet of things: A survey on the security of IoT frameworks. J. Inf. Secur. Appl. 38, 8–27 (2018) 6. Alhogail, A., Alshahrani, M.: Building consumer trust to improve internet of things (IoT) technology adoption. In: International Conference on Applied Human Factors and Ergonomics. Springer (2018). https://doi.org/10.1007/978-3-319-94866-9_33

684

M. Saeed et al.

7. Chang, K.-D., Chen, J.-L.: A survey of trust management in WSNS, internet of things and future internet. KSII Trans. Internet Inf. Syst. 6(1) (2012). https://doi.org/10.3837/tiis.2012. 01.001 8. Zhang, Q., et al.: An adaptive dropout deep computation model for industrial IoT big data learning with crowdsourcing to cloud computing. IEEE Trans. Ind. Inf. 15(4), 2330–2337 (2018) 9. Awan, K.A., et al.: Holitrust-a holistic cross-domain trust management mechanism for servicecentric internet of things. IEEE Access 7, 52191–52201 (2019) 10. Fang, W., et al.: FETMS: Fast and efficient trust management scheme for information-centric networking in internet of things. IEEE Access 7, 13476–13485 (2019) 11. Alshehri, M.D., Hussain, F.K., Hussain, O.K.: Clustering-driven intelligent trust management methodology for the internet of things (CITM-IOT). Mobile Netw. Appl. 23(3), 419–431 (2018) 12. Kowshalya, A.M., Valarmathi, M.: Trust management for reliable decision making among social objects in the social internet of things. IET Netw. 6(4), 75–80 (2017) 13. Das, R., Singh, M., Majumder, K.: SGSQOT: A community-based trust management scheme in internet of things. In: Proceedings of International Ethical Hacking Conference 2018. Springer (2019). https://doi.org/10.1007/978-981-13-1544-2_18 14. Alshehri, M.D., Hussain, F.K.: A centralized trust management mechanism for the internet of things (CTM-IOT). In: International Conference on Broadband and Wireless Computing, Communication and Applications. Springer (2017). https://doi.org/10.1007/978-3-31969811-3_48 15. Fernandez-gago, C., Moyano, F., Lopez, J.: Modelling trust dynamics in the internet of things. Inf. Sci. 396, 72–82 (2017) 16. Hui, T.K., Sherratt, R.S., Sánchez, D.D.: Major requirements for building smart homes in smart cities based on internet of things technologies. Future Gener. Comput. Syst. 76, 358–369 (2017) 17. Valera, A.J.J., Zamora, M.A., Skarmeta, A.F.: An architecture based on internet of things to support mobility and security in medical environments. In 2010 7th IEEE Consumer Communications and Networking Conference. IEEE (2010)

Sentiment Classification and Comparison of Covid-19 Tweets During the First Wave and the Second Wave Using NLP Techniques and Libraries Sareeta Mugde , Garima Sharma , Aditya Singh Kashyap(B) and Swastika Swastik

,

Welingkar Institute of Management Development and Research, Mumbai, India

Abstract. This research focuses on analysing the sentiments of people pertaining to severe periodic outbreaks of COVID-19 on two junctures – First Wave (Mar’20 & Apr’20) and Second Wave (Jun’21 & Jul’21)-since the first lockdown was undertaken with a view to curb the vicious spread of the lethal SARS-Cov-2 strain. Primarily, the objective is to analyse the public sentiment – as evident in the posted tweets - relating to the different phases of the pandemic, and to illuminate how keeping an eye on change in the tenor and tone of discussions can help government authorities and healthcare industry in raising awareness, reducing panic amongst citizens, and planning strategies to tackle the monumental crisis. Considering the daily volume of social media activity, in our project, we scoped to analyse the Tweets related to the two different pandemic stages – The First wave and the Second wave – by implementing Text Mining and Sentiment Analysis, subfields of Natural Language Processing. To manually extract tweets from the platform, we used Twitter API coupled with Python’s open-source package using a set of COVID-19-related keywords. Crucially, before finalising the project pipeline, we conducted a thorough secondary research to find the solutions and methodologies implemented in our area of interest. We listed the current works and attempted to plug the gaps in those via our experiment. We used several classification and boosting algorithms to create a framework to distinguish the textual data of the tweets. Relevant scope, limitations, and room for improvements have been discussed comprehensively in the upcoming sections. Keywords: Text mining · Sentiment classification · COVID 19 · Natural language processing · Tweet analysis · Tweepy · TextBlob

1 Introduction During a crisis, whether natural or man-made, people tend to spend relatively more time on social media than the normal. As crisis unfolds, social media platforms such as Facebook and Twitter become a dynamic source of information because these platforms break the news speedier than official news channels and emergency response agencies. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 685–699, 2022. https://doi.org/10.1007/978-3-030-96299-9_65

686

S. Mugde et al.

During such events, people usually make informal conversations by sharing their safety status, querying about their loved ones’ safety status, and reporting ground level scenarios of the event or crisis they find themselves in. This process of continuous creation of conversations on such public platforms leads to accumulating of a large amount of socially generated data. The amount of data can range from hundreds of thousands to millions. In recent times, the most used social media platforms for informal communications have been Facebook, Twitter, Reddit, etc. Amongst these, Twitter, the microblogging platform, has a well-documented Application Programming Interface (API) for accessing the data (tweets) available on its platform. Therefore, it has become a primary source of information for researchers working on the Social Computing domain. With our research, we analysed the collection of tweets during the two phases of Coronavirus waves, inducing partial to strict lockdown and other set of restrictions. Our work required around 50,000 tweets, collected over the gap of 10 months. To understand the patterns and relationships from the textual data, we implemented a few Natural Language Processing (NLP) and Text Mining techniques to perform Sentiment Analysis on the Twitter data. Sentiment Analysis is the process of identifying the sentiment, emotion, or the intention behind the textual communication. This is done via contextual mining of text, thereby extracting subjective information on the source to identify crucial patterns, relationships, and hidden chest of information useful in aiding other objectives. 1.1 Literature Review The literature review for our research spanned from 2012 to 2021, comprising of various Conference Papers, Journal Research Papers, Reports, documentations, and several other open-source resources like Kaggle, TextBlob, Tweepy, and NLTK [4, 18, 19, 25]. Additionally, we shortlisted around 20 papers from databases like IEEE, Springer, and Elsevier which aligned with the approach that we wanted to implement, as well as the available solutions in the domain of our problem statement. From our secondary research, we got to understand the various techniques and approaches followed by authors to analyse the impacts, changes, and various other factors in their own unique ways, with Kaur and Sharma [7] and Khan, Rustam, Kanwal, Mehmood and Choi [26] using TextBlob for Sentiment analysis. A plethora of methodologies – technical flow and ideas – were implemented to express their work covering various aspects of the pandemic, as Kaur, Ahassan, Alankar, and Chang [6] used deep learning to conduct Sentiment Analysis and Staszkiewicz, P.; Chomiak-Orsa [13] analysing the Contagion and Mortality owing to the virus. Through our experiments and stepwise implementations, we aimed to plug the gaps in the current work and append a new ideology in the COVID-19 research paradigm. From the secondary research of pandemic-related literature - Chintalapudi, Battineni, and Amenta (2021) [3], Alamoodi, Zaidan, Zaidan, Albahri, Mohammed, Malik, and Hameed, (2020) [23], and technical framework to support those ideas with tangible outcomes, we added a few more implementations and broadened the scope of our work. As available works - Iyer and Kumaresh (2020) [5], Raheja and Asthana (2021) [27], and Ridhwan, Hargreaves (2021) [28] – predominantly cover single wave of coronavirus

Sentiment Classification and Comparison of Covid-19

687

or vaccine related public sentiment, our project targeted two pandemic waves where we analysed how sentiments changed as per the discussions with growing time –as evidenced in the Second wave with the views of public revolving around going back to normal and getting vaccinated when compared to a year before where discussion focussed on sanitisers and staying home. In terms of technical implementation, majority of works used VADER, BERT models – as visible in works by Chintalapudi, Battineni, and Amenta (2021) [3], and Ridhwan, Hargreaves (2021) [28], and other deep-learning models used by Lamsal (2021) [20]. Through our project, we added Boosting algorithms to analyse the efficiency and outcomes in relatively undiscovered ground. The focus was also to keep the computation time on an optimal level; thereby we tried to tap into a different area of machine learning to study our subject of research. Further literature review unearthed a research void in terms of non-availability of Sentiment analysis studies in global context. 1.2 SCOPE The project focusses on tweets relating to COVID 19 posted on the microblogging site during the First and Second Wave of the Pandemic outbreak. Posts on other microblogging sites are beyond the ambit of the study. The study assumes significance on ground of its relevance and use across domains such as Customer Service, Risk Management and Spam Filtering. Customer Service – From brand’s perspective, surgery of customer’s feedback could provide key aspects about a product which customer likes or where pain points appear. The intentions and tone of the comments could also form the basis of several other takeaways. Risk Management – NLP can be used to derive the insights on industry trends and financial markets by monitoring shifts in sentiment and by extracting information from analyst reports and whitepapers. Such techniques could be handy for Banking industries as this data and reports could provide more confidence while considering or initiating investments across sectors. Spam Filtering – Text mining techniques can create a filter which protects the infrastructure from suspicious e-mails, and reducing cyber-attack on social forums, thereby providing an improved user experience. Using the COVID-19 related tweets can help in better understanding of the mass sentiment regarding the pandemic, public’s needs and demands during the restrictions. The NLP analysis can assist in knowing various trends around the most important aspect, Vaccination. A deeper understanding of the coronavirus related tweets and social posts can provide a supplementary support on insights as for Marketing and advertising teams to structure their campaigns and strategies as per the patterns and trends emerging from the data post analysis. It can also help government institutions to gauge the issues and needs during the pandemic and address the shortcomings in their operations. The Twitter sentiment analysis

688

S. Mugde et al.

can play a huge role in behavioural finance, and it can be a credible source for skimming information about the market. 1.2.1 Modus Operandi To gain a preliminary understanding of the type of data in our hands, we have presented snippets of the tweets related to the two waves of the pandemic. For analysing the context, we have run a Python code to generate the respective Polarity, Subjectivity, and Sentiment of the sample tweets from our dataset. Further breaking down of Sentiment, and the explanation of Polarity and Subjectivity lies ahead in form of detailed discussions and explanations. Below, the two snapshots present the tables containing the sample tweets from our datasets representing the two phases of COVID-19 (Figs. 1 and 2).

Fig. 1. First wave

Fig. 2. Second wave

2 Proposed Model 2.1 Data Collection For our initial analysis, we retrieved the dataset from Kaggle, containing 48,753 entries of Tweets, Location, Date, ScreenName, and Sentiment – which was the dependent variable, where only the OriginalTweet column forms the genesis of our Sentiment Analysis using Natural Language Processing techniques.

Sentiment Classification and Comparison of Covid-19

689

The dataset for second wave was generated manually as the COVID-19 related tweets were extracted using Vaccination and other pandemic related keywords. Python’s Tweepy package came to the fore as we established a connection with relevant Twitter API codes. The Tweet Listener function was left idle for 10 h to allow us with generous amount of textual data to continue our analysis. Citing memory constraints, the dataset was trimmed and only 18,798 randomly chosen Tweets were extracted for the Sentiment Analysis for the first wave dataset. Manually extracted dataset was pruned to 15,000 rows, consisting of variables like Polarity, Subjectivity, Text, timestamp, and Sentiment – Positive, Neutral, and Negative. Assigning sentiment was done by categorising Polarity range. The Polarity and Subjectivity was calculated using the relevant methods of the TextBlob library present in Python. For our complete execution of project, Google Collaboratory was used as the IDE. The TextBlob package contains an array of functions to conduct various fundamental Natural Language Processing jobs. Suiting the objectives of our project, we derived two pronounced factors, Sentiment and Polarity. Polarity is derived from textual data using dot operator to invoke the sentiment class. For each textual data the output can vary from −1 to +1. −1 defines negative, whereas +1 denotes positive, and 0 shows neutral sentiment. The lower the polarity score, the higher the pessimism in the textual data and vice-versa. Likewise, Subjectivity is derived from textual data by invoking the subjectivity class with the help of dot operator. It measures the share of information or opinion presented via the selected piece of data. Higher score means that the text contains strong opinion rather than actual facts. Gaining Polarity and Subjectivity scores aided us profoundly during the creation of dataset representing the second wave. To implement a classification model, we formed three bounds to assign three required sentiment categories – Positive, Negative, and Neutral. Polarity score between −1 and 0, 0, and 0 and 1 were categorised as Negative, Neutral, and Positive, respectively (Fig. 3). 2.2 Data Wrangling Before proceeding towards the model, the dependent variable, Sentiment, underwent feature engineering to reduce the complexity of our model. Initially, the dependent variable comprised of five sentiments – Extremely Positive, Positive, Neutral, Negative, and Extremely Negative. The distribution of dataset according to initial sentiments is depicted in the below plot (Fig. 4). To minimise the categories, a method was designed to merge the Extremely Positive and Positive to only Positive and Extremely Negative and Negative to only Negative, respectively. The Neutral category was exempt of changes. It was a twopronged approach, where 5 categories were encoded into numbers using Python’s category_encoders package, followed by the application of a function to merge categories using Panda’s apply() function with the defined method being passed as an argument. Updated sentiment distribution is displayed in below in Seaborn’s countplot() (Fig. 5).

690

S. Mugde et al.

Fig. 3. Project framework

Fig. 4. Tweet distribution sentiment wise

For self-collated dataset using Tweepy, the Sentiment categories were kept to three – Positive, Negative, and Neutral. This was done by segmenting the polarity scores and assigning them respective sentiment. The Polarity scores were derived using TextBlob package in Python (Fig. 6).

Sentiment Classification and Comparison of Covid-19

691

Fig. 5. First wave sentiment distribution

Fig. 6. Second wave sentiment distribution

Starting the data cleaning phase, the Tweet texts are imported using the ISO-8859-1 encoding – which contains 191 additional characters from the Latin script. It is better than UTF-8 as it fails to generate relevant ASCII code when it encounters the Latin Words, leading to inefficient Binary parsing. Data Wrangling was concluded when Positive, Neutral, and Negative categories were assigned integers 1, 2, and 3, respectively, to ease the process of model formation. To prepare our dataset for Sentiment Analysis, the Data Frame is checked for any missing values in the Sentiment Column or Tweet Column. As it is found out that Location attribute has a few Null entries, no changes are made since the Origin does not play a part in the Language Processing of the Tweets. 2.3 Bag of Words In natural language processing, a common technique for extracting features from text is to place all the words that occur in the text in a bucket. This approach is called a ‘Bag

692

S. Mugde et al.

of Words’ model or BoW for short. It’s referred to as a “bag” of words because any information about the structure of the sentence is lost. In order to extract useful texts from our collection of Tweets, Bag of Words (BoW) is designed to collect all the essential words from our dataset and transform them into numbers by creating a Sparse Matrix. Since our Machine Learning models can be trained only on numbers, it is imperative to sparse important words into numbers by comparing them to a list of pre-defined significant texts. Python provides number of packages to ease the process of extracting and appending words into the BoW model. Regular Expression – ˆa-zA-Z – initially, is used to replace non-letter characters by spaces – which also forms a part of further data cleaning procedure. Furthermore, the text data is imparted uniformity by converting them into lower case. Then, tweets are split into individual words and then those words are stemmed into their root form. Lastly, the processed collection of words from the tweet is appended as a list, which is then transformed into a Sparse Matrix. To attain the Sparse Matrix for the First and Second wave Tweets, CountVectorizer() and TfidifVectorizer() methods are implemented, respectively. CountVectorizer() converts a collection of text documents to a matrix of token counts: the occurrences of tokens in each document. This implementation produces a sparse representation of the counts. Term Frequency stands for the number of times a word appears in a document divded by the total number of words in the document. Every document has its own term frequency. Inverse data frequency determines the weight of rare words across all documents in the corpus. In a bid to explore multiple avenues, we decided to implement CountVectorizer and TfidfVectorizer on the corpus of first wave and second wave, respectively. 2.4 Classification Models 2.4.1 Random Forest Classifier It is an ensemble decision tree-based learning algorithm. This classifier is a set of decision trees randomly seeded from the training dataset. Random Forest uses bagging and feature randomness to create uncorrelated decision trees leading to a Forest formation. Random Forest shows higher accuracy than decision tree algorithm, as it cumulates votes from each decision tree to return final class of the test object. Python’s Scikit library contains relevant method to implement this classification model (Fig. 7). 2.4.2 Support Vector Machines (SVM) Classifier SVM is a supervised machine learning algorithm which can perform regression and classification. To perform the latter, all the data points are plotted on an n-dimensional space, and a decision line, known as hyperplane, separates the classes. The data points on either side of the hyperplane that are closest to the hyperplane are called Support Vectors which is used to plot the boundary line. In SVM Classification, the data can be either linear or non-linear. There are different kernels that can be set in

Sentiment Classification and Comparison of Covid-19

693

Fig. 7. Typical random forest

an SVM Classifier. For a linear dataset, we can set the kernel as ‘linear’. The non-linear datasets have kernel as ‘rbf’ and ‘polynomial’. Since it is complex to draw hyperplane on a 2D plane, the decision line is plotted in a 3D plane and then simplified to a lower dimension. 2.4.3 Naïve Bayes Classifier The Naïve Bayes Classifier is a probabilistic type of classification algorithm-due to this feature, it differs from the other techniques implemented in our project.Here, we have implemented Gaussian, Multinomial, Bernoulli, and Complement Naïve Bayes Classifier. The assumptions of each type of NB remains same, only being distinguished on the distribution of our data. The classification is done with a posterior probability solution. P(A|B) =

P(B|A) ∗ P(A) P(B)

(1)

P(A|B): The Conditional Probability that event A occurs, given that event B has already occurred. P(B|A): The Conditional Probability that event B occurs, given that event A has already occurred. P(A), P(B): Probability of event A and B occurring irrespective of each other. To fit the purpose of Machine Learning algorithms, A is considered as the measured variable. Whereas B is assumed to be the explanatory variable. Naïve Bayes takes the final form of: posterior =

prior ∗ likelihood evidence

(2)

P(A|B): The Conditional Probability that measured variable belongs to a particular value, given the explanatory attributes. It is also called Posterior Probability.

694

S. Mugde et al.

P(B|A): The likelihood of training data; P(A): The prior probability of measured variable; P(B): The probability of training data or the evidence. 2.4.4 Boosting Algorithms After having executed a range of classification algorithms, we decided to boost the results using boosting algorithms. Following three techniques were applied to our dataset. a. Adaptive Boosting Adaptive Boost (AdaBoost) is an ensemble learning technique where weights are reassigned at each iteration, with higher on the incorrectly classified instances. This technique can tweak the classification as well as the regression algorithms. In the project, we boosted a decision tree classifier to generate a new classification model. b. Gradient Boosting In the gradient boosting algorithm, we train multiple models sequentially, and for each new model, the model gradually minimizes the loss function using the Gradient Descent method. The Gradient Tree Boosting algorithm takes decision trees as the weak leaners because the nodes in a decision tree consider a different branch of features for selecting the best split, which means all the trees are not the same. c. Extreme Gradient Boosting (XGBoost) Unlike many other algorithms, XGBoost is an ensemble learning algorithm meaning that it combines the results of many models, called base learners to make a prediction. Just like in Random Forests and AdaBoost, XGBoost uses Decision Trees as base learners. XGBoost is known for its efficiency in scaled results (Table 1). Table 1. First and second wave accuracy summary First wave model summary

Second wave model summary

Model

Training Test set Model set

Training Test set set

Gaussian Naive Bayes

68.55%

40.02% Gaussian Naive Bayes

72.49%

57.89%

Multinomial Naive Bayes

79.00%

66.03% Multinomial Naive Bayes

86.82%

79.83%

Bernoulli Naive Bayes

81.92%

68.54% Bernoulli Naive Bayes

87.76%

80.42%

Complement Naive Bayes 82.71%

67.29% Complement Naive Bayes 87.66%

79.85%

Random Forest Classifier

97.08%

73.67% Random Forest Classifier

97.38%

88.16%

SVM Kernel

96.47%

75.37% SVM Kernel

95.64%

87.81%

Adaptive Boosting (AdaBoost)

70.58%

67.78% Adaptive Boosting (AdaBoost)

78.51%

76.90%

Gradient Boosting

77.62%

75.27% Gradient Boosting

87.00%

83.76%

XGBoost Classifier

63.09%

60.65% XGBoost Classifer

74.39%

73.24%

Sentiment Classification and Comparison of Covid-19

695

3 Results and Discussions 3.1 Word Cloud Analysis The combined Word Cloud for both waves paints a lucid picture of the sentiments and issues which were discussed heavily on the social media. It also captures certain moments which caused a stir on the micro-blogging site (Fig. 8).

Fig. 8. First wave (L) vs Second wave (R)

During the initial lockdown and the first wave, terms like pandemic, coronavirus, Covid-19, groceries, store, supermarket, highlight the theme of discussions hovering on the social space as people stacked on groceries, and talked about the global crisis as whole, even while spreading news, remedies, or even a candid discussion. In the midst of dense words, several terms like thank, food stock, panic buy, hand sanitiser, and government caught the essence of global pandemic. The frontline workers and medical staff were expressed gratitude in those tough times, and people, meanwhile, panicked for food supplies and hand sanitisers, where government was deemed responsible for the situation. Similarly, for the second wave where the paradigm shifted towards an effort to gain the normalcy back and expedite the mass vaccination program to cease the vicious spread of the deadly coronavirus. Particularly during the timescale of the collection of tweets for our dataset pertaining to the second wave, the Matt Hancock scandal was the live wire on Twitter, which is clearly evident in our above heatmap and ensuing sentiment-wise word clouds, too. The way his scandal unfolded was via email and included his office, which is thus legible in our 2nd Wave’s Wordcloud. Also, Matt Hancock was the Health Secretary of the Great Britain, which drew further wrath amid the ghastly second wave. Words like pandemic, wear mask also shows minimal prominence on the heatmap, despite it holding much higher significance than the gigantic terms floating on the textual dashboard.

696

S. Mugde et al.

This is done to signify the way massive trends or, for that matter, even single event can dominate the larger or elongated trend, such as the pandemic. Several short events can overpower the diffused trend, but the intensity remains inversely proportional to the time that theme persists on the social space. It can be a useful insight for concerned factions planning to hop on the trend and analyse the intensity to place their strategies or products accordingly in the digital space. The above three word-clouds provide sentiment wise textual heat map. As per our data pre-processing module, the 5 sentiments were compressed into three categories1: Positive, 2: Neutral, and 3: Negative. Supermarket is a prominent presence in all the three clouds. However, the negative heat map also sees the prominence of terms like panic buying, pandemic (it is used across all sentiments), crisis. The pattern clearly explains the sentiment of users tweeting about the pandemic-induced lockdown and Covid-19 in broader perspective. As discussed during the collective tweet collection, Matt Hancock and his controversy did the rounds when the second COVID-19 wreaked havoc in the United Kingdom. One takeaway from this is that regardless of doing sentiment-wise Wordcloud analysis, Matt Hancock was part of Neutral and Negative tweets. So breaking down the insights can unearth some fresh insights, which might play a part in concerned authorities taking the aid of NLP techniques. Moreover, terms like delta variant, variant, serious, limited dose, vaccination were, undeniably, direct towards the deadly delta variant and stymied jabbing at various parts of the world. Positively, vaccination drive, fully vaccinated, and India’s Vaccination drove optimistic tweets people were upbeat about achieving the herd immunity and steadily surging vaccination volume, respectively. Neutrally, Vaccination and Government set the tone for informational talk where people discuss about the prospects of attaining the biggest weapon to fight the pandemic (Fig. 9).

Fig. 9. Evaluation for classification models

Sentiment Classification and Comparison of Covid-19

697

4 Conclusion Through our array of implementations, we were able to perform sentiment analysis on the pair of datasets comprising of coronavirus-related tweets – or during the varying timescale of the two waves of the pandemic-catching the essence of Twitter via what users write and express. We were able to draw some incisive insights from our implementation, which provided policy implications and probable implication in the industry. It gave us a platform to explain the scope of our project as to how it can help in deriving strategies and decisions related to the field of the stock market, finance, marketing, etc. A combined heat map for the first and the second wave was pictured, and it depicted the pattern, explaining the common words featuring across majority of Tweets, irrespective of the sentiment. Multiple word clouds were drawn to define the pattern of words and terminologies used for Negative, Positive, or Neutral expressions, thereby underscoring the importance of sentiment-wise heatmap to gauge the various terms dominant across the emotions or specific domain. This Natural Language Processing technique and methodology can be brought into play by various management and operational paradigms to sharpen the proceedings as the impact of emotional and sentimental factor has been discussed vividly in our previous sections. Depending upon the distribution of the data and the Bag of Words model used to form the Sparse Matrix. The accuracy attained on multiple parameters ranged between 40% to 88% on the test set and 63% to 97% on the training set. Python’s CountVectorizer() method is used to create a parse matrix to train our classification model. We used stopwords package of Natural Language toolkit. These words form a relative noise in our training data, thereby affecting the accuracy of our trained algorithm. On a few occasions, removing the stopwords completely diffuses the essence of our text. For enhancing our Naïve Bayes model, we need to make our Bag of Words even more precise. Tackling this issue, Term Frequency is used to depict the relative importance of each given word. Inverse Data Frequency gauges the weight of rarely used words across the corpus. In order to enchance the accuracy of our four NB classifers, TfidfVectorizer, present in sklearn package of Python, can be used to compile the Bag of Words in the TF-IDF format. Consistently retrieved data corpus provides for the better analysis and more accurate results, as evident in manually formed Tweets dataset for the second wave. Already present dataset on COVID-19’s first wave hints toward it being collected for other objectives using other keywords. To dabble with accuracy, we can use other metrics, Precision, Recall, and F1-Score to better judge the models. Also, boosting algorithms could be hyper tuned and imbalanced datasets could be treated to make it consistent. Our project clearly depicted its scope and its wide array of applications. The wideranging accuracy can be brought into control via several machine learning algorithms.

698

S. Mugde et al.

References 1. Vijay, T., Chawla, A., Dhanka, B., Karmakar, P.: Sentiment analysis on COVID-19 Twitter data. In: IEEE International Conference on Recent Advances and Innovations in Engineering (2020) 2. Saranya, G., Geetha, G., Chakrapani, K., Meenakshi, K., Karpagaselvi, S.: Sentiment analysis of healthcare Tweets using SVM classifier. In: 2020 2nd International Conference on Power, Energy, Control and Transmission Systems, IEEE (2020) 3. Chintalapudi, N., Battineni, G., Amenta, F.: Sentimental Analysis of COVID-19 Tweets Using Deep Learning Models. Inf. Dis. Rep. 13(2), 329–339 (2021) (Multidisciplinary Digital Publishing Institute (MDPI)) 4. Coronavirus Tweets NLP – Text Classification, Kaggle. https://www.kaggle.com/datatattle/ covid-19-nlp-text-classification 5. Priya Iyer, K.B., Kumaresh, S.: Twitter sentiment analysis on coronavirus outbreak using machine learning algorithms. Eur. J. Molecul. Clin. Med. 07(03), 2663 (2020) 6. Kaur, H., Ul Ahassan, S., Alankar, B., Chang, V.: A proposed sentiment analysis deep learning algorithm for analyzing COVID-19 Tweets. Inf. Syst. Front. 23(6), 1417–1429 (2021) 7. Kaur, C., Sharma, A.: Twitter Sentiment Analysis on Coronavirus using Textblob. EasyChair 2516–2314 (2020) 8. Ezhilan, A., Rakshana, B.S., Dheekksha, R., Anahitaa, R., Shivani, R.: Sentiment analysis and classification of COVID-19 tweets. In: 2021 5th International Conference on Trends in Electronics and Informatics (ICOEI), pp. 821–828 (2021). https://doi.org/10.1109/ICOEI5 1242.2021.9453062 9. Kariya, C., Khodke, P.: Twitter sentiment analysis. Int. Conf. Emerg. Technol. 2020, 1–3 (2020). https://doi.org/10.1109/INCET49848.2020.9154143 10. Kale, S., Padmadas, V.: Sentiment analysis of Tweets using semantic analysis. In: 2017 International Conference on Computing, Communication, Control and Automation (ICCUBEA), pp. 1–3 (2017). https://doi.org/10.1109/ICCUBEA.2017.8464011 11. Rolling updates on coronavirus disease, World Health Organization (WHO). https://www. who.int/emergencies/diseases/novel-coronavirus-2019/events-as-they-happen 12. COVID-19 Portal, World Health Organization (WHO). https://www.who.int/emergencies/dis eases/novel-coronavirus-2019 13. Staszkiewicz, P., Chomiak-Orsa, I.: Dynamics of the COVID-19 contagion and mortality: country factors, social media, and market response evidence from a global panel analysis. IEEE Access 8, 106009–106022 (2020) 14. Saif, H., He, Y., Alani, H.: Semantic sentiment analysis of Twitter. In: Cudré-Mauroux, P., et al. (eds.) ISWC 2012. LNCS, vol. 7649, pp. 508–524. Springer, Heidelberg (2012). https:// doi.org/10.1007/978-3-642-35176-1_32 15. El Rahman, S.A., AlOtaibi, F.A., AlShehri, W.A.: Sentiment analysis of Twitter data. Int. Conf. Comput. Inf. Sci. 2019, 1–4 (2019). https://doi.org/10.1109/ICCISci.2019.8716464 16. Drus, Z., Khalid, H.: Sentiment analysis in social media and its application: systematic literature review. Proc. Comput. Sci. 161, 707–714 (2019). https://doi.org/10.1016/j.procs.2019. 11.174 17. Karami, A., Bookstaver, B., Nolan, M., Bozorgi, P.: Investigating diseases and chemicals in COVID-19 literature with text mining. Int. J. Inf. Manage. Data Insights 1(2), 100016 (2021). https://doi.org/10.1016/j.jjimei.2021.100016 18. TextBlob: Simplified Text Processing. https://textblob.readthedocs.io/en/dev/ 19. Tweepy Documentation. https://docs.tweepy.org/en/stable/ 20. Lamsal, R.: Design and analysis of a large-scale COVID-19 Tweets dataset. Appl. Intell. 51(5), 2790–2804 (2020). https://doi.org/10.1007/s10489-020-02029-z

Sentiment Classification and Comparison of Covid-19

699

21. Siddiqua, U.A., Ahsan, T., Chy, A.N.: Combining a rule-based classifier with ensemble of feature sets and machine learning techniques for sentiment analysis on microblog. In: 2016 19th International Conference on Computer and Information Technology (ICCIT), pp. 304– 309 (2016). https://doi.org/10.1109/ICCITECHN.2016.7860214 22. Twitter Developer Platform. https://developer.twitter.com/en/docs 23. Alamoodi, A., et al.: Sentiment analysis and its applications in fighting COVID-19 and infectious diseases: a systematic review. Exp. Syst. Appl. 114155 (2020). https://doi.org/10.1016/ j.eswa.2020.114155 24. Hermanto, D.T., Ziaurrahman, M., Bianto, M.A., Setyanto, A.: Twitter social media sentiment analysis in tourist destinations using algorithms Naive Bayes classifier. J. Phys. Conf. Ser. 1140, 012037 (2018). https://doi.org/10.1088/1742-6596/1140/1/012037 25. Natural Language Toolkit, NLTK 3.6.3 Documentation. https://www.nltk.org/api/nltk.sentim ent.html 26. Khan, R., Rustam, F., Kanwal, K., Mehmood, A., Choi, G.S.: US based COVID-19 Tweets sentiment analysis using TextBlob and supervised machine learning algorithms. Int. Conf. Artif. Intell. 2021, 1–8 (2021). https://doi.org/10.1109/ICAI52203.2021.9445207 27. Raheja, S., Asthana, A.: Sentimental analysis of Twitter comments on Covid-19. In: 2021 11th International Conference on Cloud Computing, Data Science and Engineering (Confluence), pp. 704–708 (2021). https://doi.org/10.1109/Confluence51648.2021.9377048 28. Ridhwan, K.M., Hargreaves, C.A.: Leveraging Twitter data to understand public sentiment for the COVID-19 outbreak in Singapore. Int. J. Inf. Manage. Data Insights 1(2), 100021 (2021). https://doi.org/10.1016/j.jjimei.2021.100021

Automatic Modulation Recognition Models Based on Transfer Learning and Simulated Radio Signals in AWGN Channels Jamiu R. Olasina1,3 , Emmanuel Adetiba1,2,4(B) , Abdultaofeek Abayomi5 , Obiseye O. Obiyemi6,7 , Surendra Thakur6 , and Sibusiso Moyo8 1 Department of Electrical and Information Engineering, Covenant University, Ota, Nigeria

[email protected]

2 Covenant Applied Informatics and Communication Africa Center of Excellence,

Covenant University, Ota, Nigeria 3 Department of Computer Engineering, Federal Polytechnic Ilaro, Ilaro, Nigeria 4 HRA, Institute for Systems Science, Durban University of Technology, P.O. Box 1334,

Durban, South Africa 5 Department of Information and Communication Technology, Mangosuthu University of

Technology, P.O. Box 12363, Jacobs, Durban 4026, South Africa 6 KZN e-Skills CoLab, Durban University of Technology, Durban, South Africa 7 Department of Electrical and Electronic Engineering, Osun State University,

Osogbo, Osun State, Nigeria 8 Office of DVC Research, Innovation and Engagement, Institute for Systems Science, Durban

University of Technology, Durban 4001, South Africa [email protected]

Abstract. Automatic Modulation Recognition (AMR) has been applied in several areas in wireless communication. However, out of all the developed AMR models, none has successfully catered for large number of blind detection of digital modulation schemes. In this current study, we simulated and labeled datasets for fifteen digital modulation schemes. The Signal-to-Noise Ratio (SNR) of the data ranges from −20 dB to +20 dB, over different channel impairment scenarios such as Line-of Sight (LOS) with Additive White Gaussian Noise (AWGN) and NonLine of Sight (NLOS) with AWGN. Transfer learning approach was employed to develop the AMR models using pre-trained Convolutional Neural Network (CNN) architectures. The results obtained, which ranged from training accuracy values of 91.67% (+5 dB) to 100% (−15 dB) for LOS, and 92.50% (−20 dB) to 100% (15 dB) for NLOS, proved significant. The lowest testing results for LOS and NLOS were 70% (+5 dB) and 81.67% (+10 dB) respectively, and 100% testing accuracies were attained in all cases for +20 dB SNR. The proposed AMR model is thus efficient and reliable for modulation recognition in different software-defined and cognitive radio use cases. Keywords: Automatic Modulation Recognition · Additive White Gaussian Noise · Line-of Sight · Non-Line-of Sight

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 700–712, 2022. https://doi.org/10.1007/978-3-030-96299-9_66

Automatic Modulation Recognition Models

701

1 Introduction Communication is an integral part of the society and without it, the whole world would be in dilemma. The technological advancement in information and communication technology has allowed modulation and demodulation of radio signals to enhance communication over a distance, such that the robustness of these technologies that are coupled to achieve the communication process will determine its efficient utilization. A communication system is either analog or digital depending on the type of signal that it is handling. Analog system is greatly affected by noise signal and it performs poorly with increased threshold of noise level, while digital system has successfully overcome the challenge of noise [1] due to its anti-interference properties. These properties make the digital system to be more reliable, and resilient than its analog system counterpart. Digital communication has therefore broadened the scope of radio communication as there are many digital modulation and demodulation technologies than in analogue system. Consequently, more communication techniques are now performed using digital means. The two cases of digital modulation are Cooperative Conditions (CCs) and NonCooperative Conditions (NCCs). In NCC, Automatic Modulation Recognition (AMR) is applied between signal detection and demodulation phases, and its main objective is to recognize the modulation type of the transmitting signal. NCCs became more predominant due to its many applications in both civilian and military research domains. Some of these applications include signal confirmation, identification of interference, confirmation of interference, counter-measuring radio communication in Software Defined Radio (SDR) technology, and spectrum management [1]. Therefore, the need to consider more digital modulation techniques in the domain of AMR is very important [2], and as a result, in this paper, we carefully generated (simulated) synthetic datasets for fifteen different digital modulation schemes, for nine different values of SNR, over the channel impairments of LOS-AWGN and NLOSAWGN. Up till now, to the best of our knowledge, no research study in the literature has catered for all the mentioned parameters at once. AMR is the process of modulation exploration from a received signal, which can aid Cognitive Radio (CR) system to opportunistically use any unused frequency band at a particular period of time, without a disruption to the original primary users [3, 4]. AMR consists of Likelihood-Based (LB) and Feature-Based (FB) [5, 6] recognition. The LB AMR performs modulation recognition by considering the modulation with the highest probability value, while the FB modulation recognition is dependent on characteristics such as cyclostationary, higher order statistics and etc., leveraging on Machine Learning (ML) approach for classification task [7, 8]. In this current study, datasets are generated for 15 different digital modulation schemes for SNR values of −20 dB, −15 dB, −10 dB, −5 dB, 0, +5 dB, +10 dB, +15 dB, and +20 dB for Line of Sight (LOS) - AWGN and Non-Line of Sight (NLO) AGWN using GNU radio companion. The GNU radio companion is a simulation environment for simulation of any radio communication systems, using appropriate signal flow diagrams that depict some of the real life signal characteristics. AMR has been popularly utilized in the domain of radio communication through application in CR, using the LB, FB and recently with the deep learning (DL) approach outperforming all the others including the shallow ML approach [9–11].

702

J. R. Olasina et al.

In real-life setting, the mathematical modelling of a received signal can be represented by Eq. (1) [2]:  ft x(t) = s(td (t − ft))g(fi)d (fi) + Na(t) (1) fi

where s(t) is the signal that was transmitted, td is the time deviation, g(fi) is the wireless channel for transmission, d(fi) is the characteristics response to a step input parameter, and Na(t) is the Additive White Gaussian Noise (AWGN). The signal transmission can be represented by Eq. (2) [12]  cn ej(wn t+£) k(t − nQm ) (2) s(t) = n

where wn is the angular frequency of the signal, £ is the initial phase of the frequency carrier, Qm is the periodic symbol, cn is the symbolic sequence, and k(t) is the shaping filter. A particular digital modulation can be obtained by varying the parametric properties of Eq. (2). In this work, the fifteen different digital modulations considered for AMR include: 4QAM, 16QAM, 64QAM, 256QAM, 8PSK, 16PSK, 32PSK, 64PSK, 128PSK, 256PSK, CPFSK, DQPSK, DBPSK, GMSK, and GFSK. The remaining portion of this paper is structured as follows: Sect. 2 is the related works, methodology is contained in Sect. 3 while the results and discussion are presented in Sect. 4, and the conclusion in Sect. 5 respectively.

2 Related Work The state-of-the-art Convolutional Neural Network (CNN) based DL has become handy due to its wide use in the field of radio communication and AMR studies. CNN is a highly integrated computational algorithm, with deeply stacked layers with which it performs feature extractions unlike the traditional, and shallow ML approaches where specified complex mathematical and statistical tools are required for feature extraction task. A modular classification system for the implementation of AMR was carried out in [13]. However, the model has a problem of recognizing high SNR values and also limited to only three modulations, which include 8PSK, QPSK, and GFSK. In [14], the authors simulated Orthogonal Frequency Division Multiplexing (OFDM) in multipath and noisy channels to address high-sensitivity to SNR with high Bit Error Rate (BER) values. This was achieved using equalization of synchronized and estimated channel to reduce the threshold of noise level. Eleven digital modulations were recognized using CNN by the authors in [15], attaining the highest training accuracy of 87.4%. In [9], an accuracy of 99.8% was obtained using AlexNet pre-trained model, but the work could not cater for AMR of a large number of modulation with varied SNR values. Further studies by the authors in [16] proved the improvement of deep learning based CNN over the traditional ML approaches including the Support Vector Machine (SVM), by a better training accuracy for Line of Sight (LOS) and Non-Line of Sight (NLOS) channels, using different SNR. An accuracy of 82.8% was achieved in [10] using constellations datasets. Furthermore, the authors in [17] achieved a training accuracy of 94% over SNR between −10 dB

Automatic Modulation Recognition Models

703

and +10 dB with DL-based CNN approach. Despite the remarkable improvement of deep learning based AMR over the traditional approaches, coverage of large number of digital modulation in wide SNR range is scanty in the literature. Thus, for the AMR model in the study at hand, we utilized constellation images from carefully simulated radio signals for fifteen digital modulation schemes, over −20 dB to +20 dB SNRs in LOS and NLOS channels as well as a pre-trained CNN model for both feature extraction and classification tasks.

3 Methodology The methodology applied for the realization of this work consists of the generation of simulated datasets, pre-processing and labeling of the datasets, and the training of CNN for AMR of digital radio signals. According to the Fig. 1, datasets were structured after preprocessing, before the training proper. The structuring of the datasets is a major motivation in the ongoing research work, and the recorded accuracies showed its significance as likely alternative datasets in place of real life datasets, and this would be considered in the nearest future. 3.1 Datasets Generation Datasets were carefully simulated for fifteen different digital modulation schemes for LOS-AWGN and NLOS-AWGN channels, using GNU-radio companion interactive flow diagrams. The flow diagram for typical datasets generation for a modulation scheme in this study is shown in Fig. 2. The datasets were structured into the fifteen digital modulation schemes and each modulation was further divided into nine different SNR values, in the range of −20 dB ≤ SNR ≤ +20 dB, where the SNR was computed using the formula, VN = 1/antilog(0.05 × SNR), and plugged into the dynamic channels of the GNU radio block. The LOS and NLOS channels were incorporated with various AWGN, to mimic real life signal scenario. The generated datasets were pre-processed by carefully filtering the collected datasets of size 400 by 400 by 3 pixels, and the batch processing tool was then applied to resize all datasets into 227 by 227 by 3 pixels for the AlexNet input image size. 3.2 AMR Model Development The proposed AMR model is based on an existing pre-trained neural network (i.e. AlexNet), using transfer learning approach. The AlexNet architecture was adapted by changing the last fully connected layer of the network to accommodate fifteen different classes, with each class representing a modulation scheme. In this study, 5,400 samples of simulated datasets were used to evaluate the proposed AMR model after it was trained with simulated datasets of size 21,600. All the parameters of AlexNet were fixed, except the last classification layer that was re-configured to fit the number of classes in the study at hand (i.e. fifteen classes). All the models were trained using the Deep Learning Toolbox in MATLABR2020a environment. The Stochastic Gradient Descent with Momentum (SDGM) was explored for the models and the categorical cross-entropy loss

704

J. R. Olasina et al.

Fig. 1. Block diagram of the proposed work

Fig. 2. Dataset simulation flow diagram for DBPSK modulation scheme in GNU radio companion.

Automatic Modulation Recognition Models

705

function was used. We configured ReLU activation functions for all layers except the last densely connected layer in which Softmax was utilized. The learning rate was 0.0001, batch size of 10 and validation frequency of 30.

4 Result and Discussion 4.1 Result of Generated and Pre-processed Datasets Dataset simulation parameters were set according to the procedures in [9, 18] were adopted due to their semblance to real-life radio communication scenario. Figure 3 shows samples of generated constellation diagram before the pre-processing phase. The figures include colored radio image, the gray image version, Fast Fourier Transform (FFT) of the image and bar chart.

Fig. 3. Sample of a generated raw radio signal.

The simulated datasets were then pre-processed by removing the noise around the constellation diagram pixels, using appropriate scripts in MATLAB2000a. The corresponding results generated are shown in Fig. 4. After the pre-processing stage, the images were subjected to batch processing, which involves the conversion of constellation images to 227 by 227 by 3. This is the required

706

J. R. Olasina et al.

Fig. 4. Sample of pre-processed raw radio signal

input image size for AlexNet pre-trained model. The output of the batch processing for the considered modulation schemes in this study is shown in Fig. 5. As shown, the constellation diagrams of modulation schemes belonging to different types are similar at the low SNR values. However, from 0 dB up to 20 dB, the constellations of modulation schemes (in the same classes) are very similar while those of different classes are highly dissimilar. 4.2 Results of AMR Models The validation and testing results in this study are presented in Table 1. For our model that was built with the LOS-AWGN dataset, the least validation accuracy (93.33%) and testing accuracy (80.00%) were obtained at −20 dB. This is an expected result since the low SNR culminated in distorted constellation diagrams for the different modulation schemes/classes. However, there are major improvements at +15 dB and +20 dB SNRs, with the model producing validation and testing accuracies of 100%. On the other hand, the model that was built with NLOS-AWGN dataset gave the poorest validation accuracy of 92.50% at −20 dB and least testing accuracy of 81.67% at +10 dB. For this model, the highest validation and testing accuracies of 100% were obtained only at +20 dB. These results obviously show that the NLOS-AWGN dataset provides a tougher model of real-life wireless communication environment. Evidences of these results are further presented graphically in Figs. 6, 7, 8, 9 and 10 for accuracies and losses against the

Automatic Modulation Recognition Models

707

Fig. 5. Pre-processed and labeled training datasets.

SNR values from −20 dB to +20 dB for models trained with both LOS-AWGN and NLOS-AWGN channels at Epoch 5. The figures show high accuracies and low losses at high SNRs. Result of the existing works that are based on constellation diagrams and CNN reported in [10, 17, 19] covered SNRs between −10 dB and +10 dB, and considered lower classes than the 15 modulation schemes herein considered. Despite the lower numbers of modulation classes in these previous works, the reported accuracies are lower than our result in the study at hand. Table 1. Results of our AMR models SNR (dB)

AMR models LOS-AWGN Validation (%)

NLOS-AWGN Testing (%)

Validation (%)

Testing (%)

−20

93.33

80.00

92.50

100.00

−15

100.00

94.67

100.00

100.00

−10

93.33

100.00

93.16

93.33 (continued)

708

J. R. Olasina et al. Table 1. (continued)

SNR (dB)

AMR models LOS-AWGN

NLOS-AWGN

Validation (%)

Testing (%)

Validation (%)

−5

100.00

100.00

100.00

0

Testing (%) 98.67

100.00

84.00

100.00

93.00

+5

91.67

70.00

100.00

82.67

+10

93.33

90.67

93.33

81.67

+15

100.00

100.00

100.00

94.00

+20

100.00

100.00

100.00

100.00

Fig. 6. Bar charts for training accuracy and loss against SNR for the LOS-AWGN channel at 5 Epochs (Blue Bars = Accuracy; Orange Bars = Loss)

Fig. 7. Bar charts for training accuracy and loss against SNR for the NLOS-AWGN channel at 5 Epochs (Blue Bars = Accuracy; Orange Bars = Loss)

Automatic Modulation Recognition Models

709

Fig. 8. Training accuracy and loss versus number of Epoch for the AMR Model Trained with SGDM Optimizer

Fig. 9. Bar charts for training accuracy and loss against SNR for the NLOS-AWGN Channel at 5 Epochs (Blue Bars = Accuracy; Orange Bars = Loss) and 10 Epochs (Yellow Bars = Accuracy; Purple Bars = Loss)

710

J. R. Olasina et al.

Fig. 10. Plot showing the accuracies against SNR (from −20 dB to +20 dB) for LOS-AWGN (Green = Validation; Red = Test) and NLOS-AWGN (Blue = Validation; Violet = Test)

5 Conclusion The preliminary results obtained in this study have no doubt proved the efficacy of an AMR model developed using transfer learning approach (with AlexNet pre-trained architecture) and constellation datasets for LOS-AWGN and NLOS-AWGN. Notably, all the validation accuracies are from 91.67% to 100%. These models show potentials for incorporation in a modern day wireless receiver to aid software-defined and cognitive radio functionalities. Being preliminary results of our on-going research on multi-task Wireless Signal Recognition (WSR), our future works will involve exploration of more simulated data as well as Over-the-Air (OTA) datasets in real-time wireless environments with harsh characteristics. We hope to come up with a robust multi-task WSR models from our endeavors that can be leveraged for software-defined and cognitive radio functions such as spectrum sensing, automatic demodulation and spectrum management in 2G/3G/4G/5G/beyond5G, Internet of Things (IoT) and Cloud Federation use cases. Acknowledgments. This research was carried out at Covenant University’s Advanced Signal Processing and Machine Intelligence Research (ASPMIR) laboratory. The High Performance Computing (HPC) node of the Covenant Applied Informatics and Communication African Centre of Excellence (CApIC-ACE) FEDGEN Testbed was utilized for experimentations. We also acknowledge the Office of the DVC Research, Innovation and Engagement as well as the Institute

Automatic Modulation Recognition Models

711

for Systems Science (ISS), and KZN e-Skills CoLab, Durban University of Technology, Durban, South Africa for collaboration and support for the publication of this study.

References 1. Jiang, K., Zhang, J., Wu, H., Wang, A.: Applied sciences based on deep convolutional neural network, pp. 1–14 (2020) 2. Zha, X., Peng, H., Qin, X., Li, G., Yang, S.: A deep learning framework for signal detection and modulation classification. Sensors 19(18), 4042 (2019). https://doi.org/10.3390/s19184042 3. O’Shea, T.J., Corgan, J., Clancy, T.C.: Convolutional radio modulation recognition networks. In: Jayne, C., Iliadis, L. (eds.) EANN 2016. CCIS, vol. 629, pp. 213–226. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-44188-7_16 4. Azza, M.A., El Moussati, A., Moussaoui, O.: Implementation of an automatic modulation recognition system on a software defined radio platform. In: International Symposium on Advanced Electrical and Communication Technologies, ISAECT 2018 - Proceedings, pp. 1–4 (2019). https://doi.org/10.1109/ISAECT.2018.8618837 5. Ajala, S., Adetiba, E., Ajayi, O.T., Abayomi, A., Kelechi, A.H., Moyo, S.: Automatic Modulation Recognition Using Minimum-Phase Reconstruction Coefficients and Feed-Forward Neural Network. J. Comput. Sci. Eng. (2019, accepted) 6. Hazza, A., Shoaib, M., Alshebeili, S.A., Fahad, A.: An overview of feature-based methods for digital modulation classification. In: 2013 1st International Conference on Communications, Signal Processing, and their Applications ICCSPA 2013, vol. 1, no. 08 (2013). https://doi. org/10.1109/ICCSPA.2013.6487244 7. Ramkumar, B.: Automatic modulation classification for cognitive radios using cyclic feature detection. IEEE Circ. Syst. Mag. 9(2), 27–45 (2009) 8. Kaur, M.: Automatic modulation recognition for digital communication signals. Int. J. Soft Comput. Eng. 2(2), 110–114 (2012) 9. O’Shea, T.J., Roy, T., Clancy, T.C.: Over-the-air deep learning based radio signal classification. IEEE J. Sel. Top. Signal Process. 12(1), 168–179 (2018). https://doi.org/10.1109/JSTSP.2018. 2797022 10. Peng, S., et al.: Modulation classification based on signal constellation diagrams and deep learning. IEEE Trans. Neural Networks Learn. Syst. 30(3), 718–727 (2019). https://doi.org/ 10.1109/TNNLS.2018.2850703 11. Liu, K., Gao, W., Huang, Q.: Automatic modulation recognition based on a DCN-BILSTM network. Sensors 21(5), 1–17 (2021). https://doi.org/10.3390/s21051577 12. Danse, M., Klerkx, L., Reintjes, J., Rabbinge, R., Leeuwis, C.: Unravelling inclusive business models for achieving food and nutrition security in BOP markets. Global Food Secur. 24, 100354 (2020). https://doi.org/10.1016/j.gfs.2020.100354 13. Jagannath, J., Saarinen, H.M., Drozd, A.L.: Framework for automatic signal classification techniques (FACT) for software defined radios. In: 2015 IEEE Symposium on Computational Intelligence for Security and Defense Applications, CISDA 2015 - Proceedings, pp. 54–60 (2015). https://doi.org/10.1109/CISDA.2015.7208628 14. Anjana, C., Sundaresan, S., Tessy Zacharia, R., Gandhiraj, K.P.S.: An experimental study on channel estimation and synchronization to reduce error rate in OFDM using GNU Radio. Procedia Comput. Sci. 46, 1056–1063 (2015). https://doi.org/10.1016/j.procs.2015.01.017 15. O’Shea, T.J., West, N.: Radio machine learning dataset generation with GNU radio. In: 6th GNU Radio Conference, pp. 1–6 (2016). https://scholar.google.com/citations?view_op= view_citation&hl=en&user=4S4GyXYAAAAJ&citation_for_view=4S4GyXYAAAAJ: 9Nmd_mFXekcC

712

J. R. Olasina et al.

16. Yang, C., He, Z., Yang Peng, Y., Wang, J. Y.: Deep learning aided method for automatic modulation recognition. IEEE Access 7, 109063–109068 (2019). https://doi.org/10.1109/ ACCESS.2019.2933448 17. Wang, J., Barth, J., Göttgens, I., Emchi, K., Pach, D., Oertelt-Prigione, S.: An opportunity for patient-centered care: Results from a secondary analysis of sex- and gender-based data in mobile health trials for chronic medical conditions. Maturitas 138(May), 1–7 (2020). https:// doi.org/10.1016/j.maturitas.2020.05.003 18. O’Shea, T., Hoydis, J.: An introduction to deep learning for the physical layer. IEEE Trans. Cogn. Commun. Netw. 3(4), 563–575 (2017). https://doi.org/10.1109/TCCN.2017.2758370 19. Wu, P., Sun, B., Su, S., Wei, J., Zhao, J., Wen, X.: Automatic modulation classification based on deep learning for software-defined radio. Math. Probl. Eng. 2020, 1–13 (2020). https:// doi.org/10.1155/2020/2678310

Analysis of the Access to the Financing of the Ecuadorian Companies in the Framework of the Sanitary Emergency of COVID 19 and the Economic Sectors of Unemployment Marcelo León2(B)

, Carlos Redroban3

, Vinicio Loaiza1

, and Paulina León1

1 Universidad Tecnológica Empresarial de Guayaquil, Guayaquil, Ecuador 2 Grupo de Estudio de Metodologias para Ingenieria en Software y Sistemas de Información

GEMIS, Buenos Aires, Argentina 3 Universidad ECOTEC, Guayaquil, Ecuador

Abstract. This work was framed in the sanitary COVID 19 crisis, with the aim to consider the probability of access to credit, in financial institutions properly registered by the Superintendence of Banks, and Superintendence of Popular Economy and Supportive. We used the data of the Structural Company Survey of the year 2020. Through the analysis of the logistic probability in the economic sectors, was verified the commercial and manufacturing sectors have a bigger probability to access credit, making them susceptible in the financial sector, due to the critical effects created by the measures to hold the pandemic. Measures that range from governmental politics to the administrative politics of the company may attenuate the consequences. Keywords: Financing · Companies · COVID 19 · Logistic probability analysis

1 Introduction Within a year of the declaration of the global sanitary crisis, the Ecuadorian economic agents try to overcome the uncertainty, difficulties provoked by the restrictive, and confinement measures adopted by the competent authorities. The Central Bank of Ecuador [1], foresaw the decrease of the total claim of wellbeings and services 9.97% in the year 2020, in contrast to 2019. Being the cause the decrease of the salaries in the public sector, with the increase in unemployment, the deterioration of the expectations of investors, and because of a contraction of a 22% in the remittances received from Italy, Spain and the United States. In that sense, Peñarreta indicated that the access to financing has a positive effect at emergence of new companies, and settled companies [1]. It is necessary to have the access to the data of Ecuadorian credit of the companies, according to its economic sector and activities. The purpose of the present work is to identify the probability that a company has a credit. According to the economic sector in which plays, in addition, to unbundle the information in economic activities, and the captured data by the Ecuadorian Statistics © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 713–722, 2022. https://doi.org/10.1007/978-3-030-96299-9_67

714

M. León et al.

and Census Institute in 2019 through logit models [3], which Medina [2] indicates that is used to estimate a probability, in this case, that the economic sectors have access to external financial [1]. Among the main findings, is possible to verify that the economic sector with more probability of having credit is the commercial,. Being the sector that shows more recovery. On the other hand, mining is the sector with less probability of having external financing. The activities of the health, finances, secures and wholesale and retail commerce would be the activities with more probability of access to a credit to the reactivation. The contribution of the investigation will help to determine the sectors and economic activities that require governmental politics as well as management in order to take advantage of its capacity of financing and overcome the crisis. The investigation is structured in five sections in which it is detailed the context of the study, main and recent studies related to it, data and methodology used, and last the conclusions and recommendations derived of the analysis.

2 Literature Review The financial capital is indispensable in the business environment, in particular in the development of inversions in fixed capital or innovation. In that sense, Roja mentions that in order to close the financing gap among the different economic and productive agents, the governments have begun to finance the companies through programs of development such as programs of credits to the inversion and innovation, development bank, bank guarantees or lines of financing, assimilating the experience of other countries [2]. According to Martínez et al., in the regional case of Common Market of the South, applying a methodology of regressions logit and probit, it was verify that in Argentina, Brazil and Uruguay, the micro companies that are older and play in the manufacturing sector have a 53%, 47% and 38% respectively, having more probabilities of accessing to a credit. In the case of Paraguay, the medium companies that operate in the manufacturing sector have a 73% probability of having external financing [3]. Wellalage & Locke aggregates the perspective of gender studying the access to credit, analyzing the determinants of financing in entrepreneurship managed to women through the methodology of regressions probit, indicates that women entrepreneurs have a 3% less of accessing to a credit compared to men [4]. At national level, Portal et al. [5], and Belén et al. [6], analyzed the determinants of external micro-companies financing, and the access to micro-companies banking financing, and of the Pymes of high technology respectively. Through a statistic multivariate technique of groups, in which conclude coinciding partially that the features of the microcompanies and the profile of the owner are determinant to access to credit. In the first case, they have access to an external financial. Those companies have a bigger quantity of employees and quality young owners, considering their higher level of schooling and experience. On the other hand, for Belén et al., the features are not determinant in the collection of credit for the companies of high technology [6]. León et al. [7] as well as Wellalage and Locke [4], in the Peruvian case, used a probit model, mention that the gender matters in the moment of accessing to a credit, also

Analysis of the Access to the Financing of the Ecuadorian Companies

715

highlight that the older manufacturing companies have a higher probability of accessing to a loan. Looking into in our space of study, the Ecuadorian are, Delgado & Chávez indicate that the main sources of financing are the banking credit and the own funds, however very few access to a credit, what implies its early closure [10]. Franco et al. [11], through the methodologies probit and logit, with data from the three main provinces of the country (Guayas, Pichincha, and Azuay), determines that there are between 28% and 56% of medium size companies with a higher probability of obtaining credit. Confirming that moreover a relation between the economic activity and the collection of credit. Being the activity of food and trade, those that have a higher probability of getting it, ignoring the other provinces of the country and the rest of economic activities that the official statistics consider.

3 Data and Methodology With the purpose of analyzing the access to credit in financial institutions of the Ecuadorian companies, according to the sector and economic activity in which they operate, data of the Business Structural Survey in 2019, got from 3526 companies by the Ecuadorian Statistics and Census Institute in 2019. As is shown in Fig. 1, the dependent variable is the access to financing, meanwhile, the independent variable in the first case, it refers to the economic sectors, and in the second case the productive activities of the Ecuadorian economy. The control variables are the level of actives, and the capital’s property used to layout a better set of the results with the intervention of factor when applying for financing (Table 1). Table 1. Description of variables (Dictionary of variables from the Structural Business Survey (2019).) Name

Definition

Access to financing Expresses if the company had a line of credit with a financial entity Economic sectors

Kind of variable Dichotomous variable

It describes the economic sectors such as: Nominal categorical variable trade, construction, manufacture, mining and services

Economic activities It describes the economic activities in which are divided the economic sectors

Nominal categorical variable

Total actives (log)

Tangible and intangible fixed assets year 2019 Value of the company’s total assets

Expressed in dollars of USA

Capital ownership

It describes the ownership of capital in companies: private, public and mixed

Nominal categorical variable

716

M. León et al.

According to Ordóñez et al. [12], the financial system plays an important role in the world economy. In the Ecuadorian economy, it is necessary to recognize its composition and the factors that affect the dynamics of each sector and its growth. The financial sector in Fig. 1, indicates that at the national level, 63% of the business sector has financing from a financial institution, whether private/public/cooperative/mutual, among others„ having a moderate access by companies to some type of external financing.

Fig. 1. Access to financing.

Figure 2 shows that the largest number of companies with financing are located in the provinces of Guayas and Pichincha, that is, in the most important cities with the largest number of companies. The province of Guayas concentrates most of the companies. Since it is the country’s main port for the export and import of raw materials and finished products, which means that there is a high level of economic movement; The province of Pichincha concentrates the activities of public administration as it is the province of the capital of the country, in addition to great commercial movement and high demand for professional services, far below are Azuay, Manabi and El Oro, which in comparison with the previous two provinces its concentration of companies is low, they have a great economic dynamic promoted mainly by industrial activities in the case of Azuay and Tungurahua, and commercial activities in the case of Manabi and El Oro. According to data from the Central Bank of Ecuador [1], and with the Monetary and Financial Policy and Regulation Board in its resolution 603-2020-F, Table 2 shows the maximum lending interest rates by segment for the private, public and popular and solidarity financial sector. It can be observe that the more complex the organizational structure of the company, the lower the maximum interest rate. Likewise, for companies or companies with sales of less than $ 100,000, or small-scale activities, a higher interest rate is perceive than the other segments.

Analysis of the Access to the Financing of the Ecuadorian Companies

717

Fig. 2. Companies with external financing by province.

Table 2. Maximum interest rates. Segment

% Annual

Corporate product

09.33

Business product

10.21

Productive SMEs

11.83

Retail microcredit

28.50

Simple accumulation microcredit

25.50

Extended accumulation microloans

23.50

Based to the Ecuadorian Institute of Statistics and Census (2020) for the official analysis there are five economic sectors which are manufacturing, mining, commerce, construction and services; in the Ecuadorian economy 42.60% of establishments are dedicated to commerce, 29.52% are dedicated to services, 19.12% to manufacturing, 5.81% to construction and 2. 95% to mining, however, the economic returns are not explained by the number of establishments, as shown in Table 3, since the sector with the highest average sales level is mining with approximately 67 million dollars, followed by manufacturing industry with 42 million, commerce with 26 million, construction with 3.5 million and services with one million dollars (Fig. 3).

718

M. León et al.

Fig. 3. Average sales level by economic sector.

4 Methodological Basis The present research methodology is descriptive, analytical and econometric to know the characteristics and state of financing of Ecuadorian companies in 2019, a logistic probability or logit model suggested by Medina [3] is used, according to Gujarati and Porter [12] and the logit model is suitable for the analysis of binary choices, that is, when the dependent variable is dichotomous in which the two possible outcomes are 1 and 0, being this the logarithm of the odds ratio, being the best option when it comes to analyze individual cases or at the microeconomic level, the logistic probability model is defined in the following distribution function of equation P(Y = 1Xi) = 11 − eZiPY = 1Xi = 11 − eZi

(1)

Where Yi = 1: has financing, Yi = 0: does not have financing, Xi: sectors or economic activities, P(Y = 1Xi): probability of having accessed external financing, Zi: exponent of the exponential which is a linear regression. When linearizing the distribution function, Eq. (2) is observe. Zi = β0 + β1Xi + μZi = β0 + β1Xi + μ

(2)

Where Zi: exponent of the exponential that is a linear regression, β0 is the intercept of the curve, β1 is the slope of the curve and μ is the stochastic error.

5 Results and Discussion The results shown in Table 3, indicate that the model of access to credit by economic sector, at a global level, is significant, as shown by its value of Prob > chi2 = 0.0000, that is, less than 0.05. According to Franco et al. [12], access to credit is also determined, in addition to the economic sector, by the size of the company, employee training, use of ICT’s, shareholder concentration, management experience, age of the company, level of sales, exports and product quality, As stated by Franco et al. (2019), the economic sector with the highest probability of accessing external or bank financing is commerce, in relation to this it is observed

Analysis of the Access to the Financing of the Ecuadorian Companies

719

that the construction sector is 3.06 times less likely to access credit, the manufacturing sector is 1.25 times less likely to access, on the other hand the mining sector is 2. 24 times less likely to access and finally the service sector companies have 2.19 times less likely to access financing in a financial institution, additionally it also indicates that if the company has more assets, it is more likely to access a credit. Differences are evidenced in the cone the MERCOSUR cases [6, 10]. Financing was concentrated by the manufacturing industry companies. and the external financing is mainly used by high-tech companies [9], which is a reflection of the Ecuadorian economic structure based on trade in goods. Table 3. Logit model of access to credit by economic sector. Variable

Feature

Total assets (log

Coefficient 0.1617

Odd-Ratio 1.1175

Probability

Base

0.0000 Private trade

Sector

Construction

-1.1844

0.3268

0.0000

-0.2195

0.8029

0.0420

-0.8040

0.4474

0.0000

-0.7855

0.4559

0.0000

-1.6503

0.1920

0.000

1.2234

0.588

Remarks

0.2017 3519

Prob > chi2

0.0000

Manufacturing Mining Services Capital ownership Public mixed

We can also highlight the case of the mining activity, that being the activity with the highest average sales, operates with in most cases with own capital, this because they have a series of stricter state and administrative regulations at the time of starting operations, in addition to having a great help of foreign investment. On the other hand, commercial companies depend on the constant flow of cash and on a great liquidity, to be able to meet their obligations with suppliers, human talent, accounts payable, obligations with public institutions. within the framework of the sanitary emergency and the mobility restriction and confinement measures, commercial companies are the most likely to generate higher financial expenses due to moratoriums or debt restructuring with financial institutions, reducing their future resources to innovate or venture into digital and virtual commercialization; in the case of the manufacturing sector, this is another sector that would be severely affected, since the high probability of a manufacturing company having a loan and the impossibility of operating normally would also mean a significant loss, since many of the resources are already invested in raw materials and fixed assets, in addition to the obligations acquired previously; For both cases, i.e. commercial and manufacturing companies, the decrease in demand would be added to the delicate economic situation that the Ecuadorian economy was already going through and the increase in prices in their inventories due to the withdrawal of fuel subsidies,

720

M. León et al.

complicating their situation and thus the possibility of meeting their obligations in order to obtain more financing. With respect to the construction, mining and services sectors, that are sectors with a lower probability of having a loan and being the services sector the one that has adapted the fastest to the new normality, they would have most of their resources free to face the crisis and eventually resort to bank financing. Table 4. Logit model of access to credit by economic activities. Variable Total assets (log

Feature

Coefficient 0.1755

Odd-Ratio 1.1919

Probability 0.0000

Base

Accommodation Health Administrative Services Finance and Insurance Real Estate Professional

Economic activity

Arts and Entertainment Wholesale and Retail Trade Construction Water and Sewerage

0.2206

3.0217

0.000

0.0902

1.0944

0.849

1.6418

5.1645

0.000

0.3235

1.3819

0.193

0.8989 0.5174

2.4569 1.6777

0.000 0.042

1.0304

2.8023

0.002

0.3305

1.3917

0.282

-0.9059

0.4041

0.014

Education Mining and quarrying

0.4510

1.5699

0.184

0.4337

1.5430

0.132

Manufacturing

0.4971

1.6440

0.070

Information and communication

0.0490

1.0502

0.864

Other services

0.0209

1.0211

0.940

Energy supply Transportation and storage

-0.3125

0.7315

0.414

0.3872

1.4728

0.321

Public

-2.1274

0.1191

0.000

Mixed

0.2026

1.2246

0.588

Capital Ownership

Remarks Prob > chi2 Pseudo R2

3519 0.0000 0.0559

Additionally, Table 4 shows the probability of access to credit by economic activities, which are disaggregated levels of the economic sectors analyzed above. In this case,

Analysis of the Access to the Financing of the Ecuadorian Companies

721

it’s identified that seven of seventeen economic activities are significant at the time of acquiring credit. Considered with respect to the lodging and hotel activity, activities such as health, financial and insurance services, professional services, wholesale and retail trade, water and sewerage, and manufacturing have a statistically significant probability of accessing credit, so they are activities that require large amounts of investment and working capital on a large scale.

6 Conclusions The economic structure of each country has a direct relationship with the demand for financing in the economic sectors. However, it should be emphasize that mining, despite being the economic sector that has generated the highest sales on average in 2019, is more likely to use its own funds, which are mostly foreign capital. The present research achieves the objective of analyzing the characteristics of access to credit by each sector and economic activity, among the main results it stands out that it is the commercial sector that has made more use of external financing. It differs from results presented in emerging economies such as Brazil, Argentina, Uruguay, and Paraguay, which bet on the manufacturing sector. Among the disaggregated economic activities that resort to bank loans are commerce, health, professional services, manufacturing, and financial services. It is assumed a greater credit affectation towards the Ecuadorian commercial sector had problems adapting to the virtuality and restrictions of the new normality. Unlike other activities that have been able to opt for teleworking or in turn, the demand for their services has increased during the current health crisis, it is necessary that the relevant government entities establish legal or administrative measures to help companies to maintain adequate liquidity such as moratoriums, suspensions of debtor dues that mitigate the credit crisis in this activity without significantly deteriorating the financial stability of the national economy. Likewise, company managers will establish internal policies in each company that will allow them to have the necessary resources to meet their employer, tax, and operational obligations, dispensing with expenses that are not extremely necessary, and making use of mobile and virtual technologies to provide better quality products and service.

References 1. BCE: El Covid-19 Pasa Factura A La Economía Ecuatoriana: Decrecerá Entre 7,3% Y 9,6% en 2020. Banco Central del Ecuador, June 2020. https://www.bce.fin.ec/index.php/boletinesde-prensa-archivo/item/1369-el-covid-19-pasa-factura-a-la-economia-ec 2. Peñarreta, M.: Acceso al financiamiento y emprendimiento en la región 7 de Ecuador. Revista Publicando 4(13), 1–19 (2017) 3. INEC: Intituto Ecuatoriano de Estadísticas y Censos. ENESEM (2020). https://www.ecuado rencifras.gob.ec/encuesta-a-empresas/ 4. Medina, E.: El Uso de los Modelos de Elección Discreta para la Predicción de Crisis Cambiarías: El Caso Latinoamericano. Universidad Autónoma de Madrid, Madrid (2003) 5. Rojas, L.: Situación del financiamiento a Pymes y empresas nuevas en América Latina. CIEPLAN, Santiago de Chile. CIEPLA, Santiago de Chile (2017)

722

M. León et al.

6. Martinez, L.B., Guercio, M.B., Corzo, L., Hernán, V.P.: Determinantes del financiamiento externo de las PyMEs del MERCOSUR. Revista Venezolana de Gerencia 22(80), 672–692 (2018) 7. Wellalage, N., Locke, S.: Access to credit by SMEs in South Asia: do women entrepreneurs face discrimination. Res. Int. Bus. Financ. 41, 336–346 (2017) 8. Portal, M., Feito, D., Ramirez, N.: Determinantes del financiamiento externo en microempresas mexicanas. Revista Espacios 39(18), 1–13 (2018) 9. Belén, M., Martinez, L., Vigier, H.: Limitations to bank financing of high-tech SMEs. Estudios Gerenciales 33(142), 3–12 (2017) 10. León, J., Sánchez, M., Jopen, G.: Acceso y uso de microcréditos para la MYPE: la clave está en la diferenciación. Economía y sociedad 92(6), 6–11 (2017) 11. Delgado, D., Chávez, G.: Las Pymes En El Ecuador Y Sus Fuentes De Financiamiento. Revista Observatorio de la Economía Latinoamericana, pp. 1–18 (2018) 12. Franco Gómez, M.D.C., Gómez Gutiérrez, F., Serrano Orellana, K.: Determinantes del acceso al crédito para la PYME del Ecuador. Conrado 15(67), 295–303 (2019) 13. Granda, E.M.O., Zurita, I.N., Álvarez, J.C.E.: El sistema financiero en Ecuador. Herramientas innovadoras y nuevos modelos de negocio. Revista Arbitrada Interdisciplinaria Koinonía 5(10), 195–225 (2020) 14. Gujarati, D., Porter, D.: Econometría, 5th edn. McGrawhill, New York (2010)

State of the Art of Wind and Power Prediction for Wind Farms Ricardo Puga1 , José Baptista1 , José Boaventura1,2 , Judite Ferreira3,4(B) , and Ana Madureira3,4 1 Universidade de Trás-os-Montes e Alto Douro (UTAD), Vila Real, Portugal

{baptista,jboavent}@utad.pt

2 Centre for Robotics in Industry and Intelligent Systems, INESC TEC, Porto, Portugal 3 ISEP/IPP, Porto, Portugal

{mju,amd}@isep.ipp.pt 4 ISRC - Interdisciplinary Studies Research Center, Porto, Portugal

Abstract. There are different clean energy production technologies, including wind energy production. This type of energy, among renewable energies, is one of the least predictable due to the unpredictability of the wind. The wind prediction has been a deeply analysed field since has a considerable share on the green energy production, and the investments on this sector are growing. The efficiency and stability of power production can be increased with a better prediction of the main source of energy, in our case the wind. In this paper, some techniques inspired by “Biological Inspired Optimization Techniques” applied to wind forecast are compared. The wind forecast is very important to be able to estimate the electric energy production in the wind farms. As you know, the energy balance must be checked in the electrical system at every moment. In this study we are going to analyse different methodologies of wind and power prediction for wind farms to understand the method with best results. Keywords: Wind prediction · Computational fluid dynamics · Support vector machine method · Random theory · Power prediction

1 Introduction The wind prediction has been a deeply analyzed field since has a considerable share on the green energy production, and the investments on this sector are growing. The efficiency and stability of power output of the wind farms can be increased with a better prediction of the main source of energy, in our case the wind. The study of wind power prediction has different ways to be drilled into, a deep analysis of different methods is done to understand which one has better results. In this study we are going to analyze different methodologies of wind and power prediction for wind farms to understand the method with best results. In the past decades the fossil fuels were the main source of energy, this must be mitigated since this source is highly finite and leads to air pollution contributing to the © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 723–732, 2022. https://doi.org/10.1007/978-3-030-96299-9_68

724

R. Puga et al.

ozone depletion and global warming. The accuracy and precision of wind/power prediction is the core of making the wind a more reliable source of energy. A better prediction will result in an increase of efficiency and stability. There are several techniques and mathematical models that allow to minimize the impact of the wind and power prediction problem [1]. Some of these models and mathematical techniques are inspired by biological algorithms. This paper presents some studies of wind power prediction. There are different ways to be drilled into, a deep analysis of different methodologies is done. So, to understand which one has better results we analyzed: the Fuzzy Logic base Algorithms, Wavelet Theory, Back Propagation Neural Network, Back Propagation Neural Network with ARMA.

2 Biological Inspired Optimization Techniques 2.1 Fuzzy Logic Base Algorithms Fuzzy logic approaches the truth or false in a different way, where 0 and 1 are the extreme of the scope, in between these exist different degrees of truth. There are some ways to affect that fuzzy logic: first fuzzification where fresh inputs go exchanged to blurred inputs; secondly these inputs go processed with blurred principles to produce a blurred signal; and lastly defuzzification which leads to a degree of the result as in fuzzy logic there can be more than one result with different degrees. Figure 1 represented the diagram of fuzzy logic technique.

Fig. 1. Fuzzy logic structure

Adaptive Neuro-Fuzzy Inference System: ANFIS is an adaptive multi-layer feed forward networks used for nonlinear prediction, where past data samples are used to predict the future data samples, aligned with self-learning ability of neural networks with the linguistic expression function of fuzzy inference system. A type of ANFIS network is a Takagi-Sugeno fuzzy inference system mapped to a neural network structure with five layers. Each layer contains several nodes characterized by the node function as seen in Fig. 2. The study presented in refence [2] shows the implementation of a double stage (Fig. 2) hierarchical adaptive neuro-fuzzy inference system model for short-term wind power prediction of a micro-grid wind farm.

State of the Art of Wind and Power Prediction for Wind Farms

725

Fig. 2. Development pipeline for the proposed algorithms.

The first stage implements the wind the forecast at the turbine site, the second stage model the wind speed to the power output of the turbine. This combination results in the next day power production forecast. Study presented in [2], a new hybrid approach is proposed for short-term wind power perdition using ANFIS with two hierarchical stages. It was compared the mean absolute percentage error (MAPE), the sum squared error (SSE), the root mean squared error (RMSE), and the standard deviation of error (SDE) for every season of the year. The average MAPE was 8.1133% giving this next day forecast a great positioning regarding performance when compared with other forecasting methods [2]. 2.2 Neural Network Base Algorithms Wavelet Theory Wavelet Transform is a new tool for time-frequency analysis. Multi-resolution approximation by wavelet basis functions is a technique for representing a function on many different scales, which are formed by scaled and translated mother wavelet. The Continuous Wavelet Transform (CWT) of a signal x(t) is defined as (Eq. 1) [3].    +∞ t−b 1 dt = x(t), ψa,b(t)  x(t)ψ ∗ (1) WT x (a, b) = √ a a −∞ Where ψ(t) is the mother wavelet, and other wavelets:   t−b 1 ψa,b(t) = √ ψ a a

(2)

Are its dilated and translated versions, where a and b are the dilation parameter and translation parameter respectively. Wavelet theories use the mathematical method Wavelet Transform, using this, it is decomposing the original signal into several time series that have simpler frequency components, achieving a new level of analysis of the initial wave. The wavelet theory has a continuous (Eq. 1, Eq. 2) and a discrete form, being the last used in the practical applications. To break a signal into many lower resolution components the decomposition process used is the wavelet decomposition tree (Fig. 3).

726

R. Puga et al.

Fig. 3. Mallat wavelet decomposition tree [3]

This method aligned with an artificial neuronal network based on was phase space reconstruction method was used to predict the power output in the wind turbine. The results achieved are better than ARMA method [3]. The max power of the wind turbine during the test was around 850 KW. Table 1. MAE of power output prediction for wind turbine 1 (KW) [3] Model

1 h ahead

3 h ahead

ARMA

60.97

120.10

No WT

58.58

110.88

Hybrid

55.63

111.04

Table 2. MAE of power output prediction for wind turbine 2 (KW) [3] Model

1 h ahead

3 h ahead

ARMA

3.20

5.22

No WT

3.15

5.35

Hybrid

2.97

5.21

The principal’s changes compared the prediction results of the Table 1 and Table 2 can have been caused by the alteration in the amount of dataset for each case. The wind turbine 1 (Table 1) had 1 month of training data and 1 month of data for testing, on the other hand the wind turbine 2 (Table 2) had 3 months of training data and 3 months of data for testing [3]. Back Propagation Neural Network The traditional prediction process is based on a training data set and a testing part, the results of these predictions usually don’t have any error minimization process being the first algorithm iteration the final one [4]. The back propagation neural model is a multi-layer feed- forward neural network with an error back propagation training algorithm being able to approximate any nonlinear mapping using mentor training method [4].

State of the Art of Wind and Power Prediction for Wind Farms

727

Back Propagation (BP) Neural Network with Error Data The back propagation neural model for this study has one hidden layer (Fig. 4).

Fig. 4. BP neural network with one hidden layer

The input layer has M neurons and each of these neuros is expressed by m. The hidden layer has I neurons and each of these neurons is expressed by I. The output layer has J neurons and each neuros is expressed by j 4. The synaptic weight of input layer and hidden layer is expressed by wmi (m = 1, 2, … , M; i = 1, 2, … , I), the synaptic weight of hidden layer and output layer is expressed by wij (i = 1, 2, … , I; j = 1, 2, … , J). The output of each neuron is:   I M (3) ykj(n) = wij (n)ϕi wmi (n)xkm i=1

m=1

The error signal of each neuron of the output layer is: ekj (n) = dkj (n) − ykj (n)

(4)

The data was divided in 3 data sets, main data training, error data training and test. The results of the improved back propagation neural network method had a bigger accuracy than the normal method (Fig. 5). To test this theory another prediction base algorithm was used, Support vector machine [4]. The improved method showed better results (Table 3) when compared with the normal method. Back Propagation Neural Network Aligned with Storage System The goal of the wind prediction is ultimately, in my study, to achieve a better power supply to the grid from energy provenience from wind power. A way to achieve this goal is to have Energy Storage System (ESS) to be able to store this resource when the production is above a certain level as show in the Fig. 6. It is obtained that Pw = Pes + Pg based on power equilibrium. Where Pw is the output of wind farm, Pes is the observing energy of the ESS and the Pg is the power inject to Grid. A study based on back propagation model with six input variables and one hidden layer was created to predict the power output of a wind turbine [5].

728

R. Puga et al.

Fig. 5. Improved method with BP as base [4] Table 3. Comparison between improved and normal method with BP as base [4] Method

MAE/kW MRE/% RMSE/kW Time/s

Traditional 44.75 pred.

30

56.48

65.34

Improved pred.

22

34.32

86.27

28.56

Fig. 6. Scheme of wind power structure with power storage [5].

Fig. 7. Curve of the real power and the predicted power [5].

Fig. 8. Curves of power grid [5]

State of the Art of Wind and Power Prediction for Wind Farms

729

As we can see in the Fig. 7 the wind prediction of the wind power output is poorly accurate but even with a low-quality prediction quality, we are able to output a smooth power to the grid as we can see in the Fig. 8. The improvement of the accuracy of the wind prediction would provide the possibility of a lower energy storage by the system and a bigger power output to the grid. Back Propagation Neural Network with ARMA The expression of the zero mean stationary sequence for ARMA (p, q) method is: ϕ(B)xt = θ (B)at

(5)

Where ϕi, θi, I = 1, … , p and j = 1, … , q. The autoregressive polynomial is ϕ(B) = 1 − ϕ1 (B) − · · · − ϕp (B)

(6)

The moving average polynomial is ϕ(B) = 1 − θ1 (B) − · · · − θp (B)

(7)

In this work a BP neural network residual correction is approach is studied aligned with the ARMA method. The data for the network is from a wind farm with a total installed capacity of 17.56 MW. The data is submitted to a differential processing where afterwards the ARMA model is applied. Secondly is established the BP neural network and the prediction value or ARMA is superimposed on the output of the network being the final prediction achieved [6]. To best compare the results of the method a side-by-side analysis is made between the full model and ARMA method only as in the Table 4. The mean absolute error and mean square error are 0,1880 and 0,0468 for the complete method, when compare with the ARMA method the result has an improvement of 26.83%, 27.02% and 1.42% for MAE, MSE and MAPE respectively. Table 4. Table with the model errors [6] Error Índex

Arma (4, 5)

BP-ARMA

MAE

0.4563

0.1880

MSE

0.3170

0.0468

MAPE

0.0154

0.0012

Grey Model The wind power prediction model shown in Fig. 9 is supported by the grey model. The grey model is used for systems with a lot of uncertain and poor information, with low data requirements is suited for wind prediction. On the other hand, it has a low prediction accuracy. For this study in analysis where used an improved grey model, a discrete grey model and fraction grey model [7].

730

R. Puga et al.

Fig. 9. Power prediction model [7]

To combine all these inputs was used a neuronal network since it has the perks of self-adaptation and self-learning being a fit for grey information models. The results of the study were bench marked comparing to Auto-regressive Integrated Moving Average (ARIMA). The error of grey combination model for wind power prediction is shown in Fig. 10.

Fig. 10. Results of power prediction model [7]

Table 5 shows that the “combination” has great results, comparing the results with the benchmark, ARIMA, the improvement of the power output prediction for MAE, MAPE and RMSE was 37.7%, 34.9% and 34.4%, respectively [7].

State of the Art of Wind and Power Prediction for Wind Farms

731

Table 5. Wind speed prediction error of different models results of power prediction model [7]. Improved DGM 1/2 order 1/4 order 2/3 order Combination ARIMA GM MAE (kW)

76.9

89.4

106.6

101.2

104.6

43.4

69.7

MAPE %

15.7

18

21.1

19.7

20 .8

9.7

14.9

129.6

143.7

142.3

139.7

54.2

82.6

RMSE (kW) 109.4

3 Conclusion and Future Work We describe the Fuzzy Logic base Algorithms, Wavelet Theory, Back Propagation Neural Network, Back Propagation Neural Network with ARMA. Fuzzy Logic base Algorithms presented in [2], a new hybrid approach is proposed for short-term wind power perdition using ANFIS with two hierarchical stages. It was compared the mean absolute percentage error (MAPE), the sum squared error (SSE), the root mean squared error (RMSE), and the standard deviation of error (SDE) for every season of the year. The average MAPE was 8.1133% giving this next day forecast a great positioning regarding performance when compared with other fore-casting methods [2]. The wavelet theory aligned with an artificial neuronal network based on was phase space reconstruction method was used to predict the power output in the wind turbine. The results achieved are better than ARMA method [3]. We conclude too that used a combination of the grey model, with discrete grey model and the fraction grey model this “combination” has great results, comparing the results with the benchmark, ARMIA, the improvement of the power output prediction for MAE, MAPE and RMSE was 37.7%, 34.9% and 34.4%, respectively [7]. There are other techniques that could be used but cannot be presented in this paper, such as: E. Genetic Algorithms [8]; F. Empirical Mode Decomposition [9]; Random Theory Base Algorithms; Computational Fluid Dynamics Method Base Algorithms [10], etc.

References 1. Ji, G., Dong, Z., Wang, D.-F., Han, P., Xu, D-P.: Wind speed conformal prediction in wind farm based on algorithmic randomness theory, July 2008 2. Eseye, T., Zhang, J., Zheng, D., Ma, H., Jingfu, G.: A double-stage hierarchical ANFIS model for short-term wind power prediction. In: 2017 IEEE 2nd International Conference on Big Data Analysis (ICBDA), Beijing, China, pp. 546–551 (2017). https://doi.org/10.1109/ ICBDA.2017.8078694 3. Wang, L., Dong, L., Hao, Y., Liao, X.: Wind power prediction using wavelet transform and chaotic characteristics. In: WNWEC 2009 - 2009 World Non-Grid-Connected Wind Power and Energy Conference, pp. 1–5 (2009). https://doi.org/10.1109/WNWEC.2009.5335780 4. Mao, M., Cao, Y., Chang, L.: Improved fast short-term wind power prediction model based on superposition of predicted error. In: 2013 4th IEEE International Symposium on Power Electronics for Distributed Generation Systems (PEDG), pp. 1–6 (2013)

732

R. Puga et al.

5. Xinyu, Z., Lei, D.: A smooth scheme of wind power generation based on wind power prediction. In: Proceedings 2011 International Conference on Transportation, Mechanical, and Electrical Engineering (TMEE), pp. 958–961 (2011) 6. Chenhong, Z., Penghui, W., Yuan, Z., Yagang, Z.: Wind speed prediction research based on time series model with residual correction. In: 2017 2nd International Conference on Power and Renewable Energy (ICPRE), Chengdu, China, pp. 466–470 (2017) 7. Zhang, Y., Sun, H., Guo, K.Y.: Wind power prediction based on PSO-SVR and grey combination model. IEEE Open Access J. (2019) 8. Vaitheeswaran, S.S., Ventrapragada, V.R.: Wind Power Pattern Prediction in time series measurement data for wind energy prediction modelling using LSTM-GA networks. In: 2019 10th International Conference on Computing, Communication and Networking Technologies (ICCCNT), pp. 1–5 (2019) 9. Zhou, B., Sun, B., Gong, X., Liu, C.: Ultra-short-term prediction of wind power based on EMD and DLSTM. In: 2019 14th IEEE Conference on Industrial Electronics and Applications (ICIEA), Xi’an, China, pp. 1909–1913 (2019) 10. Li, L., Wang, Y., Liu, Y.: Wind velocity prediction at wind turbine hub height based on CFD model. In: International Conference on Materials for Renewable Energy and Environment (ICMREE) (2013)

State of the Art on Advanced Control of Electric Energy Transformation to Hydrogen Ricardo Puga1 , José Boaventura1,2 , Judite Ferreira3,4(B) , and Ana Madureira3,4 1 Universidade de Trás-os-Montes e Alto Douro (UTAD), Vila Real, Portugal

[email protected], [email protected]

2 Centre for Robotics in Industry and Intelligent Systems, INESC TEC, Porto, Portugal 3 ISEP/IPP, Porto, Portugal

{mju,amd}@isep.ipp.pt 4 ISRC - Interdisciplinary Studies Research Center, Porto, Portugal

Abstract. The need for sustainable power production has led to the development of more innovative approaches to production and storage. In light of this hydrogen production through wind power has emerged as sufficient in ensuring that the objectives of the Paris Agreement are made. This paper discusses the state-ofart models and controls used in ensuring that greater efficiency is achieved in the processes of energy to hydrogen transformation. The paper concludes with a comparison of the models and determination of one which suffices in ensuring that hydrogen/energy transformation is more efficient. Keywords: Hydrogen · Wind · Control modelling · Mathematical model · Integrated model

1 Introduction After the Paris Agreement, the development of more sustainable energy technologies has become critical and necessary in ensuring that the actions of climate change are tacked. As such, renewable hydrogen has emerged as one of the most promising and potential medium of energy storage. This means that it can be employed to guarantee the security of energy systems and also to cover the energy demands during the low availability periods of VREs which include wind and solar power [1]. 1.1 Principles The main principles of a hydrogen production technology by wind power system include; the hydrogen produced through electrolysis will not be carried by the grid, rather, it goes directly to storage and is produced by the power generated by the wind turbines, photovoltaic, nuclear and others [2, 3, 4]. Secondly, the energy possibly lost initially through wind curtailment should maximally be absorbed. It is based in this principle that periods of insufficient wind are be managed by having the grid connection function as the source © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 733–742, 2022. https://doi.org/10.1007/978-3-030-96299-9_69

734

R. Puga et al.

of backup electricity. This ensures that the electrolyze always functions at full capacity [5]. Finally, the excessive electric power provided by the turbines should be used for hydrogen production through the electrolysis process. This is done as an alternative to curtailing electricity, an approach commonly employed in traditional production processes. 1.2 Added Processes Processes added to the system include the pressurization of hydrogen before it is stored. When pressurized, hydrogen can be used in a range of applications. Enhanced efficiency of hydrogen fuel cells can ensure that the needs pertaining to power supply in most industries and residential areas are addressed [6–8]. It is worthwhile to know that currently, technologies that facilitate hydrogen production are not mature. The reason for this is that there still exists significant challenges in enhancing the efficiency of most aspects of the production technologies. A significant production approach employed today is that recognized as the offgrid technology of wind power generation [9]. Through it, the possible impacts that wind generated power can bring to a grid system will be eliminated due to an indirect connection [10]. It also guarantees that better efficiency is achieved because wind power is utilized at a lower cost. The voltage from transformers is adjusted to the voltage required and rectifications made on the alternating current to direct current. 1.3 Production Costs It is critical to recognize that costs incurred in hydrogen production are significantly less compared to when the connection between the unit and the grid is direct [11]. Offgrid systems of power generation are more efficient especially when problems related to power consumption are considered. The generated hydrogen is also clean and renewable. 1.4 Hydrogen to Electric Power Conversion For hydrogen to be converted back to electric energy, fuel cells are currently the most used tools. They are electrochemical devices that function as alternative sources of power because they are effective in converting hydrogen to electric power. The cells produce electricity by effecting a reaction between hydrogen and air [2]. In addition to the above type of fuel cell, there also is another regarded as the Polymer Electrolyte Membrane (PEM) fuel cells. As the name suggests, these cells function based on the principle of electrochemical reaction [3]. The reaction occurs between oxygen and hydrogen and is aided with a catalyst such as platinum. In the cells, hydrogen is channelled through field flow plates which exist on the anode side of the cell. On the other hand, oxygen from the atmosphere is passed through the cathode side of the cell. The platinum catalyst at the positive cell side helps to create hydrogen electrons and protons. The polymer electrolyte membrane (PEM) type of fuel cells can produce greater densities of current in addition to having fast start-up and shutdown times. They also operate at temperatures between 60 to 80 °C.

State of the Art on Advanced Control

735

When the Alkaline and PEM electrolysers are compared, it can be noted that the previous technology of electric power generation from hydrogen is already quite mature. Nevertheless, the PEM technologies are emerging rapidly with constant reductions in their Capital Expenditures (CAPEX) being constantly reported [3]. It also is important to recognize that the general lifetime of the alkaline (ALK) electrolyser currently is more than twice as long and might even achieve greater life in the coming decades [12–16]. On the other hand, the state-of-the-art PEM electrolysers are highlighted to have a more flexible operation capability and reactivity that is greater than that currently achieved in the ALK technologies. Both systems, however, can be maintained with little power consumption in standby mode. They also can operate for short periods and at a greater capacity compared to the nominal load. PEM electrolysers, for example, can offer greater Frequency Containment Reserve (FCR) without having to sacrifice the capacity of production available due to the upward and downward regulation capabilities.

2 Modelling and Control Methodologies 2.1 Generic Systems Models The efficient transformation of hydrogen to energy and vice versa depends on the effectiveness of the modelling approach used. In the generic systems models, the systems typically will include four primary component pasts. The first is the use of renewable energy technology as the primary energy source. In the case of wind, weather data is always needed for the most appropriate control mechanisms to be determined. The second element of the model is an electrolyser that is used for hydrogen production [17]. Excess energy generated by renewables is used to ensure that hydrogen and oxygen are produced from supplied water. The third element is hydrogen from the electrolyser which is stored in pressurized gas tanks. It is also possible for a compressor to be included. The fourth element is a fuel cell that transforms the stored hydrogen to electric energy whenever the renewable power supplied to the grid is not sufficient [17]. Fuel cells often can be replaced by hydrogen engines. For the electrolyser to be connected to the fuel cell, and finally, to the controller of the entire system, power conditioning components must be included. It also is typical for the systems to include the data I/O components. The aim of this work is to provide boundary conditions to the model and also for results extraction. Data extracted can be used both for economic analysis of each system and also in the process of optimization. 2.2 Mathematical Models Mathematical modeling is a primary example of a method used in determining the efficiency of the hydrogen to energy transformation process. Within it, the advanced alkaline electrolyser model is included. It is based on thermodynamics, transfer of heat, and the empirical electrochemical equations [18]. It is presumed that splitting water into oxygen and hydrogen is made possible when an electric current is directly passed between two electrodes with an aqueous electrolyte separating them. The electrolyte typically has a good ionic conductivity example being the aqueous potassium hydroxide.

736

R. Puga et al.

The potassium positive ions and hydroxide’s negative ions facilitate the molecule movements. Kinetics in the electrodes of an electrolyser cell is modelled based on the U-I relationships. This means that ohmic resistance and excess voltages are significant in ensuring that I-U curves are created. ‘U’, represents Voltage and current is represented by ‘I’. As per the law developed by Faraday, hydrogen’s rate of production within an electrolyser cell is equal to the electric current. Nevertheless, energy efficiency can only be stated to exist in the event less heat is generated. For this reason, it is calculated from the cell and thermoneutral voltage. Every given rise in hydrogen production at a provided temperature leads to a rise in the cell voltage level. Even so, at a specific density of the current, energy efficiency will increase with the rise in the cell’s temperature. Finally, the compressed gas model is a mathematical model under which the pressure within the storage is calculated either based on the Ideal Gas Law or van der Waal’s equation [19–21]. The model simply conducts a mass balance between the gas getting into the storage and that leaving it. The pressure that matches with the H2 mass in the tank is hence calculated to determine whether it is above a specified level and if so, excess hydrogen has to be pumped out and converted back to energy. 2.3 Data-Driven Control-Oriented Models for PEM Fuel Cells The model includes the use of machine learning and other forms of regression algorithms to create a more robust controls-oriented model. Performance typically is compared using the Explained Variance score ratio. This is k-fold methodology can be used to crossvalidate a model using 5-folds. A fixed seed is established to guarantee reproducibility [22]. When data is presented on a graph, it is possible to observe that the algorithm representing the least variation and greater accuracy is the gradient boosting aggressor. Nevertheless, such an aggressor only will reach a score of 0.86 which is somehow less for control. Within the same fuel cell modeling approach, neural networks can also be considered. These are categorized as per their respective behaviors in time. They either can be static or dynamic, the previous having the capacity to model with greater accuracy the PEM fuel cell’s performance. Even so, time as a variable negatively impacts the output voltage and also the current even when placed in steady-state conditions [22–25]. When neural networks are dynamic, they also will consider time and hence their structures can be used as system control generic models. The model also includes the approach regarded as artificial neural network and fuzzy logic used whenever adding heuristic rules ends up causing difficulties. Within this formulation, electrolysers and the fuel cell can achieve various regimes of operation including working at a partial load or more steady power, hence ensuring that a system can achieve greater efficiency levels [25]. The feed-forward neural networks and other intelligent algorithms guarantee that the control includes forecasts on how to enhance a plant’s performance.

State of the Art on Advanced Control

737

2.4 Integrated Model The integrated modelling approach involves several models put together to facilitate the functioning of a system. In this case, the model includes steam forming which is designed as a Continuous Stirred-Tank reactor (CSTR) whose operation is ideal. The mechanism of the reaction uses the Ni-AL catalyst and is describe with competing reactions in the form of Equations [26–30]. The rates of reaction can also be calculated using a kinetic model. It is presumed that the ability of a PEM fuel cell to produce electricity is founded on reactions that occur on the catalytic anode and cathodes summarized as H2 = 2H+ + 2e−

(1)

0.5O2 + 2H+ + 2e− = H2 O

(2)

and

respectively [30]. A simple but dynamic model of a fuel cell regards the behaviours of electronic chemicals as parameterized while the semi-empirical model considers both the molar balances and those of hydrogen and oxygen on the positive and negative parts of the cell. Founding or guiding equations are those adopted from the research by Khan and Iqbal. Even so, two exceptions have to be considered. First, in the electro-chemical models, concentration overpotential is a term added in the equations [31]. In the integrated system of fuel processing, hydrogen simply is fed directly to the fuel cell from the Water Gas shift reaction (WGS) rather than using the hydrogen tank. For this reason, hydrogen’s balance equation will slightly deviate from what is represented in the study by Khan and Iqbal. As a modular dynamic model for an integrated fuel processing, it helps in facilitating the functions such as the formation of ethanol, purification of hydrogen through a WGS reactor membrane, and finally, a PEM fuel cell which helps convert hydrogen to electric power [31]. Control over fuel feed is done using an MPC to minimize delays and also ensure that fuel starvation is avoided. 2.5 Autoregressive Moving Average eXogenous (ARMAX) Model For control strategies to be created for the cell reactant flows, the ARMAX model often is used. From a systems point of view, hydrogen is considered as an input variable fed at a flow rate of NH that can be adjusted. Oxygen also is considered as a significant input and is represented by NA especially when the fuel cell uses the Oxygen air content. Current and voltage then are perceived as system outputs [32]. This also is represented a highly standardized MIMO system. The relationship between the flow of oxygen and the fuel cell’s voltage will be positive assuming that the flow of hydrogen is always constant. For the non-linear and time-differentiated dynamic identification of the fuel cell to be included, the ARMAX can be expanded to also include recursive least squares algorithm through which the identification proves can predict system output as per the information obtained previously. The model is linear but piecewise, it is non-linear meaning that it is characterized by temporary linearization [32]. It also is befitting to state that the model also includes an element of power forecasting. Forecasting has over time received

738

R. Puga et al.

significant attention because of its significance when it comes to planning the operations and efficiency of hydrogen-supported power grids. When compared to other techniques of forecasting, the model’s time series does not need the meteorological forecasts which often are complicated. Because of its simplicity, the model has widely been defined as a statistical model through which power from the hydrogen systems can be forecasted. Nevertheless, the model also is data-driven and hence does not consider climatic information [32]. This happens although the information is valuable in ensuring that forecast accuracy is improved. The model is flexible and more general when its practical use is considered. And for this reason, it significantly enhances forecast accuracy. 2.6 Compressor Model The model is separated into two distinct parts. The first includes a static compressor map that helps determine the rate of hydrogen flow through a compressor. Thermodynamic equations are then used to calculate the temperature of exit air and the compressor power required. In the second part, the inertia between the compressor and motor is represented in addition to the compressor speed’s definition. Consequently, the speed is used in the map to determine the flow rate of the air mass [31]. The compressor speed is the sole dynamic state included in the model. Its inputs include the pressure of inlet air, temperature, the compressor’s motor voltage command, and also the downstream pressure, which also is stated as the supply manifold pressure; Pcp, out = Psm. Inlet air, in this case, primarily is atmospheric. Its temperature and pressure are assumed to be a contestant at Patm = 1 and Tatm = 25 °C. One of the fuel system’s inputs is the motor command. The model of manifold supply is employed to determine the downward pressure. It also is significant to recognize that the air mass flow rate is ascertained through the compressor’s flow map. The considered elements include the ratio of pressure across the compressor and compressor speed. Nevertheless, it is not appropriate to supply the compressor flow map in the form of a look-up table especially when the dynamic simulations of a system are used. In addition to this exploration is unreliable and routines of standard interpolation typically are not differentiable. As such, the non-linear curve method can be used to model the characteristics of the compressor [18]. The compressor model is significant in ensuring that greater system efficiencies are achieved. The reason for this is that greater but balanced compressor energies are typically required to ensure that hydrogen tanks are appropriately filled because of heating that comes about in the course of fast fills. 2.7 Resume Table The above discussion introduces modelling approaches currently used to transform hydrogen to electric power. In this modelling Table 1, the approaches are compared in terms of their advantages and disadvantages. The comparison is meant to ensure that the most efficient or desirable approach is determined.

State of the Art on Advanced Control

739

Table 1. Comparison of modelling approach Modeling approach

Pros

Cons

Generic systems

Simple and straightforward Few components part considered Use of meteorological data

The system may require significant sums of sensors A single controller is used for the entire system

Mathematical models

Leads to the development of more advanced control strategies due to the equations developed. It considers different parameters including design, operational aspects. The equations consider other significant laws of mathematics including the Faraday law

Time-consuming The models are highly complex Some physical externalities can be missed

Data-driven control oriented models for PEM fuel cells

Integrates machine learning and regression algorithms. Integrates a five-fold approach to cross-validate the model features. It is less time consuming when compared to the fully mathematical models It is more accurate

They still focus on feedback control with fixed references. They have to be designed to consider interactions between different subsystems. In practice, it typically is required that the adaptiveness and robustness of the integrated controls are emphasized

Integrated models

It considers several models put together include kinetic models that can be used to calculate the rates of reaction

The model is typically quite complex to use. It is time-consuming due to the vast sum of variables that have to be considered. It requires a vast use of sensors to obtain data

ARMAX model

It is possible to forecast power from hydrogen systems does not need meteorological forecasts. It is flexible and more general when practicality is considered

Do not use climate information which is more accurate when wind power is considered. Characterized by temporary linearization which makes it complex Some models do not consider previously fed information (continued)

740

R. Puga et al. Table 1. (continued)

Modeling approach

Pros

Cons

Compressor model

Includes a compressor map and thermodynamic equations. The modeling approach guarantees that energy loss through heating is avoided

Only emphasizes the pressure of hydrogen as critical in energy production Data cannot be illustrated in a look-up table during simulation It assumes that the pressure and temperature of atmospheric air fed to hydrogen during transformation are always constant

3 Conclusion This paper discusses the different models used to ensure that hydrogen is effectively converted to energy and vice versa. In addition to enhancing the efficiency of the energy production systems, they also are primary in ensuring that system deficiencies are determined and corrected. From their comparison, the data-driven control models emerge as more effective when employed for PEM cells to reduce the control costs. To be specific, the integration of neural controls ensures that the highest levels of accuracy are achieved because of their ability to map out the complex relationships that evidence linearly. The modelling approach is sufficient because it is founded on the premise that the voltage provided by a fuel cell does not just depend on the reactant supply as well as the transient conditions; rather, it is the load that significantly affects the performance of fuel cells. It saves time when compared to the sole use of a mathematical model and also ensures that the sum of sensors and controls included in the system is reduced.

References 1. Rezaei, M., Khozani, N., Jafari, N.: Wind energy utilization for hydrogen production in an underdeveloped country: an economic investigation. Renew. Energy 147, 1044–1057 (2020) 2. Office of Energy Efficiency Renewable Energy. Hydrogen Production: Electrolysis. https:// www.energy.gov/eere/fuelcells/hydrogenproduction-electrolysis 3. Zhou, T., Francois, B.: Modeling and control design of hydrogen production process for an active hydrogen/wind hybrid power system. Int. J. Hydrogen Energy 34, 21–30 (2009) 4. Mostafaeipour, A., Qolipour, M., Goudarzi, H.: Feasibility of using wind turbines for renewable hydrogen production in Firuzkuh. Iran. Front. Energy 13(3), 494–505 (2019). https:// doi.org/10.1007/s11708-018-05346 5. Fernandez-Guillamon, A., Das, K., Cutululis, N.: Offshore wind power integration into future power systems: overview and trends. J. Mar. Sci. Eng. 7, 399 (2019). https://doi.org/10.3390/ jmse7110399

State of the Art on Advanced Control

741

6. IRENA: Hydrogen from renewable power: Technology outlook for the energy transition, International Renewable Energy Agency, Abu Dhabi (2018) 7. IRENA: Hydrogen: A renewable energy perspective, International Renewable Energy Agency, Abu Dhabi (2019) 8. Gondal, I., Masood, A., Khan, R.: Green hydrogen production potential for developing a hydrogen economy in Pakistan. Int. J. Hydrogen Energy 43, 6011–6039 (2018) 9. Fasihi, M., Breyer, C.: Baseload electricity and hydrogen supply based on hybrid PV-wind power plants. J. Clean. Prod. 243, 118466 (2020) 10. Zhang, Y., Sun, H., Guo, Y.: Integration design and operation strategy of multi-energy hybrid system including renewable energies, batteries, and hydrogen. Energies 13, 5463 (2020). https://doi.org/10.3390/en13205463 11. Li, Z., Guo, P., Ham, R., Sun, H.: Current status and development trend of wind power generation-based hydrogen production technology. Energy Explor. Exploit. 37(1), 5–25 (2019) 12. Mitra, P., Zhang, L., Harnefors, L.: Offshore wind integration to a weak grid by VSCsHVDC links using power-synchronization control: a case study. IEEE Trans. Power Deliv. 29, 453–461 (2013) 13. Reed, G.F., Al Hassan, H.A., Korytowski, M.J., Lewis, P.T., Grainger, B.M.: Comparison of HVAC and HVDC solutions for offshore wind farms with a procedure for system economic evaluation. In: Proceedings of the 2013 IEEE Energytech, Cleveland, OH, USA, 21–23 May 2013, pp. 1–7 (2013) 14. Guidi, G., Fosso, O.: Investment cost of HVAC cable reactive power compensation offshore. In: Proceedings of the 2012 IEEE International Energy Conference and Exhibition (ENERGYCON), Florence, Italy, 9–12 September 2012, pp. 299–304 (2012) 15. Hur, D.: Economic considerations underlying the adoption of HVDC and HVAC for the connection of an offshore wind farm in Korea. J. Electr. Eng. Technol. 7, 157–162 (2012) 16. Sharma, R., Rasmussen, T.W., Jensen, K.H., Akamatov, V.: Modular VSCs converter based HVDC power transmission from offshore wind power plant: compared to the conventional HVAC system. In: Proceedings of the 2010 IEEE Electrical Power Energy Conference, Halifax, NS, Canada, 25–27 August 2010, pp. 1–6 (2010) 17. Chen, H., Johnson, M.H., Aliprantis, D.C.: Low-frequency AC transmission for offshore wind power. IEEE Trans. Power Deliv. 28, 2236–2244 (2013) 18. Negra, N.B., Todorovic, J., Ackermann, T.: Loss evaluation of HVAC and HVDC transmission solutions for large offshore wind farms. Electr. Power Syst. Res. 76, 916–927 (2006) 19. Stoutenburg, E., Jacobson, M.: Optimizing offshore transmission links for marine renewable energy farms. In: Proceedings of the OCEANS 2010 MTS/IEEE SEATTLE, Seattle, WA, USA, 20–23 September 2010 20. Bresesti, P., Kling, W.L., Hendriks, R.L., Vailati, R.: HVDC connection of offshore wind farms to the transmission system. IEEE Trans. Energy Convers. 22, 37–43 (2007) 21. Gomis-Bellmunt, O., Liang, J., Ekanayake, J., King, R., Jenkins, N.: Topologies of multiterminal HVDC-VSCs transmission for large offshore wind farms. Electr. Power Syst. Res. 81, 271–281 (2011) 22. Legorburu, I., Johnson, K.R., Kerr, S.A.: Multi-use maritime platforms North Sea oil and offshore wind: opportunity and risk. Ocean Coast. Manag. 160, 75–85 (2018) 23. Zhang, Y., Ravishankar, J., Fletcher, J., Li, R., Han, M.: Review of modular multilevel converter based multi-terminal HVDC systems for offshore wind power transmission. Renew. Sustain. Energy Rev. 61, 572–586 (2016) 24. De Alegrıa, I.M., Martin, J.L., Kortabarria, I., Andreu, J., Ereno, P.I.: Transmission alternatives for offshore electrical power. Renew. Sustain. Energy Rev. 13, 1027–1038 (2009) 25. Rourke, F.O., Boyle, F., Reynolds, A.: Marine current energy devices: current status and possible future applications in Ireland. Renew. Sustain. Energy Rev. 14, 1026–1036 (2010)

742

R. Puga et al.

26. Chou, C.J., Wu, Y.K., Han, G.Y., Lee, C.Y.: Comparative evaluation of the HVDC and HVAC links integrated in a large offshore wind farm—An actual case study in Taiwan. IEEE Trans. Ind. Appl. 48, 1639–1648 (2012) 27. Colmenar-Santos, A., Perera-Perez, J., Borge-Diez, D., de Palacio Rodríguez, C.: Offshore wind energy: a review of the current status, challenges and future development in Spain. Renew. Sustain. Energy Rev. 64, 1–18 (2016) 28. Erlich, I., Shewarega, F., Feltes, C., Koch, F.W., Fortmann, J.: Offshore wind power generation technologies. Proc. IEEE 101, 891–905 (2013) 29. Segura, I., Perez-Navarro, A., Sanchez, C., Ibañ, E.Z.F., Payá, J., Bernal, E.: Technical requirements for economical viability of electricity generation in stabilized wind parks. Hydrog. Energy 32(16), 3811–3819 (2007) 30. Fernandez, R.D., Battaiotto, P.E., Mantz, R.J.: Impact of wind farms voltage regulation on the stability of the network frequency. Hydrog. Energy 33(13), 3543–3548 (2008) 31. Thanapalan, K., Guwy, P.: Model-based controller design for hydrogen fuel cell systems. In: The Proceedings of IFAC World Congress 2008, Seoul, Korea, pp. 4636–4641 (2008) 32. Yanting, L., Yan, S., Lianjie, S.: An ARMAX model for forecasting the power output of a grid-connected photovoltaic system. Renew. Energy 66(C), 78–89 (2014)

Dynamic Modelling of a Thermal Solar Heating System José Boaventura-Cunha1,2(B) and Judite Ferreira3,4 1 Universidade de Trás-os-Montes e Alto Douro (UTAD), Vila Real, Portugal

[email protected]

2 Centre for Robotics in Industry and Intelligent Systems, INESC TEC, Porto, Portugal 3 ISEP/IPP, Porto, Portugal

[email protected] 4 ISRC - Interdisciplinary Studies Research Center, Porto, Portugal

Abstract. Nowadays the world faces the challenge to rapidly diminish the use of fossil fuels in order to reduce pollutants and the emission of greenhouse gases and to mitigate the global warming. Renewable energies, such as solar radiation, among others, are playing a relevant role in this context. Namely, the use of thermal energy storage systems in buildings and industry is increasing enabling to reduce operational costs and carbon dioxide emissions. Heat storage systems based in solar thermal panels for heating water in buildings are industrially mature but some improvements can be made to improve their efficiencies. In this work are presented the methods and the results achieved to model the dynamic behavior of the hot water temperature as function of the weather, operating conditions and technical parameters of the thermal solar system. This type of dynamic models will enable to optimize the efficiency of this type of systems regarding the use of auxiliary energy sources to heat the water whenever the temperature in the storage tank falls below a defined threshold level. As future work it is intended to use adaptive control algorithms to reduce the use of backup power sources (electricity, oil, gas) by using the information of the system status as well predictions for hot water consumption profiles and solar radiation. Keywords: Adaptive control · Modeling · Predictive control · Renewable energy · Solar radiation · Thermal solar panel

1 Introduction Renewable energy sources are playing a relevant role in reducing the fossil-based energy consumption in buildings and so in the pollutants and carbon dioxide emissions. However, at the present there is a potential to increase the number of thermal energy storage systems in buildings and industry facilities as well to improve the controllers used to operate the backup power sources. It was estimated that in Europe around 1.4 MGWh/year can be saved and 400 Mtons of carbon dioxide emissions reduced by more extensive use of thermal storage systems [1]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 743–750, 2022. https://doi.org/10.1007/978-3-030-96299-9_70

744

J. Boaventura-Cunha and J. Ferreira

The main objective of this work is related with the modeling and control of solar thermal systems for heating domestic water. These systems use solar collectors where the incident solar radiation provides the heating of fluids that circulates through a pipe closed circuit linked to a heat-exchanger placed in a storage tank. This circulation of the fluid is done when there is solar radiation incidence (day) in the absorber zone and if the water temperature in the storage tank is less than the temperature of the circulating fluid in the solar collector. Normally the collector is black to achieve a better solar radiation absorption and so a better heat transfer to the fluid [2–5]. There are different technologies of solar collectors that can be used to achieve the mentioned heat transfer process. Generally, in domestic applications are used flat plate stationary collectors oriented to south with an angle regarding the horizontal plane that is equal to the latitude of the location. Collectors with other designs and materials and/or with control orientation of one or two axis are more expensive and restricted to industrial applications [2, 5, 6]. Table 1 show different technologies organized by different categories regarding stationary collectors where the concentration indexes of the collectors are defined as the ratio between the opening area and the absorber area. Table 1. Types of solar stationary collectors Collector

Concentration index

Operating temperature

FPC-Flat Plate Collector

1

30–80 °C

CPC-Compound Parabolic Collector

1–5

60–240 °C

ETC-Evacuated Tube Collector

1

50–200 °C

As mentioned, the aim of this work is to model and improve the control of a thermal solar system for domestic use which main components are: a FPC collector, a storage tank, a heat exchanger, a backup energy power source and a circulating pump, as showed in Fig. 1. The efficiency of the collector efficiency is defined as the rate of thermal energy used against incident solar radiation (Eq. 1) and depends on several factors [7–10]. For instance, the transmittance τ(θ) of a glass cover and the absorptance α(θ) of the black plate for solar radiation are dependent on the angle of incidence of the Sun radiation in the solar panel, θ. These parameters range from 0 to 0.9 and 0 to 0.92, respectively, when solar radiation is parallel or perpendicular to the solar panel, and influences the heat power absorbed. Also, there are thermal losses that are dependent of the weather conditions (air temperature, wind speed) and the insulation materials used in the panel. n=

QN SR

(1)

Dynamic Modelling of a Thermal Solar Heating System

745

Fig. 1. Schematics of a thermal solar system for domestic use (a) and of a FPC collector (b). Adapted from www.alternative-energy-tutorials.com/solar-hot-water/flat-plate-collector.html

where QN = EN -QV represents the available thermal power (W/m2 ), EN absorbed solar radiation (which depends on the glass transmission and absorption coefficients of the panel collector, QV the thermal losses and SR the radiation that reaches the glass cover (W/m2 ). The thermal losses (Eq. 2) is a function of the difference between the temperatures of the absorptive surface (TAS ) and the outside air (Tair ) and the heat transfer coefficient of the collector, UL (W/m2 °C). QV = UL (TAS − Tair )

(2)

Losses also occur in the storage tank and in the closed circulating pipes of the fluid, but in general these heat losses could be neglected if a proper thermal insulation is done. To compute the water temperature in the storage tank (TWS ), and assuming that no stratification occurs, it can be used the following equation: TWS =

Qheat mCH2O

(3)

where Qheat is the amount of transferred heat to the water (Wh), m is the mass of water to be heated (kg) and CH2O denotes the specific heat of the water, 1.16 (Wh/kg°C). In the next section it will be used this simplified approach to compute the water temperature in the storage tank as function of the weather conditions and hot water consumption.

2 Materials and Methods Here, are presented the hardware and software tools used to model a domestic thermal solar system as depicted in Fig. 1. The actuators used in the system are a fluid pump (250 W nominal power) and an auxiliary electrical backup heater (nominal power of 3, 5 KW) that is turned on to heat water when the water temperature is below 45 °C, with a hysteresis of 1 °C, by using a thermostat. The operation of this heating resistor

746

J. Boaventura-Cunha and J. Ferreira

was later modified with the aim of improving the efficiency of the system. Namely, the actuation state on was modified taken in account the predictions of the solar radiation and water temperature in order to reduce the amount of the electric energy supplied. The fluid closed circuit pump is activated by a PLC controller whenever the closed fluid temperature is greater or equal to TWSB + 4 (°C), being TWSB the water temperature in bottom of the storage tank. The pump is turned off when TWSB reaches 60 °C. The measurements of air and water temperatures were made using PT100 sensors and the solar radiation was measured with a SPN1 pyranometer sensor from delta-T devices. The sensors were connected to a data acquisition system linked to a computer via USB. The algorithms to model the panel solar system and to perform the predictive control of the auxiliary energy system were performed in the computer using a time sampling of 1 min and a storage time of 5 min. The developed methods can be applied to distinct thermal solar systems that use Flat Plate Collectors. The results presented were achieved using a commercial thermal solar system composed by 2 Vulcano solar panels, model FKT - 1S/1 W, with an effective area of 2 * 2.3 m2 , an optical transmittance of 0.9 and an absorption coefficient of 0.92. The hot water storage tank as a volume of 400L and the coefficient of heat losses is 3.1 WºC−1 m−2 . The methods employed to predict the solar radiation are detailed in the works [11, 12]. The algorithm applied to model the water temperature which can be used in the future to perform the predictive control of the auxiliary heater system is showed in Fig. 2. Input technical data of the thermal solar system

Read sensors (air and water temperatures, solar radiaƟon, heat water consumpƟon…)

Compute predicƟons for the water temperature (based on sensor readings, actuator states and thermal models)

Decision to turn on or oŏhe auxiliary heater actuator

Fig. 2. Algorithm used to model TWS and to control the auxiliary electrical heater.

Besides the mentioned parameters and coefficients used to predict TWS over a future horizon of 1 h it was used the hot water consumptions mean values for each of the weekdays. The following table shows a typical mean consumption flow of the hot water (L/min) over time for a weekday.

Dynamic Modelling of a Thermal Solar Heating System

747

Table 2. Typical consumption for a weekday in march Mean output flow of heat water from the storage tank

Hours of consumption

6 L/min

07 h 15 to 08 h 00

3 L/min

13 h 00 to 14 h 00

3 L/min

19 h 30 to 20 h 30

3 Results and Conclusions In this section are presented the results achieved for typical weekday periods of march and june 2020 using the described solar thermal system installed in a building located in Vila Real, Portugal. The data used in the following simulations and control strategies are showed in Fig. 3 for a 5 days period regarding the outside air temperature, global solar radiation measured over a horizontal plane and hot water consumption displayed in Table 2. During these 2 time periods the mean temperatures of the water supplied by the network were 10.5 °C and 19 °C, respectively.

Fig. 3. Relevant weather variables and hot water consumption measured from 9 to 13 march 2020.

In Fig. 4 are plotted the results achieved by applying the mentioned techniques and models. As it can be observed the simulation models describe well the dynamic behavior of the temperature of the hot water, being the mean absolute error less than 2.4 °C.

748

J. Boaventura-Cunha and J. Ferreira

Fig. 4. Water temperature computed with the panel solar model (black) and measured (red) and energies supplied to the water in the storage tank by the solar panel and the electric backup system from 9 to 13 march 2020.

Figures 5 and 6 show the variables measured and simulated for the time period from 7 to 11 of june 2020. Also, in this case it can be observed that the simulation models describe well the dynamic behavior of the temperature of the hot water, being the mean absolute error less than 2.2 °C. As future work it is intended to implement real-time adaptive control algorithms to reduce the use of backup power sources (electricity, oil, gas) by using the information of the system status as well on-line algorithms to predict hot water consumption profiles and the solar radiation [11, 12]. Simulation results of the adaptive controller based on the presented model and using the predictions of solar radiation and water consumption to minimize the cost function, J, showed that it can be achieved a reduction in the consumption of the backup energy system of about 11%. J=

hp i=1

a1 [ε(k + i)]2 +

hu i=1

a2 [u(k + i)]2

(4)

Where 1 is the predicted error, i.e., the difference between the simulated and desired temperatures of the hot water, u, the control effort, hp = 12 the prediction horizon, hu = 2 the control horizon and a1 = a2 = 0.5 the weighing factors.

Dynamic Modelling of a Thermal Solar Heating System

749

Fig. 5. Relevant weather variables and hot water consumption measured from 7 to 11 june 2020.

Fig. 6. Water temperature computed with the panel solar model (black) and measured (red) and energies supplied to the water in the storage tank by the solar panel and the electric backup system from 7 to 11 june 2020.

750

J. Boaventura-Cunha and J. Ferreira

Acknowledgements. This work is financed by National Funds through the Portuguese funding agency, FCT - Fundação para a Ciência e a Tecnologia, within project UIDB/50014/2020.

References 1. Sarbu, I., Sebarchievici, C.A.: Comprehensive review of thermal energy storage. Sustainability 10, 191 (2018) 2. Kalogirou, S.A.: Solar thermal collectors and applications. Progress Energy Combust Sci. 30(3), 231–295 (2004) 3. Kalogirou, S.: Solar Energy Engineering: Process and Systems. Elsevier, London (2009) 4. Kicsiny, R., Farkas, I.: Improved differential control for solar heating systems. Sol. Energy 86(11), 3489–3498 (2012) 5. Kosti´c, L.T., Pavlovi´c, Z.T.: Optimal position of flat plate reflectors of solar thermal collector. Energy Build. 45, 161–168 (2012) 6. Ahmad, L., Khordehgah, N., Malinauskaite, J., Jouhar, H.: Recent advances and applications of solar photovoltaics and thermal technologies. Energy 207, 118254 (2020) 7. Exell, R.: Flat-Plate Solar Collectors. King Mongkut’s University of Technology Thonburi (2000) 8. Kovacs, P.: Quality assurance in solar heating and cooling technology - a guide to the standard EN 12795. SP – technical research institute of Sweden (2012) 9. Zambolin, E., Col, D.D.: An improved procedure for the experimental characterization of optical efficiency in evacuated tube solar collectors. Renewable Energy 43, 37–46 (2012) 10. Zambolin, E., Col, D.D.: Experimental analysis of thermal performance of flat plate and evacuated tube solar collectors in stationary standard and daily conditions. Sol. Energy 84(8), 1382–1396 (2010) 11. Boaventura-Cunha, J., Moura Oliveira, P.B.: Solar radiation prediction using classical and evolutionary techniques. In: Computación Aplicada a la Industria de Processos, CAIP’2001, Campos de Jordão, Brasil (2001) 12. Coelho, J.P., Pinho, T., Boaventura-Cunha, J.: Hidden Markov Models: Theory and Implementation using Matlab, CRC Press, Inc. (2019)

A Review of Unpredictable Renewable Energy Sources Through Electric Vehicles on Islands Juliana Chavez1 , Jo˜ ao Soares1 , Zita Vale2 , Bruno Canizes1 , and S´ergio Ramos1(B) 1

GECAD - Research Group on Intelligent Engineering and Computing for Advanced Innovation and Development Polytechnic of Porto, School of Engineering (ISEP), Rua Dr. Antonio Bernardino de Almeida, 431, 4200-072 Porto, Portugal {chave,jan,bmc,scr}@isep.ipp.pt 2 Polytechnic of Porto, School of Engineering (ISEP), Rua Dr. Antonio Bernardino de Almeida, 431, 4200-072 Porto, Portugal [email protected]

Abstract. The development of new technologies such as renewable energy sources, energy storage devices, and electric vehicles has changed the structure of the distribution grid to an active grid with bidirectional power flow. This paper introduces the concept of a smart island and its challenges and an isolated smart microgrid mode. Furthermore, this review provides an overview of the challenges that the penetration of electric vehicles has on islands through various types of charging modes. Finally, it is analyzed the impact of vehicle-to-grid on islands and the decarbonization of transportation sectors. Keywords: Electric vehicles · Micro grid · Plug-in electric vehicles Renewable energy sources · Smart island · Vehicle-to-grid

1

·

Introduction

Islands play a crucial role because they have been described as ideal locations to develop and test new strategies and solutions that will drive the transition to the continent. Islands on small geographic proximity provide the potential for the development of 100% renewable island energy systems by developing their grid interconnections [1]. In the past, since power plants were utterly manageable while the load was unpredictable, the flexibility of the grid was provided by traditional power plants. Nowadays, due to the variety of renewable energy sources (RES), variability and unpredictability shifted to the generation side, and the opposite change happened to the flexibility agents. One possible solution to solve this issue is to introduce electric vehicles (EVs). Day by day, the number of electric vehicles on the road has been increasing. However, to meet the fundamental objective of EVs, reducing air pollution, reducing fossil fuel dependency, electricity storage, c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 751–760, 2022. https://doi.org/10.1007/978-3-030-96299-9_71

752

J. Chavez et al.

grid services, and an increase in energy security, the electric energy needed to charge the EVs must come from RES. Therefore, RES plays a crucial role in energy transition, although they are more likely to be used for demand supplements due to their high efficiency and environmental features [2]. EVs have brought about real advantages to the transportation system by charging their only energy user role to real-time active elements, such as Vehicleto-Grid (V2G). V2G is introduced to minimize the heavy impact of EVs in smart islands, a concept that has been developed over time. However, integrating clean and renewable energy sources and integrating electric mobility will pose challenges to their grids. Their charging demand may impose a sudden charge on the grid and disrupt the systems’ normal operation [3]. Therefore, this paper provides four types of battery charging: uncoordinated, coordinated, smart charging for Plug-in Hybrid Electric Vehicle (PHEV), and charging using photovoltaic for EV. In this section, the mathematical model of PHEVs charging types is presented along with the advantages and disadvantages that each one entails according to its functionality. The transition from fossil fuel-based economy and emissions to renewable approach the issue of climate change which is a crucial part addressed at the end of this article. The rest of this paper is organized as below: Sect. 2 briefly introduces the concept of a smart island. The details of electric vehicle charging modes optimization are elaborated in Sect. 3, then Sect. 4 verifies the integration of V2G. Section 5 analyzes the impact of decarbonization of transportation sectors and Sect. 6 presents a discussion to the literature review. Finally, some conclusions are presented in Sect. 7.

2

The Concept of Smart Island

A Smart island (SI) is the definition of an insular territory that can implement integrated solutions to manage infrastructures and natural resources. It is an area that is isolated from the main electrical grid, and it’s known to provide its demand by using RES in a local manner [4]. Therefore, within the smart islands, it has been introduced a platform for the EVs to be charged and discharged without the use of a common grid. 2.1

Challenges of the Smart Island and the Distributed Optimization

Due to its isolation feature, the demand supplement problem is one of the main challenges of the smart island, but this issue can be handled by using RES. In [4], the authors investigated a multi-objective deep adversarial-learning-based approach to address the optimal energy management of microgrids. Here, the aim was to minimize the total cost, pollutant emissions, and uncertainty factors modeled using an effective scenario-based method. The smart island usually faces several challenges due to its different load demands because it can not be supplied only by using transportations and microgrid systems. A way to resolve this issue is through the use of energy hubs (EH),

A Review of Unpredictable RES through EV on Islands

753

as authors in [5] discussed. The EH is a multi-carrier energy system that includes different generation units based on several layers, multiple sources, and functions and can meet the load requirements using effective methodologies. Based on the studies mentioned above, smart islands can be operated and managed on a centralized basis. However, centralized approaches are not yet economically justified, and if the smart island agents are operated independently, this will be a problem. In order to solve this problem, a distributed base method can be used instead of the centralized ones. The distributed approach refers to the methods through which the optimal energy and operation management schemes are obtained by proper energy and data transactions between the agents. In the same paper mentioned previously [4], the distributed optimization is analyzed by using a primal-dual method of multipliers which has shown a better performance when compared to the alternating-direction method of multipliers. Here, distributed approaches have proven better than centralized methods in terms of robustness, low process time, and fewer communication links. The main advantages of distributed methods are higher convergence and decentralization, which reduces the risk of attacking the center and destroying the entire system. 2.2

A Smart Island as an Isolated Smart Microgrid

The microgrid is considered to be the most viable option for a small island power system. Microgrid (MC) is an independent electric system that includes distributed energy sources and electric loads in small electric grids. They can be operated in an islanded mode which provides higher reliability, higher power quality, and power loss reduction. One of the critical primary challenges in MC is energy management for island modes. A radical change in the energy system is required to increase the use of RES and solve the carbon dioxide problem. Authors from [5] provided a newer model for the operation of the smart island, which is represented as an isolated smart microgrid integrated with smart transportation systems residing of EV and the smart EH. Due to the high uncertainty effects on the smart island operation problem, the point estimate method (PEM) approach is used to handle the uncertainties for EVs as one of the study cases implemented.

3

Electric Vehicles Charging Optimization Modes on Islands

Since the conventional demand for an electric vehicle is somewhere between 10 kWh and 100 kWh per charge, the cumulative charging of EVs will impact grid performance and stability. Given that EVs have a high capital cost, charging occurs at home or work during the evening and nighttime. Thus, isolated overloading of the grid may happen in the early stages of electric vehicle adoption. One of the main concerns of electrification of the transport sector is the impact of EVs on isolated electric grids. Distribution transformers can quickly become overloaded since an electric vehicle can increase the charging place

754

J. Chavez et al.

demand, which can lead to several problems [6]. Many charging modes are analyzed to minimize these issues. The overall operation of the microgrid has been divided into four operating modes: uncoordinated, coordinated, smart charging for PHEV, and charging using photovoltaic for EV. 3.1

The Uncoordinated Charging

The uncoordinated charging of a large number of EVs can compromise the grids reliability, security, efficiency and economy. In references [7,8] it is assumed that PHEVs leave their home in the morning and return in the evening, which is around 6:00 PM. A probability distribution function (PDF) with a uniformly distributed feature to model this scenario is defined in (1). f (ts) = (1/b − a) ∗ a ≤ ts ≤ b, a = 18, b = 19

(1)

where ts is the charging start time of the PHEV, a, and b are a constant for specifying the shape of a logarithmic spiral. 3.2

The Coordinated Charging

Coordinated charging or charge management is the simplest way to execute and is most suitable in the early stages of EV adoption. It is also ideal for low electric vehicle penetration rates. Also, in [7,8], PHEV owners tend to charge PHEVs in low-demand hours, during off-peak hours. The charging time is delayed, mainly after 9:00 PM, where the market price is low. The probability density function models the second charging pattern, which can be defined by (2). f (ts) = (1/b − a) ∗ a ≤ ts ≤ b, a = 21, b = 24

(2)

As was mentioned in [6], coordinated charging can be implemented using unidirectional chargers with programmable timers, which can be set to charge the vehicle at a predetermined time of day. This method can help ensure that no additional generating capacity is required, and it also minimizes the impact on the daily demand profile. Optimization of charging times can help reduce daily electricity costs, while coordinated charging can help flatten the load curve. 3.3

Smart Charging

In smart charging (SC), access to data from the charging device is essential to implement SC technology. The EV and the charging device share connection data, and then the charging device shares the data with the charging station operator. However, this data is only shared with the charging station operator with the consent of the EV user. The only disadvantage in SC is its high complexity and cost [9]. Authors in [7] refer that the primary goal of the SC scheme is to schedule the charging pattern of PHEVs in the most beneficial way possible. Vehicle charging

A Review of Unpredictable RES through EV on Islands

755

occurs when there exists a mutual interest between PHEV owners and utilities. In addition, the vehicle starts charging when there is over generation capacity available, and the electricity price is low. In equation (2), a normal probability density function is considered. f (ts) = (1/(ψ ∗

√ 2 2 ∗ π)) ∗ e−1/2∗(ts−μ/σ) , μ = 1, ψ = 3

(3)

where ψ is the covariance, and μ is the mean value of the signal strength data set, and σ is the standard deviation. The vehicle battery starts charging once it is plugged into the grid. Equation (4) models the probability distribution function of the daily driven miles on the vehicle. √ 2 2 f (m) = (1/(m ∗ ψ ∗ 2 ∗ π)) ∗ e(− ln (m)−μ )/2∗ψ ) , m > 0 (4) where m is the number of observations in a sample set with type H1 (where the signal source is an adversary). Authors in [3] refer to the fact that smart charging vehicles are connected to the grid and allow charging and discharging of the batteries to reduce energy import and export while simultaneously reducing production from conventional power plants. The outcomes show an increase in smart charging on EVs and a reduction in hydrogen vehicles. In [10] Islands of Vis, Korkula, Lastovo, and Mljet are investigated. Here the authors propose that interconnections of a group of islands can integrate the production from locally available RES. Besides that, EVs are connected to the grid using smart charging systems and the V2G concept, and those vehicles can be considered potential storage systems for variable energy production. The results, modeled with EnergyPLAN, showed that the interconnections increased the share of energy from RES and reduced the total excess of electricity production. At the same time, the V2G concept enabled the exploitation of synergies between sectors. 3.4

Charging Using Photovoltaic

Out of many RES, the solar photovoltaic (PV) based charging station (CS) is easily accessible and a possible solution. The main goal of designed CS is to utilize the PV array energy maximally to charge EVs. Therefore, its controller is designed to operate CS in an island model. In [11] it is presented a real-time energy management scheme for EV charging using photovoltaic (PV) in the Uligamo island’s context. The charging algorithm provides a decentralized coordinated method based on the heuristic strategy to optimize energy flow within the microgrid. The results show that EV charging using a PV-based microgrid is more economical than the autonomous charging generator. Through this scheme, the burden on the microgrid is reduced significantly. Despite the various advantages of using the PV array for EV charging, the solar irradiance variation affects the charging station operation. Authors in reference [12] have discussed the effects of irradiance variations on EV charging.

756

J. Chavez et al.

The effect of irradiance variation becomes worse if the charging station operates in only island mode. Since the solar PV array power is not available at night, the charging station needs to be assisted with the storage battery [13]. Even under a step-change in solar irradiance level, sudden connection and interruption of EVs do not affect the performance of the control and charging of other EVs. In order to counteract this difficulty, a storage battery can be used to optimize the management of photovoltaic production to smooth electricity consumption during peak periods [14].

4

Integration of V2G

The innovation of V2G has some advantages, features, cost-effectiveness, and technical needs. When an electric vehicle’s battery reaches between 70% and 80% of its original storage capacity, it is considered insufficient for use. In this case, vehicleto-grid services can be used [6]. Battery systems are essential to homeowners on the small islands since this method can reduce electric bills, improve reliability and offer security, especially in power outages due to natural hazards. 4.1

Impact of V2G and V2H Penetration on Islands

Extreme events can damage power systems, like causing power loss for many customers and long outage periods. Therefore, during such circumstances, an electric vehicle can be used to power a house directly, through vehicle-to-home (V2H) or injects energy back into the grid, through vehicle-to-grid (V2G), where EV serves as a mobile energy storage system [15]. In this case, the EV creates a balance in the network by discharging the energy stored in the battery during peak hours and recharging during offpeak hours [9]. In [16], Shin and Baldick optimize the V2H system to provide maximum “backup duration,” meaning the time duration of V2H supports the residential load without experiencing a critical load reduction during island mode. However, if the restoration continues, the amount of energy EVs can contribute to limited stored energy without local power generation. In [4] an emergency power supply strategy featuring scheduled EV charging isolated systems is proposed. Here, by replenishing batteries with secured energy sources, EVs can transport electricity to the island system. To solve this issue, an optimization problem, the genetic algorithm (GA), is formulated to maximize the adequacy of the isolated system offer during the outage period and minimize the total loss of load. 4.2

V2G Charging

Engaging in V2G services can shorten the useful life of the EV by increasing the rate of battery degradation through the constant changes between energy injection into the grid and consumption [9]. However, only services requiring large amounts of power that lead to a significant battery discharge depth can significantly reduce battery life.

A Review of Unpredictable RES through EV on Islands

757

In [17] the energy transition of Hawaii is analyzed by underlining the interconnection of RES, battery storage, and V2G as two of the main components of the future smart grid. Authors in [18] model the potential impact of PHEVs technology on the island of S˜ ao Miguel, A¸cores. They considered this island as an example of an isolated island with high renewable energy potential but largely dependent on fossil fuels incurring high import costs. To this end, the authors employ The Integrated MARKAL-EFOM System (TIMES), where they discuss one-way grid to electric vehicles (G2V) charging strategies for different scenarios of electric vehicles with varying levels of PHEVs penetration and conclude that 32% of them from this island’s vehicle fleet could be realized. The results obtained indicate that the PHEVs integration into the local grid system could become a reality since it bears the potential to yield significant benefits to the energy mix, reducing thus the environmental impact of their heavy fossil-fuel dependency through allowing more intermittent renewable energy onto the grid. Authors in reference [19] analyzed the introduction of EVs in the Caribbean island of Barbados to ease the integration of RES with a predominance of photovoltaic. In this research, two EVs operation modes were analyzed: scheduled charge and V2G, concluding that V2G results are the best solution with the best marginal cost. Article from [20] study presented some insights about the impact of V2G on the island of Korcula. The outcomes show how EVs would increase the total electricity exchanged with the mainland without affecting the peak power exchanged. Likewise, reference [21] obtains similar and better results in terms of primary energy demand reduction. They realized from the scenarios made that the V2G scenario is the one with the lowest annual cost, demonstrating that Smart Energy Systems might offer economically better solutions.

5

Decarbonization of Transportation Sectors on Islands

To promote the decarbonization of transportation sectors, transitioning to EV should develop in tandem with increasing renewable generation. The combination of EVs and RES makes possible the reduction of the dependency on fossil fuels and of the gas emissions. However, it is noted that the decrease in emission values leads to an increase in the total planning cost [22]. Many electric vehicle owners are often motivated to decarbonize their energy consumption and invest in renewable energy systems that offset their household and electric vehicle use. Electric mobility must be carried out in conjunction with the energy sector to ensure that emissions are reduced. Islands need alternative green energy scenarios to balance their high dependence on fossil fuel imports for electricity generation. In [23] authors made scenarios that included hydrogen and battery storage for small islands. Through configurations of energy systems using HOMER software, they demonstrated the decarbonization of electricity and road transport sectors in an environmentally sustainable way. Authors in reference [24] examined the case study of two islands in the Azores, Pico, and Faial islands, to outline possible paths for 100% renewable energy systems [25]. The outcomes show that considering the islands with

758

J. Chavez et al.

independent power systems, Pico island can achieve the primary goal. However, Faial island only reaches 70%. On the other side, the island’s connection allows one to accomplish a 100% renewable energy system in both islands. In [7] authors used SC electric vehicles and V2G, reducing a large part of the hydrogen ones. They came to realize that smart charges in this scenario increase 15% of the total transport demand. On the other side, hydrogen vehicles witness a reduction of 45% in their use in the regional vehicle fleet. Besides that, the grid’s capacity to battery connection must be extended to 3000 MW to make up for the growing number of vehicles that need charging and use these vehicles as storage units for electricity.

6

Discussion

The increasing impact coming from the combination of EV and RES makes it possible to reduce worlds pollution. Therefore, EVs are the perfect solution to solve smart island challenges. To solve these challenges, a microgrid system alongside with EH is the most viable solution on islands. Additionally, the microgrid system is managed through distributed optimization, which has proven to be better than centralized methods. To minimize the problems that EV penetration causes on isolated electrical networks, four charging modes are referenced. In Table 1, it is possible to see a summary of the main advantages, disadvantages, and functionality of the operating modes discussed before. Overall, coordinated and uncoordinated charging modes have more disadvantages than smart charging since they can compromise the main grid. However, despite smart charging having a low electricity price and promoting a reduction of hydrogen vehicles, charging by using a PV is the most accessible and less polluting solution. The drawback is that charging through PV depends on solar irradiation variety, meaning there is no production at night. A way to counteract this disadvantage is by assisting EVs with a storage battery. Table 1. Summary of operation modes. Operating modes Functionality

Advantages/Disadvantages

Uncoordinated

PHEV leave their home in the morning and return in the evening, which is around 6:00 PM

Can compromise grid’s reliability, security, efficiency, and economy

Coordinated

The charging time is delayed, mainly after 9:00PM

Can help ensure that no additional generating capacity is required and it also minimizes the impact on the daily demand profile

Smart charging

The vehicle battery starts charging The outcomes show an increase in once it is plugged into the grid the smart charging on EV and a reduction of hydrogen vehicles

Photovoltaic

Through the use of PV array energy maximally to charge EV

Easily accessible and a possible solution although it depends on solar irradiation variance

A Review of Unpredictable RES through EV on Islands

7

759

Conclusion

This paper provides solutions that improve the ability of the grid to cope with unpredictable RES in the insular contexts by using EV. Here, several solutions, such as uncoordinated, coordinated, smart charging for PHEV and charging using photovoltaic for EV have been made, concluding that battery energy systems through charging using photovoltaic for EV are the most used technologies in the islands. This review discussed the concept of a smart island platform for the EVs to be charged and discharged without using a typical grid, along with the several challenges they faced. Besides that, it examined the topic of EVs dealing with the different operating modes such as V2G, which is not very feasible since not enough developments have been found in order to create a battery technology necessary for the success of this mode of operation. The results depend on the island’s size and the charging mode. Funding. This work has received funding from FEDER Funds through COMPETE program and from National Funds through FCT under the project BENEFICEPTDC/EEI-EEE/29070/2017 and UIDB/00760/2020 under CEECIND/02814/2017 grant.

References 1. Groppi, D., Pfeifer, A., Garcia, D.A., Krajaˇci´c, G., Dui´c, N.: A review on energy storage and demand side management solutions in smart energy islands. Renew. Sustain. Energy Rev. 135, 110183 (2021) 2. Verma, A., Singh, B.: Multimode operation of Solar PV array, grid, battery and diesel generator set based EV charging station. IEEE Trans. Ind. Appl. 56(5), 5330–5339 (2020) 3. Calise, F., Duic, N., Pfeifer, A., Vicidomini, M., Orlando, A.M.: Moving the system boundaries in decarbonization of large islands. Energy Conv. Manag.234, 113956 (2021) 4. Mohamed, M.A., Jin, T., Su, W.: Multi-agent energy management of smart islands using primal-dual method of multipliers. Energy 208, 118306 (2020) 5. Mohamed, M. A., Almalaq, M. A., Mahrous Awwad, E.,El-Meligy, M. A.,Sharaf, M. & Z. M. Ali.: An Effective energy management approach within a smart island considering water-energy Hub. IEEE Trans. Ind. Appl., 1-1 (2020) 6. Gay, D., Rogers, T., Shirley, R.: Small island developing states and their suitability for electric vehicles and vehicle-to-grid services. Utilit. Policy 55, 69–78 (2018) 7. Dabbaghjamanesh, M., Kavousi-Fard, A., Zhang, J.: Stochastic modeling and integration of plug-in hybrid electric vehicles in reconfigurable microgrids with deep learning-based forecasting. IEEE Trans. Intell. Transp. Syst. 22(7), 1–10 (2020) 8. Lei, M., Mohammadi, M.: Hybrid machine learning based energy policy and management in the renewable-based microgrids considering hybrid electric vehicle charging demand. Int. J. Elect. Power Energy Syst. 128, 106702 (2021) 9. Almeida, J., Soares, J.:Integration of electric vehicles in local energy markets. Elsevier , Amsterdam (2021) 10. Pfeifer, A., Dobravec, V., Pavlinek, L., Krajaˇci´c, G., Dui´c, N.: Integration of renewable energy and demand response technologies in interconnected energy systems. Energy 161, 447–455 (2018)

760

J. Chavez et al.

11. Bhatti, A.R., Salam, Z., Ashique, R.H.: Electric Vehicle charging using photovoltaic based microgrid for remote islands. Energy Procedia 103, 213–218 (2018) 12. Islam, M.S., Mithulananthan, N., Lee, K.Y.: Suitability of PV and battery storage in EV charging at business premises. IEEE Trans. Power Syst. 33(4), 382–4396 (2018) 13. Mahmood, H., Jiang, J.: Autonomous coordination of multiple PV/battery hybrid units in islanded microgrids. IEEE Trans. Smart Grid 9(6), 6359–6368 (2018) 14. Ramos, S., Foroozandeh, Z., Soares, J., Tavares, I., Faria, P., Vale, Z.: Shared PV production in energy communities and buildings context. Renew. Energy Power Qual. J. 19(19), 459–464 (2021) 15. Xu, N.Z., Chan, K.W., Chung, C.Y., Niu, M.: Enhancing adequacy of isolated systems with electric vehicle-based emergency strategy. IEEE Trans. Intell. Trans. Syst. 21(8), 3469–3475 (2020) 16. Shin, H., Baldick, R.: Plug-in electric vehicle to home (V2H) operation under a grid outage. IEEE Trans. Smart Grid 8(4), 2032–2041 (2017) 17. Lee, T., Glick, M.B., Lee, J.H.: Island energy transition: assessing Hawaii’s multilevel, policy-driven approach. Renew. Sustain. Energy Rev. 118, 109500 (2020) 18. Ioakimidis, C.S., Genikomsakis, K.N.: Introduction of plug in hybrid electric vehicles in an isolated island system. Adv. Build Energy Res. 12(1), 66–83 (2018) 19. Taibi, E., Fern´ andez del Valle, C., Howells, M.: Strategies for solar and wind integration by leveraging flexibility from electric vehicles: the Barbados case study. Energy 164, 65–78 (2018) 20. Doroti´c, H., Doraˇci´c, B., Dobravec, V., Pukˇsec, T., Krajaˇci´c, G., Dui´c, N.: Integration of transport and energy sectors in island communities with 100% intermittent renewable energy sources. Renew. Sustain. Energy Rev. 99, 109–124 (2019) 21. Meschede, H.: Increased utilisation of renewable energies through demand response in the water supply sector - a case study. Energy 175, 810–817 (2019) 22. DeLima, T.D., Franco, J.F., Lezama, F., Soares, J., Vale, Z.: Joint optimal allocation of electric vehicle charging stations and renewable energy sources including CO2 emissions. Energy Inform. 4(Suppl2) (2021) 23. Groppi, D., Astiaso Garcia, D., Lo Basso, G., Cumo, F., De Santoli, L.: Analysing economic and environmental sustainability related to the use of battery and hydrogen energy storages for increasing the energy independence of small islands. Energy Conv. Manag. 177, 64–76 (2018) 24. Alves, M., Segurado, R., Costa, M.: On the road to 100% renewable energy systems in isolated islands. Energy 198, 117321 (2020) 25. Alves, M., Segurado, R., Costa, M.: Increasing the penetration of renewable energy sources in isolated islands through the interconnection of their power systems. The case of Pico and Faial islands, Azores. Energy 182, 502–510 (2019)

Fish Control Process and Traceability for Value Creation Using Blockchain Technology Joao C. Ferreira1,2,5

, Ana Lucia Martins1,3 , Ulpan Tokkozhina1,3,5(B) and Berit Irene Helgheim4

,

1 Instituto Universitário de Lisboa (ISCTE-IUL), 1649-026 Lisbon, Portugal

{jcafa,almartins,ulpan_tokkozhina}@iscte-iul.pt

2 Information Sciences and Technologies and Architecture Research Centre (ISTAR-IUL),

Lisbon, Portugal 3 Business Research Unit (BRU-IUL), Lisbon, Portugal 4 Faculty of Logistics, Molde University College, 6410 Molde, Norway

[email protected] 5 Inov Inesc Inovação—Instituto de Novas Tecnologias, 1000-029 Lisbon, Portugal

Abstract. Currently, the globe and, more specifically, corporate processes are evolving towards digitization and the elimination of time-consuming procedures. As origin knowledge and safety assurance of products became more important in establishing end-consumer confidence, traceability solutions are acute. A Blockchain (BC) is one of the most exciting disruptive technologies emerging today. The purpose of this study is to analyze the possibility of using BC to allow end-to-end supply chain (SC) management traceability and to provide a proposition of a BC application for fish tracking and control in SC contexts. The proposition demonstrates that BC provides the use of a single information system by several players across SC, resulting in more secure and efficient data sharing. A future understanding of the significance of BC technology usage as a traceability supplier and a detailed testing of a proposed architecture is identified as a diretion for future study. We, therefore, study current applications in supply chain management, with a big emphasis on the fish control process. Keywords: Blockchain · Traceability · Supply chain · IoT · Fish

1 Introduction Seafood is one of the planet’s biggest industries, representing 12% of livelihoods worldwide, where the wealth of 1 out of every 10 people on the planet derives from seafood and aquaculture [1]. More fish than ever is consumed by humans, global consumption per capita has almost doubled in the past 50 years - complete market size figures are upwards of USD $500 billion [2]. Pressure on marine ecosystems is high. Illegal, unreported, or unregulated (IUU) fishing activities pose a threat to the sustainability of ecosystems and can seriously affect the global economy. It is estimated that, on a global scale, IUU fishing represents about 30% of the total catch of global fisheries, severely penalizing coastal economies ecosystems and depleting © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 761–773, 2022. https://doi.org/10.1007/978-3-030-96299-9_72

762

J. C. Ferreira et al.

fish stocks natural capital at unsustainable rates. Annual losses worldwide due to IUU fishing are estimated at between 10 and 23.5 billion US$, representing between 11 and 26 million tons of fish products [3]. The loss of economic, social and environmental value is therefore severe. Seafood traceability is rapidly becoming a focal point for addressing the entry into the supply chains (SC) of goods that are illegally and unethically made or caught. Increasingly, experts see complete traceability and accountability in the supply chain as the only way to avoid the continued market-entry of seafood products produced or fished illegally or unethically. Blockchain (BC) technology can be an important component of the solution, providing global transparency and traceability and, consequently, creating value for both SC entities and final consumers. The problem of IUU fishing in the supply chains could theoretically be solved by fully open and traceable supply chains of seafood, enabled by blockchain technology, creating value for all SC stakeholders. Blockchain technology, also defined as an encrypted digital ledger [4] is based on a decentralized peer-to-peer system [5], that is able to create a continuous, visible and sharable record of products transactions and movements around SC in a distributed manner [6]. BC is a set of chain block, that altogether represent a permanent and tamper-proof sequence of records and transactions that can be verified in the future. Keys and encryption secure the process, and each stakeholder is identified by their key [7]. This network is build based on the consensus achieved by different voting mechanisms and the chain is extended with a new block when the majority of participants agree with it [8]. Blockchain stores any data transaction or exchange that takes place on the network, potentialy reducing the need for third parties and/or intermediaries by offering a way to distribute access to all parties on the network. While other technology options exist to help manage SCs, blockchain provides another arrow in the quiver, one that can bring together different parties that have not directly established trusted relationships with one another. BC stores every transaction or exchange of data that occurs in the network, potentially reducing the need for third parties and intermediaries by providing a means where all parties in the network may share access to the same data, including what is added to the data, and by whom, chronologically. Data cannot be removed, thus by enabling each party to see the same data, in near real-time, and assure that ‘you see what I see’ from a data perspective, blockchain can help eliminate complex and costly data reconciliation required by most systems in the world today. Application on fishing SC will be a challenge but will allow bringing together different stakeholders without any previous trustworthy relationships among them in transparent data exchange. Benefits and value drivers of a blockchain traceability solution could include more streamlined data sharing and improved confidence in catch data (e.g., to know if fish was caught in legal waters), company brand enhancement for IUU compliant prod-ucts, and a clearer business case for investing in such enabling technologies as Radio Frequency Identification (RFID) tagging that could improve automation and streamline operations (and that could be justified for higher value catch such as a fish business). Nowadays all countries are trying to implement different tools and technologies to have access to this critical information, from the caught to the consumer, but currently due to the diversity of stakeholders and associated information systems there is no control

Fish Control Process and Traceability

763

over the process. In this scenario, blockchain implemented in a large scale will allow mitigating this process. Blockchain allows value creation by providing a new crossindustry collaboration to share supply chain data, and from this data analytics processes allows future improvements. The food transportation process is essential to be the focus on the safety, quality, and the certification of producers in a global market driven by profit. [9] highlights the rising number of problems related to food safety and contamination risks. When it comes to fish, product traceability from origin producer to the final consumer is an essential problem to solve. BC disruptive technology can give solutions to this problem by managing identity of process stakeholders and associate immutable transaction block of product transactions [10]. Attributing food traceability as part of logistics management highlights the fact that quality assurance, food safety and overall efficiency of SC depend highly on logistics operations [11]. Tsolakis and colleagues [12] argue BC is well suited to address United Nation Sustainable Development Goals within the fish industry. Thus, the purpose of this study is to analyze the possibility of using BC to allow end-to-end supply chain (SC) management traceability and to introduce a proposition of a BC application for fish tracing in SC contexts. The paper is organized as following: Sect. 2 reviews the literature and provides the main highlights found in literature on blockchain technology features and contraints, focusing its attention on a traceability feature. Section 3 introduces the developed BC-based proposition for supply chains, identifying possible entities that may participate in case and revealing the vision of a final consumer. Section 4 provides a discussion of a proposed arcitecture based on mindful technology adoption frameworks; Sect. 5 is therefore concluding the study, also suggesting possible ideas for future studies.

2 Literature Review The interest in innovative technology solutions for business processes applications is growing rapidly. Considering emergency pandemic events of 2020, the efficiency of operations and digitalization is emergent as never before. [13] highlight the main criteria of supply chains in the light of these events: trustworthiness, transparency, and real-time accurate information sharing—all this would assure the safety of global populations. The global supply chain is an industry that is running two-thirds of the global economy [14], bringing to consumers everything that surrounds us: what we eat, wear and use for everyday life activities. One of the most emergent, promising technologies that has the potential to disrupt and transform supply chain activities is Blockchain [15]. Blockchain (BC) is an emerging technology, with potential applications to everyday life, from voting systems and digital identity recognition to healthcare and legal contracts [16, 17]. The distributed nature of BC, immutability of records, persistence and the ability to follow decentralized logic through smart contracts use make BCbased products and services significantly different from those previously developed and based on the Internet (e.g. Industry 4.0 sectors and supply chains) [18]. BC is assumed to become a “next holy grail for the enterprise”, as it holds enormous potential for supply chain improvement in the ways of manufacturing, production, orders placing, transportation, delivery and consumption [19]. Though, with a few exceptions, SCs are

764

J. C. Ferreira et al.

not considered to be a priority on the agenda of most countries with BC initiatives, even though interest is growing [20], empirical research is relatively limited due to the lack of knowledge among SC practitioners about the potential of this technology [21]. In the adoption of BC in supply chain management (SCM), [22] argues that everything should start with answering to two key questions: “What to adopt” and “Where to start”. Under the scope of the first question, [23] argues in favour of adopting a use case, nonetheless the second question still requires more detailed approach. Verhoeven and colleagues [24] suggest a framework for the mindful adoption of BC, which was extended by [23]. Further [23] made a guide proposition for the “mindfulness of technology adoption” under the context of BC technology. The food industry has experienced many quality drawbacks in recent years. Distrust of society in the provenance of seafood products and some conservation operations is growing [25] and BC technology might be helpful to overcome such distrust. As so, following on [24] and [23] suggestion of starting the adoption of BC in SC with pilots and use cases, this study is focused on the specific SC for a sensitive by its perishability and origins product - fish. Consequently, the overall goal of the study is to see and assess the value creation of the adoption of BC technology in SC trust in the fish industry. 2.1 Blockchain Technology and Its Featurs for Supply Chain Mangement Operations within BC are known for being fully decentralized, thus there is no need to rely on an intermediary or trusted third-party. Queiroz and colleagues [26] state transactions can be verified with smart contracts. Excepional features of smart contracts include, but are not limited to tamper-proof systems, automated processes and [27] together with the distributed nature of BC, improve upon overall synchronization of business operations and optimization of ownership value [28]. Smart contracts are expected to play a key role in managing partnership efficiency—due to the information immutability, it results in transparency and improvement of SC collaboration [29]. Due to self-executing codes, that are preventing potential corruptions or tampers withing the execution of a given contract, every party is an equal custodian of the full contract terms; this saves both time and costs in terms of contract revision, registration and verification [27]. However, building and writing a high-quality smart contract is even a bigger challenge rather than creating a traditional one, since expertise in this field is not so common yet [14] and as a result, poor, error containing coding of smart contracts leads to problems, that can be explained by immutability feature of BC-based solution [30]. Figure 1 represents the difference that smart contracts are able to create in the process of contract agreements and execution, when compared to traditional centralized systems. By its nature, BC is highlighting transparency throughout the SC, in this way providing reliance in origins and confidence of products’ provenance [21]. BC-based encrypted ledgers provide a unified version of truth through consensus protocols [31], thus enhancing the performance of SC, eliminating need to establish trust relationship among participants since everyone is a holder of full information flow existing around SC [27]. Transparency of information regarding products and processes empowers suppliers to get engaged with further activities and decisions, such as strategies development and innovation support [32].

Fish Control Process and Traceability

765

Fig. 1. Traditional (centralized) system of contracts

Even at the era of the digital economy, SCs are still cyber-vulnerable: they can be subjects to attacks due to their insecurity and are constrained with issues of trust both among suppliers as well as between supplier and consumer [33]. It is assumed, that BC holds a great potential to decentralize traditional SC and generate new networks of value combining it together with artificial intelligence, additive manufacturing, and Internet of Things (IoT) [14]. BC IoT framework is expected to be the main driver that will boost SCs to the next level of analytics, enabling data democracy, and therefore improving firms’ productivity and performance [19]. So, the next step of the digitalization should be the digitalized transformation of industrial companies, enabling the exchange of data and services in a trusted manner, and introduction of smart contracts as a unified tool for the value transfer. BC implementation is potentially applicable to any sector from construction engineering [27] and smart city collaborative gamification [34] to diamond authentication [35] and the music recording industry [36]. BC applications are commonly implied to be used together with IoT solutions, e.g. using BC as a decentralized platform for IoT-based low-cost smart meters for energy consumption [37] or for handling charging processes of electric vehicles through mobile application [38]. BC use is actually “only limited by our imagination” [14]; one of the most well-known logistics BC effectuations is the collaboration between IBM and Maersk—the use-case for maritime container shipping [5]. Walmart is testing BC for food SC [14], some studies focus on conceptual models’ applications, the case for agro-food is explored at [39], electronic components at [40]. Safety is also an challenge that was explored under the food business perspective in [41]. General applications discussion is performed at [42]. Also, wine traceability is studied at [43] and vegetables’ traceability at [44]. 2.2 Contraints of Blockchain Technology for Supply Chain Mangement Use BC technology, undoubtely looks attractive for both scholars and practitioners, however, there is still a considerable amount of challenges for its integration into SC context. Numerous infrastuctural, institutional, technical and regulatory challenges need to be embraced before BC-based solutions can reach their maturity stage [45]. Among others, challenges like organizational readiness, technical expertise, scalability [46] high cost of the technology and further dispute-resolving regulation issues [47] may arise when

766

J. C. Ferreira et al.

implementing BC in the SC. Security issues of permissionless access BC [48] and management procedures for BC used by multi-actor SCs [49] need to be addressed in future studies. A lot of BC pilots have difficulties in emerging from the pioneering phase [50] and in the majority of cases, organizational changes are needed to be undertaken before this technology can be successfully adopted [11]. Generally, all of these challenges of a BC implementation imply a high risk of emergent technology adoption from scratch that also implies big costs [45]. Moreover, the literature on BC technology for SC management needs “theoretical substance and a theoretical foundation” [51] that could refine the understanding of such an emergent phenomenon [25]. BC is claimed to be useful for traceability of goods throughout SCs, improving thus the overall transparency, however, organization and preparation of SCs themselves is essential before BC can be implemented [11]. [52] states “although some global marine conservation organisations and seafood producers have found practical solutions in disruptive technologies like blockchain, riding this wave will only prove worthwhile if coastal communities and artisanal fishers are on board and stand a chance of landing a fair share of the benefits”. Kamilaris and colleagues [53] highlight the fact that many barriers and challenges still exist, which hinder blockchain wider popularity among farmers and SC systems. These challenges involve education, technical aspects and literacy, policies and regulatory frameworks. 2.3 Traceability for Blochchain-Based Food Supply Chains Food quality is a big concern to society, thus of us, and assuring quality throughout global and complex supply chains is very challenging for food and beverage industries. Also, issues as legal regulations, food standards and corporate social responsibility, including environmental sustainability concerns, should be considered at the highest priority [54]. Like this, product traceability from origin producer to the end consumer is an essential problem to adress. Such a disruptive technology as BC, can provide solutions to this problem by managing the identity of process stakeholders and associate immutable transaction block of product transactions, allowing food retailers to keep a track and react rapidly for recalls, in this way assuring safety concerns and reducing the chance of illnesses caused by food [6]. BC and smart contracts can handle transactions in a SC process without a central control entity (third party). According to [55], BC-based solutions for food SCs could be crucial in pandemic times, as complex and lengthy overseas SCs made it very challenging for agricultural exporters to get the usual guarantees and maintain cashflow. On the example of Australia’s surplus of seafood and agriculture that Chinese market used to order, [55] explain that BC-based solution could give an ability for every participant of a SC to confirm the types of products shipped, track its real-time location, and whether it has been stored under required conditions (e.g. temperature, humidity etc.).

3 Blockchain-Based Fish Traceability Proposition This study focuses attention on traceability feature of BC technology, as well as the recognition of this attribute in the B2B (business-to-business) relations in the fish SC

Fish Control Process and Traceability

767

together with the trust of the final customer in the product available. This proposition aims to develop a holistic approach that permits the traceability of sustainable fish from moment of capturing it, all the way to the final consumer using blockchain/distributed ledger technologies, and with the use of a scan on the mobile device, it will allow users to obtain all the information: where and when the fish was caught, by which vessel and the fishing method, like this creating social and economic value. Figure 2 represents the proposed blockchain application for a fish SC. This proposition is only focused on general idea of main players, thus there are no limitations for adding players to a specific fish pilot. We suggest the Block 0 to be created from the moment fish was captured by a fishing vessel, so that the next party (in this case a fish wholesaler entity, who receives fish from vessels) can consequently create the Block 1. Like this fish is changing ownership and through its locomotion is being represented by each consequitve block created. One of the central ideas is that final consumer, once choosing the fish from a supermarket fridge, would be able to scan a code and see the full information on locomotion and transactions connected to the chosen fish – which will allow to have a full traceability until the origins of a product.

Fig. 2. Blockchain proposition for a fish supply chain

If added to the proposed network, competent state authorities can also have a possibility to follow process of fishing procedures starting from fishing vessels level, and thus combat illegal fishing, generating both economic and environmental value. A combination of RFID, QR codes or even fish DNA tacking technique will be used to capture information throughout the supply chain. An RFID tag will be affixed when the fish enters the vessel, which will follow it and will automatically register on the various devices positioned on the vessel, the dock, and the processing facility. Once the product enters the processing facility and is split into several new products, it will receive a QR code, or potentially in the future a Near Field Communication (NFC) device, that will accompany the product to its final destination at the end of the supply chain. Additionally, this network will have a potential to solve the instability of most seafood SCs, enabling independent industry stakeholders to leverage the blockchain power using a shared protocol so that data can be trusted, open, and protected. Inherent accountability and traceability of BC could effectively reduce the pressure posed on natural resources

768

J. C. Ferreira et al.

by removing illicit or unethical seafood from the SC, when preventing it from reaching the market. This would create sustainable value for all stakeholders in the SC. By encouraging seafood buyers and consumers to make well-informed commercial choices based on verifiable facts, it would achieve this unprecedented result. It would also deliver an effective control process and increase trust among the stakeholders, while protecting and enhancing natural capital.

4 Discussion BC is expected to help dealing with complexity of food industry sector in terms of full traceability of SC networks [32]. Most of traceability standards are based on the ability to follow the main characteristics of a product information from origin to the final process destination throughout the SC [11]. The typical food SC consists of many members, e.g. suppliers, producers, manufacturers, distributors, retailers, consumers and certifiers; once connected together on a unified BC platform, every single one of them will be able to update, add and verify the real-time information about products [56]. Since every transaction is visible in the BC network, it is easy to trace backward of the supply of each product or service with authenticity from a compliance or quality assurance perspective [27]. Traceability feature of BC brings the knowledge of the authenticity and origins of a product, together with footprints of products’ movements throughout the SC, which may result in both commercial benefits in terms of brand corporate social responsibility and serious safety measures [6]. Business requirements for BC-enabled traceability systems from the SC’s focal companies view were addressed in [14] claiming that specific business requirements and technological evaluation of the business case development should be accurately analyzed. Some of the central challenges in SCM is product traceability and supplier dependability, satisfaction and trust; it will impact on the performance of the entire SC [57]. Issues, such as traceability of products supplier dependability and end-to-end time and quality of service are crucial for the success of SCs. It is visible in the literature, that traceability is a relevant issue to the several parties in a SC; however it may not be recognized for having similar relevance for every element of the SC. The extend of this relevance is yet to be discussed, as well as the impacts or utility of traceability. In the context of BC technology adoption to SCM and logistics, [24] completed four mindful dimensions of technology adoption by [58, 59] and introduce the fifth one; those dimensions are as follows: (1) Engagement with the technology—Are the technological features named clearly? (2) Technological novelty seeking—Is there reasoning for the necessity of blockchain technology or can the business problem be solved with existing technology? (3) Awareness of local context—How specifically will the use case fit into the supply chain context? (4) Cognizance of alternative technologies—Are alternatives considered? (5) Anticipation of technology alteration—Are use cases adaptable?

Fish Control Process and Traceability

769

The listed dimensions were considered under the lens of key high-level SC performance objectives presented earlier by [60], which include cost reduction, speed, dependability, risk reduction, flexibility and sustainability. Since [24] were concerned with a risk of “a solution looking for a problem”, [23] expanded those five principles and added one more dimension, which is a “contribution to high-level supply chain objectives”. This dimension contributes to elimination of the unsuitability risk of BC technology in a potential use-case [23]. By mean of this substantial dimension, in future, it will save resources and time for SC that will search for latter-day technology to adopt. [61] analyses the adoption of blockchain technology in Norwegian fish supply chains and concludes that it potentially might benefit from blockchain technology. This is mainly a conclusion drawn by the fact that Norwegian aquaculture producers that were viewed to have lean SCs and have functional products in the material flows, and that the SC management objectives, cost, quality and sustainability, with the cost and quality being integral to lean SC practices, were found to get the highest advantages. In addition, blockchain technology was found to benefit lean SCs in general, especially food SCs. As can be seen from our proposition, a product with fragile quality and high value as fish, shows to be a good example for a pilot making this use case adaptable in the context of emergent technology, which therefore contributes to a high-level SC objectives, such as quality.

5 Conclusions In view of the present pandemic situation, stronger controls on food origin and safety are required at each step of SCs. By commencing a BC-based traceability process, SC actors may prevent fraudulent behaviors and possible corruption while also establishing confidence with the end-consumer via the provision of health and safety that can be verified by the customer, when choosing the product to buy. A disruptive solution of this kind has the potential to drastically decrease illicit fishing, while maintaining and even increasing corporate value and social responsibility. Current SC risks related with a lack of supplier responsibility and process transparency might be mitigated with the installation of a BC-based traceability system. BC offers a complete backward trace audit of data by monitoring transactions, processes, ownership change and critical data connected with a product and establishes a permanent encrypted platform for transaction and record-keeping across SC. Concerning the BC effect assessment in the framework of the SC, the expansion of traceability relevance to each SC element has to be further enhanced. Given the difficulty of establishing trust at the end customer stage, a BC app may be created in the future to give ultimate customers with product traceability. However, asserts that cellphones and BC alone are insufficient for accurate tracking and monitoring of caught and processed fish; hence, other kinds of sensors and trackers, such as IoT devices, remote sensors, and portable DNA sequencers, may assist alleviate this worry. The use of a case study technique has certain drawbacks, including the inability to generalize results. Nonetheless, the instance serves as a model for the proposition’s evolution. As a result, the descriptive, detailed framework might be tried in subsequent pilots or case studies with a comparable SC structure.

770

J. C. Ferreira et al.

Acknowledgement. This work was supported by EEAGrants - Programa Crescimento Azul (Aviso #5) - Candidatura EEA.BG.CALL5.005.2020 – Fish2Fork.

References 1. Our Ocean. http://ourocean2016.org/sustainable-fisheries 2. Holland, J.: UN: The world is producing and consuming more seafood, but overfishing remains rife (2020). https://www.seafoodsource.com/news/supply-trade/un-the-world-is-pro ducing-and-consuming-more-seafood-but-overfishing-remains-rife 3. Food and Agriculture Organization of the United Nations. https://www.fao.org/fao-stories/ article/en/c/1136937/ 4. Qian, X. (Alice), Papadonikolaki, E.: Shifting trust in construction supply chainsthrough blockchain technology. Eng. Constr. Architect. Manag, May (2020). https://doi.org/10.1108/ ECAM-12-2019-0676 5. O’Leary, D.E.: Configuring blockchain architectures for transaction information in blockchain consortiums: The case of accounting and supply chain systems. Intell. Syst. Account. Fin. Manag. 24(4), 138–147 (2017) 6. Wang, Y., Han, J.H., Beynon-Davies, P.: Understanding blockchain technology for future supply chains: a systematic literature review and research agenda. Supply Chain Manag. 24(1), 62–84 (2019). https://doi.org/10.1108/SCM-03-2018-0148 7. Li, D., Du, R., Fu, Y., Au, M.H.: Meta-key: a secure data-sharing protocol under blockchainbased decentralized storage architecture. IEEE Netw. Lett. 1(1), 30–33 (2019) 8. Liu, J., et al. A data storage method based on blockchain for decentralization DNS. In: 2018 IEEE Third International Conference on Data Science in Cyberspace (DSC), pp. 189–196. IEEE, June 2018 9. Jing, Z.: Application of information technology in food storage and trans portation safety management and establishment of information network integration platform for food storage and transportation safety management. In: 2018 International Confer ence on Information Management and Processing (ICIMP), pp. 125–129. IEEE January, 2018 10. Shi, P., Wang, H., Yang, S., Chen, C., Yang, W.: Blockchain-based trusted data sharing among trusted stakeholders in IoT. Softw. Pract. Exp. (2019) 11. Behnke, K., Janssen, M.F.W.H.A.: Boundary conditions for traceability in food supply chains using blockchain technology. Int. J. Inf. Manag. 52 101969 (2019). https://doi.org/10.1016/j. ijinfomgt.2019.05.025 12. Tsolakis, N., Niedenzu, D., Simonetto, M., Dora, M., Kumar, M.: Supply net-work design to address United Nations Sustainable Development Goals: a case study of blockchain implementation in Thai fish industry. J. Bus. Res. (2020). https://doi.org/10.1016/j.jbusres.2020. 08.003. ISSN 0148-2963 13. Tapscott, D., Tapscott, A.: Blockchain Solutions in Pandemics A Call for Innovation and Transformation in Public Health (2020) 14. Tapscott, D., Tapscott, A.: Blockchain Revolution. How the Technology Behind Bitcoin and Other Cryptocurrencies Is Changing the World. 2nd edn. Penguin Business, Penguin Business, London (2019) 15. Nayak, G., Dhaigude, A.S.: A conceptual model of sustainable supply chain management in small and medium enterprises using blockchain technology. Cogent Econ. Fin. 7(1), 1667184 (2019) 16. Allen, M.: How blockchain could soon affect everyday lives (2017). http://www.swissinfo. ch/eng/joining-the-blocks_how-blockchain-could-soon-affect-everyday-lives/43003266

Fish Control Process and Traceability

771

17. Sharma, T.: How Blockchain Can Benefit You In Your Daily Life (2018). https://www.blockc hain-council.org/blockchain/how-blockchain-can-benefit-you-in-your-daily-life/ 18. Rejeb, A., Keogh, J.G., Treiblmaier, H.: Leveraging the internet of things and blockchain technology in supply chain management. Fut Internet 11(7), 161 (2019) 19. Sachdev, D.: enabling data democracy in supply chain using blockchain and Iot. J. Manag. 6(1), 66–83 (2019). https://doi.org/10.34218/jom.6.1.2019.008 20. van Hoek, R.: Exploring blockchain implementation in the supply chain. Int. J. Oper. Prod. Manag (2019) 21. Montecchi, M., Plangger, K., Etter, M.: It’s real, trust me! Establishing supply chain provenance using blockchain. Bus. Horiz. 62(3), 283–293 (2019). https://doi.org/10.1016/j.bushor. 2019.01.008 22. Dobrovnik, M., Herold, D.M., Furst, E., Kummer, S.: Blockchain for and in logistics: what to adopt and where to start. Logistics 2(3), 18 (2018) 23. Van Hoek, R.: Developing a framework for considering blockchain pilots in the supply chain – lessons from early industry adopters. Supply Chain Manag. Int. J. 25(1), 115–121 (2020). https://doi.org/10.1108/SCM-05-2019-0206 24. Verhoeven, P., Sinn, F., Herden, T.T.: Examples for blockchain implementa- tions in logistics and supply chain management: exploring the mindful use of a new tech- nology. Logistics 2(3) (2018) 25. Howson, P.: Building trust and equity in marine conservation and fisheries supply chain management with blockchain. Marine Policy 115, 103873 (2020) 26. Queiroz, M.M., Telles, R., Bonilla, S.H.: Blockchain and supply chain man- agement integration: a systematic review of the literature. Supply Chain Manag. (2019). https://doi.org/ 10.1108/SCM-03-2018-0143 27. Wang, J., Wu, P., Wang, X., Shou, W.: The outlook of blockchain technology for construction engineering management. Front. Eng. Manag. 4, 67–75 (2017) 28. Chang, S.E., Chen, Y.C., Lu, M.F.: Supply chain re-engineering using block- chain technology: a case of smart contract based tracking process. Technol. Forecast. Soc. Change 144, 1–11 (2019) 29. Kim, J.S., Shin, N.: The impact of blockchain technology application on supply chain partnership and performance. Sustainability (Switzerland), 11(21) (2019): https://doi.org/10.3390/ su11216181 30. Stevenson, M., Aitken, J.: Blockchain technology: implications for operations and supply chain management. Supply Chain Manag. Int. J. 24(4), 469–483 (2019). https://doi.org/10. 1108/SCM-09-2018-0309 31. Schuetz, S., Venkatesh, V.: Blockchain, adoption, and financial inclusion in India: research opportunities. Int. J. Inf. Manag. 52, 101936 (2020) 32. Huang, Y., Han, W., Macbeth, D.K.: The complexity of collaboration in supply chain networks. Supply Chain Manag. 25(3), 393–410 (2020). https://doi.org/10.1108/SCM-11-2018-0382 33. Kshetri, N., Voas, J.: Supply chain trust. IT Prof. 21(2), 6 (2019). https://doi.org/10.1109/ MITP.2019.2895423 34. Ferreira, J., Martins, A., Gonçalves, F., Maia, R.: A blockchain and gamification approach for smart parking. In: Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering, pp. 3–14. Springer, Guimarães (2019) 35. Choi, T.M.: Blockchain-technology-supported platforms for diamond authentica- tion and certification in luxury supply chains. Transp. Res Part E Logist. Transp. Rev. 128, 17–29 (2019) 36. Chalmers, D., Matthews, R., Hyslop, A.: Blockchain as an external enabler of new venture ideas: Digital entrepreneurs and the disintermediation of the global music in- dustry. J. Bus. Res. 125, 577–591(2019)

772

J. C. Ferreira et al.

37. Ferreira, J.C., Martins, A.L.: Building a community of users for open market en ergy. Energies 11(9) (2018) 38. Martins, J.P., Ferreira, J., Monteiro, V., Afonso, J.A., Afonso, J.L.: IoT and blockchain paradigms for EV charging system. Energies 12, 2987 (2019) 39. Rue, H., Martino, S., Chopin, N.: Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations. J. R. Stat. Soc. Ser. B (Statistical Methodology) 71(2), 319–392 (2009) 40. Caro, M.P., Ali, M.S., Vecchio, M., Giaffreda, R.: Blockchain-based traceability in agri-food supply chain management: a practical implementation. In: 2018 IoT Vertical and Topical Summit on Agriculture - Tuscany (IOT Tuscany), Tuscany, pp. 1–4 (2018), https://doi.org/10. 1109/IOT-TUSCANY.2018.8373021.https://ieeexplore.ieee.org/abstract/document/8373021 41. Figorilli, S., et al.: Blockchain Implementation prototype for the electronic open source traceability of wood along the whole supply chain. Sensors 18, 3133 (2018). https://www.mdpi. com/1424-8220/18/9/3133 42. Tian, F.: A supply chain traceability system for food safety based on HACCP, blockchain & Internet of things. In: 2017 International Conference on Service Systems and Service Management, Dalian, p. 6 (2017). https://doi.org/10.1109/ICSSSM.2017.7996119. https://ieeexp lore.ieee.org/abstract/document/7996119 43. Dujak, D., Sajter, D.: Blockchain applications in supply chain. In: Kawa, A., Maryniak, A. (eds.) SMART Supply Network. E, pp. 21–46. Springer, Cham (2019). https://doi.org/10. 1007/978-3-319-91668-2_2 44. Biswas, K., Muthukkumarasamy, V. & Tan, WL. (2017). Blockchain based wine supply chain traceability system. Future Technologies Conference (FTC) 2017, 56–62. United Kingdom:The Science and formationganiza- tion.https://researchbank.acu.edu.au/flb_pub/ 989/ 45. Kamble, S.S., Gunasekaran, A., Sharma, R.: odeling the blockchain enabled traceability in agriculture supply chain. Int. J. Inf. Manag. 52, 101967 (2019). https://doi.org/10.1016/j.iji nfomgt.2019.05.023 46. Zhang, J.: Deploying blockchain technology in the supply chain. Blockchain and distributed ledger technology (DLT) [Working Title], 1–16 (2019). https://doi.org/10.5772/intechopen. 86530 47. Zhao, G., et al.: Blockchain technology in agri-food value chain management: a synthesis of applications, challenges and future research directions. Comput. Ind. 109, 83–99 (2019) 48. Liu, Z., Li, Z.: A blockchain-based framework of cross-border e-commerce sup-ply chain. Int J. Inf. Manag, 52, 102059. https://doi.org/10.1016/j.ijinfomgt.2019.102059 49. Sternberg, H.S., Hofmann, E., Roeck, D.: The struggle is real: insights from asupply chain blockchain case. J. Bus. Logist. 42, 1–17. https://doi.org/10.1111/jbl.12240 50. Higginson, M., Nadeau M.C., Rajgopal, K.: Blockchain’s Occam Problem (2019). https:// www.mckinsey.com/industries/financial-services/our-insights/blockchains-occam-problem 51. How the blockchain enables and constrains supply chainperformance. Int. J. Phys. DIstrib. Logist. Mang. 49(4), 376–397. https://doi.org/10.1108/IJPDLM-02-2019-0063 52. Howson, P.: Building trust and equity in marine conservation and fisheries supply chain management with blockchain. Marine Policy 115, 103873 (2020), ISSN:0308-597X. https:// doi.org/10.1016/j.marpol.2020.103873s 53. Kamilaris, A., Fonts, A., Prenafeta-Bold, ´ F.X.: The rise of blockchain technology in agriculture and food supply chains. Trends Food Sci. Technol. 91, 640–652, ISSN:0924-2244. https://doi.org/10.1016/j.tifs.2019.07.034 54. Thematic Report - Building better supply chains with blockchain. https://www.eublockchain forum.eu/reports 55. Braue, D.: How blockchain could lead exports recovery (2020). https://ia.acs.org.au/article/ 2020/how-blockchain-could-lead-exports-recovery.html

Fish Control Process and Traceability

773

56. Tian, FA supply chain traceability system for food safety based on HACCP, blockchain & Internet of things. In: 2017 International Conference on Service Systems and Service Management, pp. 1–6. IEEE, June 2017 57. Molamohamadi, Z., Babaee Tirkolaee, E., Mirzazadeh, A., Weber, G.-W. (eds.): LSCM 2020. CCIS, vol. 1458. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-89743-7 58. Sun, H., Fang, Y., Zou, H.: (Melody) choosing a fit technology: understanding mindful- ness in technology adoption and continuance. J. Assoc. Inf. Syst. 17, 377–412 (2016) 59. Langer, E.J.: The Power of Mindful Learning. Addison-Wesley, Boston (1997) 60. Ksherti, N.: 1 Blockchain’s roles in meeting key supply chain management objectives. Int. J. Inf. Manage. 39, 80–89 (2018) 61. Mathisen, M.: The application of blockchain technology in norwegian fish sup- ply chains a case study. Masters’ thesis (2018). http://hdl.handle.net/11250/2561323

Impact of Socio-Economic Factors on Students’ Academic Performance: A Case Study of Jawahar Navodaya Vidyalaya Kapila Devi1 , Saroj Ratnoo1 , and Anu Bajaj1,2(B) 1 Guru Jambheshwar University of Science and Technology, Hisar, India 2 Machine Intelligence Research Labs (MIR Labs), Auburn, USA

Abstract. Quality education is an essential tool for students’ cognitive, intellectual, social, and personal development and to makes them responsible citizens of any nation. Identifying students’ strong and weaker intellectual skills and subject areas is essential for a continuous learning cycle and educational policymaking. So, to enhance the student’s overall growth as an individual, we need to analyse and predict the students’ academic performance. This paper studies the impact of socioeconomic factors on students’ academic performance using Jawahar Navodaya Vidyalaya (JNV) case study. JNVs are the schools established to provide quality education to the unprivileged and rural students. We collected 257 students’ data from (JNV) Khunga Kothi, Jind, Haryana, India, of five successive batches to examine their academic achievements from admission in 6th to their passing 10th class. The results show that the students’ socioeconomic variables, such as caste, residence, and father occupation, impact their academic performance in the 6th class but cease to do so after five years of their residential study. Moreover, the female students performed significantly better than the male students. Furthermore, we observed the difference in the students’ performance from admission to five years. The results indicate improvement in performance and a strong correlation between the 6th and 10th class marks. Therefore, we proposed a regression model that predicts the students’ performance to help the students at an early stage. Observations also suggest that delivering better learning opportunities to the students belonging to the unprivileged classes can improve their academic performance. Keywords: Students’ academic performance prediction · Regression analysis · Socio economic status · SES · Academic performance-Jawahar Navodaya Vidyalaya

1 Introduction Education is the stepping stone for the development of any nation. Various kinds of developments, such as cognitive, intellectual, social, personal, are possible through quality education. Before the independence of India, quality education had been a privilege and entitlement for rich and affluent households. The rural families remained impoverished © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 774–785, 2022. https://doi.org/10.1007/978-3-030-96299-9_73

Impact of Socio-Economic Factors on Students’ Academic Performance

775

and struggled even to obtain a proper education. After four decades of an Independence, then Prime Minister of India, Sh. Rajiv Gandhi aimed to establish a network of schools in every Indian district to facilitate quality education and assure students’ overall development. Jawahar Navodaya Vidyalaya (JNV) was the name given to such schools. JNVs were founded in 1986 as a result of the education policy of 1986 [1]. Since then, it has shown to be a silver lining in the realm of school education. JNVs select children and give them the highest quality education and interpersonal skills. Furthermore, schooling, housing, lodging, clothing, and textbooks are all provided free of charge. Students in Grades 9 to 12 would, nevertheless, be required to pay a monthly fee of Rs. 1500/-. Fees are waived for students belonging to the Scheduled Caste and Scheduled Tribe categories, and females, differently-abled pupils, and families living below the poverty level. JNVs show uniqueness in many ways. It was founded on the principles of inclusion and social justice, with a focus on impoverished students from rural regions. They also address the educational inequalities that emerge from the socioeconomic status (SES) of students’ families. SES can be determined by the social impact of a person and is often assessed as a combination of family’s income, educational and occupational attainment [2, 3]. In addition to the SES variables, student absence (unexplained), age, and gender can significantly predict the students’ academic performance. The relationship between students’ family background and achievement is often seen as an important topic in regard to equality and equity of educational provision [2]. Many studies show a strong association between SES and students’ academic performance. The SES like father’s education, mother’s occupation and family income are the good predictor of students achievements [9]. In [4] the authors reported a moderate to a strong association between SES and educational attainment. It was found that the school location and minority social status are the main factors influencing the relationship.Moreover, SES is not only associated with academic performance at school but also at higher educational institutions across the countries [3, 5–10]. As of 31st March 2020, 661 JNVs are functional in the country, with about 2 87,987 students enrolled. Performance of the JNVs is among the top-ranked central board of school education (CBSE) schools, having a pass percentage of 98.66% in 10th grades and similarly in 12th standard. Table 1 indicates the pass percentage of different educational boards like CBSE, Board of school education Haryana (BSEH), Kendriya Vidyalaya (KV) etc., in the span of the last few years. Most of the time, JNVs are on the top compared to the other boards and schools in India. Predicting student performance is an important issue as it depends on the student’s participation, engagement, effort and behavior [11–13]. The grades assigned to the students are the best way to assess the student performance as they reflect the student’s participation, engagement, effort and behaviour compared to the standardized test in mathematics and reading [12]. As reported by students, there is also a relationship between mean expected grades and the level of difficulty/workload in courses, which unduly influence student rating instruction [14]. In addition to this, grades also help identify the factors that serve as good indicators of whether a student will drop out or fail the program [15]. The student’s performance can be predicted based on pre-to-post-requisite skill relationships [16]. Moreover, the major subject papers also play a significant role in predicting the over-all

776

K. Devi et al. Table 1. Pass Percentage of 10th class of different boards and school

Years

CBSE

BSEH

JNV

KV

2015

97.87

30.32

99.76

99.33

2016

96.21

39.72

98.87

98.85

2017

93.06

43.06

99.68

99.66

2018

86.07

44.27

97.1

97.8

2019

91.1

52.71

98.57

99.47

2020

91.46

64.59

98.66

99.23

Table 2. Literature review Author

Year

Methodology/ technology

[17]

2018

Linear regression (LR), Random forests, GPR, Nearest neighbors regression, Regularized linear regression (LASSO)

[9]

2018

Multiple regression

[8]

2019

[18]

[19]

Data Set (No of students)

Performance parameter

Experimental results

Mean Square Error (MSE)

Grades in each semester can be predicted by using previous semester grades

823

Analysis of Variance (ANOVA)

SES are the good predictor of student academic achievement

Statistical analysis

548

Two-way ANOVA, chi-square test, t-test

Important variables with respect to learning are modality, gender, and class rank in environmental science concepts

2020

Multiple regression

492

Residual, R square

Gender, relation with parents, medium of school and first semester marks are the significant variable to predict the performance of the student

2021

LR, regression using deep learning

3889

10140

MSE, Mean Absolute Deep learning model Error (MAE) can be applied for smaller dataset and perform better

(continued)

Impact of Socio-Economic Factors on Students’ Academic Performance

777

Table 2. (continued) Author

Year

Methodology/ technology

[12]

2011

Multi-dimensional scaling

[15]

2013

Principal component analysis

Data Set (No of students) 4520

90

Performance parameter

Experimental results

Person product moment correlation

Grades assigned to the students are the best way to assess the student performance Mathematics courses are the best indicators of student performance

performance of the students [19]. [17] used machine learning techniques to predict the grades in each semester by using previous semester grades and found that the courses related to previous semester courses were correctly predicted. Another model revealed that the four variable gender, relation with parents, medium of school and first semester marks are the significant variable to predict the performance of the student [18]. Table 2 provides summary of related work that includes the methodology used in the study, size of the dataset, validating parameters and experimental results.

2 Methodology This section describes the case study, analysis methods and tools used for the experiment as follows: 2.1 Case Study There are 661 JNVs located all over India. The data is collected from Jawahar Navodaya Vidyalaya Khunga Kothi, Jind, Haryana. This JNV is randomly selected out of 21 JNVs located in Haryana (India). The JNVs mainly admit children in class 6th through a national Jawahar Navodaya Vidyalaya selection test carried out by CBSE and provide quality education to the selected students in residential environment up to 12th class. In this case study, we have collected five successive admission years’ records (from 2006–2007 to 2010–21011) of the 257 students. It is captured from the admission form filled by the students at the time of 6th class admission. It includes attributes related to the students’ socio-economic status and academic performance, as shown in Table 3. We took the results of the 6th and 10th classes as the indicator of academic performance of corresponding batches. The 6th class results are in percent, but the 10th class result is in grades as the CBSE provides grades. We then converted the 10th class grades into a percentage by multiplying them with 9.5 as defined by the board. Out of the 21 attributes, seven attributes are categorical, and fourteen attributes are numerical.

778

K. Devi et al. Table 3. Description of attributes used for analysing academic performance

S. No.

Attribute name

Description

1

Caste

Caste of the student i.e. SC, Gen, OBC

2

Gender

Gender of the student i.e. Male, Female

3

Residence

Locality of the Student i.e. Rural, Urban

4

Father.Edu

Father education of the student i.e. Illiterate, literate

5

Mother.Edu

Mother education of the student i.e. Illiterate, literate

6

Father.Occup

Father occupation of the students i.e. Agriculture, Labourer

7

Mother.Occup

Mother occupation of the students i.e. Agriculture, Labourer

8

Family_Income

Income of the student’s family

9

Family_Size

Number of family member

10

Hindi.6

Final Hindi Subject marks in 6th class

11

English.6

Final English Subject marks in 6th class

12

Maths.6

Final Mathematics Subject marks in 6th class

13

Science.6

Final Science Subject marks in 6th class

14

Social Science.6

Final Social Science Subject marks in 6th class

15

Hindi.10

Final Hindi Subject marks in 10th class

16

English.10

Final English Subject marks in 10th class

17

Maths.10

Final Maths Subject marks in 10th class

18

Science.10

Final Science Subject marks in 10th class

19

Social Science.10

Final Social Science Subject marks in 10th class

20

Percent.6

Final marks of 6th class in percent

21

Percent.10

Final marks of 10th class in percent

2.2 Analysis Methods and Tools Open source platforms R and RStudio are used to implement regression analysis and statistical analysis. R provides many inbuilt functions for hypothesis testing based on classical statistical tests, linear and non-linear predictive modelling and clustering analysis etc. We have applied multiple linear regression to observe the impact of 6th class marks as independent variables and the students’ 10th class marks as the response variable. The lm () function in R automatically uses dummy coding for categorical variables to build the model. The mean-variance inflation factors (VIFs) test is used to validate if the assumption of absence of multi-collinearity is met for the multiple linear regression model. The formula of the multiple linear regression model is given below: Y = βx + ε

(1)

Impact of Socio-Economic Factors on Students’ Academic Performance

779

In Eq. (1), Y represents the 10th class marks (in percent), is the estimated coefficients for the 6th class marks, and is the random error that accounts for the unaccounted variables.

3 Results and Analysis This section analyses the results obtained using descriptive, inferential and the multiple regression technique specifically for academic performance predictive model. Table 4 shows the descriptive statistics of students’ 6th and 10th class marks, which is further grouped by the different socio-economic variables and sub-variables described in the student data. It also presents the percent improvement, calculated from the mean marks of the 6th and 10th classes. The students who belonged to the scheduled caste improved by 20.60%, followed by general (17.74%) and OBC students (15.28%) while they transitioned from 6th class to 10th class. In addition to this, if we inspect the mean marks of these students in the 10th class, it can be said that the performance of non-general students is approximately equivalent with other categories of students, which was lower in the 6th class when they enrolled with the JNV. Also, the female students outperformed the male students with a gain of 21.14% compared to males (16%). Table 4. Descriptive statistics of variables for 6th and 10th class Attributes Subcategory Marks in percent of 6th class Q1 Caste

Gender U.R

Mean Median Q3

Marks in percent of 10th class Max Q1

Mean Median Q3

% Max improve

GEN

26.0 59.5

69.4

78.2 92.0

47.5 73.2

79.8

87.4 95.0

17.74

OBC

37.4 60.7

70.0

75.6 94.0

58.9 73.2

77.9

88.4 95.0

15.28

SC

26.0 55.8

65.4

72.9 92.4

53.2 72.2

77.9

83.6 95.0

20.60

Female

26.0 58.6

68.8

76.8 90.2

55.1 74.1

81.7

87.4 95.0

21.14

Male

26.0 56.7

67.0

76.7 94.0

47.5 70.3

77.9

83.6 95.0

16.83

Rural

26.0 56.4

67.0

76.4 94.0

47.5 72.2

79.8

85.5 95.0

19.59

Urban

45.8 59.8

69.7

78.7 90.4

58.9 70.3

77.9

84.6 95.0

12.35

Father. Edu

ILTR

40.8 54.8

63.9

75.2 88.4

53.2 68.4

73.2

85.5 95.0

18.69

LT

26.0 57.2

68.4

77.7 94.0

47.5 74.1

79.8

85.5 95.0

18.59

Mother. Edu

ILTR

26.0 55.2

65.4

75.8 92.0

53.2 70.3

76.0

83.6 95.0

17.67

LT

26.0 58.4

70.0

78.2 94.0

47.5 74.1

81.7

87.4 95.0

18.75

Father. Occup

AGR

38.8 60.0

69.7

78.2 92.0

53.2 72.2

79.8

87.4 95.0

16.15

LBR

26.0 54.6

63.8

73.2 91.2

47.5 70.3

76.0

83.6 95.0

21.22

SEMP

57.2 64.0

71.5

74.0 86.8

64.6 74.1

81.7

83.6 91.2

16.82

SVR

26.0 57.4

68.1

79.4 94.0

55.1 72.2

80.8

87.4 95.0

19.88

AGR

40.6 59.4

66.6

78.2 88.4

57.0 72.2

77.9

83.6 95.0

19.02

HW

26.0 57.5

67.6

77.7 94.0

47.5 72.2

79.8

85.5 95.0

17.20

LBR

47.0 52.8

67.4

70.2 80.4

53.2 76.0

79.8

84.6 95.0

26.49

SVR

37.4 52.4

67.3

71.6 82.4

58.9 77.9

79.8

81.7 93.1

25.66

Mother. Occup

780

K. Devi et al.

At the time of admission, the urban students have higher mean marks (59.8%), but in the 10th class results, the rural students score high (72.2%). We have also compared the performance of the students based on the literacy status of their parents. It is observed that the students of literate parents performed better as compared to the illiterates. Comparison on the father’s occupation revealed that the students of labourer parent showed more improvement and better performance than other categories of father’s profession. Table 5 discusses the different socioeconomic variables that influence the student’s academic performance when they join the school in 6th class and later after five years, i.e., in 10th class. Caste, residence and father occupation significantly affect the results in the 6th class as they spent more time at home than the school. But these variables didn’t influence the 10th results because the school is residential and provide equal facility to all the students. Gender is the variable that affects the student performance at 10th as the female student improves more than the male student. An educated parent can better motivate and guide their children. Therefore, the father and mother education significantly influence the performance of the student in the 10th class. Father occupation affected the student’s results, but mother occupation didn’t impact the student performance as most mothers are housewives. Table 5. Comparative results of t-test for 6th and 10th S. No.

Attributes

Percent.6 (p-value)

Percent.10 (p-value)

1

Caste

0.0359

0.08602

2

Gender

0.5105

0.002326

3

U.R

0.04104

0.6751

4

Father.Edu

0.1112

0.03974

5

Mother.Edu

0.05101

0.0009796

6

Father.Occup

0.0324 (lbr-Agr)

0.207

7

Mother.Occup

0.584

0.953

Table 6 compares the mean marks of different subjects to the overall mean marks in 6th and 10th classes. English with a mean of 60.16 significantly impacts the 6th class result as most of the children are from a rural area and have poor English. Therefore, English has the least mean in all the subjects, but the student performed better in Maths (69.58) and Science (71.28). In the 6th class, the mean marks in the three subjects English, Math, and Science are significantly different that the overall mean marks. In the 10th class, the academic performance is significantly better in Hindi (85.91), poorer in Math (76.33), and Science (75) when compared to the overall mean marks.

Impact of Socio-Economic Factors on Students’ Academic Performance

781

Table 6. Comparative results of subject-wise of 6th and 10th class S. No.

Attributes

Percent.6

Percent.10

Mean (66.4)

p-value

Mean (78.6)

p-value

1

Hindi

66.82422

0.7393

85.90661

2.2e−16

2

English

60.16406

3.52e−06

76.96109

0.05507

3

Maths

69.57812

0.02393

76.33268

0.0244

4

Science

71.27734

0.0002648

75.00195

0.0002256

5

Social study

64.14453

0.07887

79.10506

0.6305

Figure 1 pictorially represents the subject-wise comparison of the 6th and 10th marks. The students improved more in Hindi, English, and Social Science compared to Maths and Science. However, they were good in Maths and Science at the time of admission in the 6th class. It leads to further analysis of the impact of pairwise subject marks (see Table 7).

100 90 80 70 60 50 40 30 20 10 0

85.9 77 66.8

Hindi

69.6

76.3

71.3

60.2

English

Maths Percent. 6

79.1

75

Science

64.1

Social Study

Percent.10

Fig. 1. Subject-wise comparison of 6th and 10th class

It is observed from Table 7 that there is a significant difference between the mean marks of Hindi (66.82) with English (60.16) and Science (71.28) as the rural students are weak in English, but they are good at their mother tongue Hindi. They also have good performance in Science. But there is no significant difference between the mean marks of Hindi with Maths (69.58) and Social Science (64.14). If we compare English with all other subjects, there is a significant difference in their mean as English has the lowest mean (60.16) in all five subjects. If we compare Science and maths, there is an insignificant difference between the two. The pairs (Maths, Social Science) and

782

K. Devi et al. Table 7. t-test result of each subject with others for 6th class

S. No.

Attributes

Hindi

English

Maths

Science

1

Hindi

-

2

English

3.613e−06

-

3

Maths

0.06534

1.871e−09

-

4

Science

0.001816

1.863e−13

0.2694

-

5

Social study

0.05219

0.005509

0.0003124

7.872e−07

(Science, Social Science) are significantly different. The students are good in Maths and Science as compared to Social Science. Table 8. t-test result of each subject with others for 10th class S. No.

Attributes

Hindi

English

Maths

Science

1

Hindi

-

2

English

2.2e−16

-

3

Maths

2.2e−16

0.5469

4

Science

2.2e−16

0.0498

0.239

-

5

Social study

1.397e−15

0.02211

0.01014

7.73e−05

-

On the other hand, Table 8 provides the pairwise comparison of the subject results in the 10th class. It is found that Hindi is the top-scoring subject followed by Social Science, English, Maths and Science. Up to the tenth class, Social Science is in Hindi medium. So there is a significant difference between the mean of Hindi and all other subjects, as shown in Table 8. It is also observed that all language subjects are on top and quantitative subjects are at the bottom. However, Maths and Science are not significantly different. From the descriptive analysis, it is observed that there is a direct association between the overall marks of 6th and 10th class with socio-economic variables. It motivate us to check whether there is any relationship between the 6th and 10th class marks. Therefore, we analysed the correlation between the two using regression analysis (see Fig. 2). The experimental results show that the marks of the 6th class are strongly associated with the 10th class marks. Hence, we can formulate a mathematical equation to predict the marks of the 10th class based on 6th class marks as shown in Eq. (2). 10th(marks) = 0.53 × 6th(marks) + 43.80 + 

(2)

This early prediction helps the teacher to enhance the performance of the vulnerable students by providing them necessary education and holistic environment.

Impact of Socio-Economic Factors on Students’ Academic Performance

783

Fig. 2. Regression analysis results

4 Conclusion Education is the stepping stone for the development of any nation. Various kinds of developments, such as cognitive, intellectual, social, personal, are possible through quality education. JNVs were set up with inclusivity, social justice, primarily for underprivileged students from rural regions, and to address the educational inequalities that emerge from students’ families’ socioeconomic status (SES) that affect sustainable and equitable development. We analysed the 257 students’ records of JNV Khunga Khothi Jind. From the descriptive analysis, we found that the students enter the school with relatively low academic performance. Due to the equal facilities and opportunities, the underprivileged students improve their academics better than their counterparts with high socioeconomic status. At the entry-level, the students have low language skills but have better analytical skills. Three variables, namely caste, residence and father occupation, significantly impact students’ performance in the 6th class. Still, these variables did not affect their performance in the 10th class as the school is residential and separate them from their home background. It can be concluded from this case study that if equal and quality education is provided to the underprivileged and rural students by separating them from

784

K. Devi et al.

their background successfully, then the student can improve rigorously and perform similar to their privileged counterparts.

References 1. National Policy on Education (1986). http://psscive.ac.in/assets/documents/Policy_ 1986_eng.pdf 2. Brese, F., Mirazchiyski, P.: Measuring Students’ Family Background in Large-scale Education Studies. IERI (2013) 3. Considine, G., Zappalà, G.: The influence of social and economic disadvantage in the academic performance of school students in Australia. J. Sociol. 38, 129–148 (2002) 4. Sirin, S.R.: Socioeconomic status and academic achievement: a meta-analytic review of research. Rev. Educ. Res. 75, 417–453 (2005) 5. Broer, M., Bai, Y., Fonseca, F.: A review of the literature on socioeconomic status and educational achievement. In: Socioeconomic Inequality and Educational Outcomes. IRE, vol. 5, pp. 7–17. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-11991-1_2 6. Humble, S., Dixon, P.: The effects of schooling, family and poverty on children’s attainment, potential and confidence—evidence from Kinondoni, Dar es Salaam. Tanzania. Int. J. Educ. Res. 83, 94–106 (2017) 7. Li, Z., Qiu, Z.: How does family background affect children’s educational achievement? Evidence from contemporary China. J. Chinese Sociol. 5(1), 1–21 (2018). https://doi.org/10. 1186/s40711-018-0083-8 8. Paul, J., Jefferson, F.: A comparative analysis of student performance in an online vs. faceto-face environmental science course from 2009 to 2016. Front. Comput. Sci. 1, 7 (2019). https://doi.org/10.3389/fcomp.2019.00007 9. Singh, R., Choudhary, S.: An empirical study on the predictive role of SES on high school students’ achievement in learning ESL. Int. J. Learn. Chang. 10, 163 (2018). https://doi.org/ 10.1504/IJLC.2018.090925 10. Thomson, S.: Achievement at school and socioeconomic background—an educational perspective. npj Sci. Learn. 3(1), 1–2 (2018). https://doi.org/10.1038/s41539-018-0022-0 11. Atkinson, R., Geiser, S.: Reflections on a century of college admissions tests. Educ. Res. 38, 665–676 (2009). https://doi.org/10.3102/0013189X09351981 12. Bowers, A.J.: What’s in a grade? The multidimensional nature of what teacher-assigned grades assess in high school. Educ. Res. Eval. 17, 141–159 (2011). https://doi.org/10.1080/ 13803611.2011.597112 13. Khan, A., Ghosh, S.K.: Student performance analysis and prediction in classroom learning: a review of educational data mining studies. Educ. Inf. Technol. 26(1), 205–240 (2020). https:// doi.org/10.1007/s10639-020-10230-3 14. Centra, J.A.: Will teachers receive higher student evaluations by giving higher grades and less course work? Res. High. Educ. 44, 495–518 (2003). https://doi.org/10.1023/A:102549 2407752 15. Li, K.F., Rusk, D., Song, F.: Predicting student academic performance. In: Proceedings - 2013 7th International Conference on Complex, Intelligent, and Software Intensive Systems, CISIS 2013, July 1 (2013). https://doi.org/10.1109/CISIS.2013.15 16. Adjei, S.A., Botelho, A.F., Heffernan, N.T.: Predicting student performance on post-requisite skills using prerequisite skill data: an alternative method for refining prerequisite skill structures. In: Proceedings of the Sixth International Conference on Learning Analytics & Knowledge, pp. 469–473. Association for Computing Machinery, New York, NY, USA (2016). https://doi.org/10.1145/2883851.2883867

Impact of Socio-Economic Factors on Students’ Academic Performance

785

17. Backenköhler, M., Wolf, V.: Student performance prediction and optimal course selection: an MDP approach. In: Cerone, A., Roveri, M. (eds.) SEFM 2017. LNCS, vol. 10729, pp. 40–47. Springer, Cham (2018). https://doi.org/10.1007/978-3-319-74781-1_3 18. Kumar, A., Eldhose, K.K., Sridharan, R., Panicker, V.: Students’ academic performance prediction using regression: a case study. In: International Conference on System, Computation, Automation and Networking (ICSCAN), 3 July 2020. https://doi.org/10.1109/ICSCAN 49426.2020.9262346 19. Hussain, S., Gaftandzhieva, S., Maniruzzaman, M., Doneva, R., Muhsin, Z.F.: Regression analysis of student academic performance using deep learning. Educ. Inf. Technol. 26(1), 783–798 (2020). https://doi.org/10.1007/s10639-020-10241-0

Techno-Economic Feasibility Analysis and Optimal Design of Hybrid Renewable Energy Systems Coupled with Energy Storage Shirin Cupples1 , Amir Abtahi1,3(B) , Ana Madureira2,3 , and José Quadrado2,3 1 Florida Atlantic University, Boca Raton, FL, USA

{scupples2015,abtahi}@fau.edu 2 ISEP/P.PORTO, Porto, Portugal [email protected], [email protected] 3 ISRC - Interdisciplinary Studies Research Center, Porto, Portugal

Abstract. Renewable energy sources such as solar and wind are now competitive with traditional fossil and nuclear power “when generating” but that is just the challenge. When “not generating” can be a problem for grid integration and the main challenge to the widespread acceptance and dissemination of solar and wind, and the focus of research for the next generation of energy engineers. “Intermittent”, the adjective most associated with solar and wind energy has been and continues to be the focus of research by power engineers, AI professionals, and system scientists from the late 20th century and is the critical factor in the design of the future power grids, The most obvious solution is energy storage but then the choice of the storage method and size are complex problems. Will best solutions involve pumped hydro, Li-Ion batteries, or hydrogen generation? Or will next-generation ultra-capacitors, or high-speed flywheels gyros, or some yet to be discovered device will be the dominating technologies? The primary objective of the storage designs will be based on what’s best for the reliability and efficiency of the grid, and simultaneously optimizing cost and environmental impact functions. Socio-economic and geopolitical considerations must also be considered to satisfy local or regional constraints. There is also the question of purpose: will it be sized for grid stability, or medium, or long-term storage. This factor will depend on the specific grid requirements. The focus of this paper is to study multi-source renewable energy systems that include storage called HRES or Hybrid Renewable Energy with Storage. This study describes the development of a behind-the-meter Energy Management System (EMS) for an HRES, under the assumption that Real-Time Pricing (RTP) is offered by a utility supplying power to a medium-size office complex. A cost function to be minimized is introduced to optimize the contribution of each energy source. Also, this work develops the basis of a platform for decision-makers to evaluate the viability of the optimized system in conjunction with the financial feasibility analysis. Keywords: Renewable energy · Distributed generation · Battery storage · Real-time pricing

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 786–796, 2022. https://doi.org/10.1007/978-3-030-96299-9_74

Techno-Economic Feasibility Analysis and Optimal Design

787

1 Introduction HRES has been proposed to be an integral component in the enable higher penetration of sustainable and clean energy generation [1]. The most important advantage of HRES is to make the best use of the renewable power generation technologies’ operating characteristics and to obtain efficiencies higher than could be obtained from a single power source. It can also address limitations in terms of fuel flexibility, efficiency, reliability, emissions, and economics. Various aspects must be considered when working with hybrid systems for the generation of electricity. Reliability and cost are two of these aspects [2]. There are two ways to increase the resiliency of HRES and distributed energy resources (DERs): Combing two or more renewable energy sources to supply the targeted load and; Connection to a reliable grid and/or energy storage [2, 3]. With RTP tariffs, electricity consumers are charged prices that reflect updated and current prices for short time intervals, typically ¼ hourly or hourly, and are quoted one day or less in advance. RTP differs from conventional retail tariffs, which are based on prices that are fixed for months or years, inflation-adjusted, reflecting fairly constant embedded supply costs. In recent years, a resurgence of interest in RTP has occurred. Economists recognize that providing consumers with price incentives to reduce their usage when wholesale prices rise would improve the performance of wholesale electricity markets in two important ways: mitigating suppliers’ ability to exercise market power and moderating price volatility. In grid-connected HRES, if RTP of electricity is implemented, an end-user with a local microgrid would typically prefer to get the lowest-priced electricity from the HRES at peak hours when the price on the market is high. Energy storage will not only help to achieve this goal, but it can also enhance the reliability and resiliency of the grid through short-term storage for peak-shaving and power quality purposes. For grid-connected HRES, energy storage becomes a critical tool of compensating for generation fluctuations of wind and solar energy, on timescales ranging from seconds to hours. The rest of this paper is organized as follows: The next sections are a brief stateof-the-art that covers important aspects and results reported by other authors while pursuing similar objectives. Then we present in detail the methodology that was implemented for the proposed approach and proceed with the computational study and the performance assessment. Finally, we present the main conclusions and some ideas for future developments.

2 Literature Review There is substantial evidence an HRES with integrated storage increases the overall system efficiency and reliability of an associated grid [4]. HRES can be defined as a subgrid of interconnected loads and generation units that can seamlessly connect/disconnect with the main grid but also operate in a standalone mode. Maleki et al., have shown that systems which rely on multiple sources of generation are more reliable than those that rely on a single intermittent source [5]. Furthermore, for remote (power) islands and offgrid communities, HRES have also proven to be economically superior to the existing electric grid [6].

788

S. Cupples et al.

The design of the HRES with an optimum combination of renewable sources is crucial to overcoming intermittency and ensuring economic feasibility. The optimum design and size of an HRES rely on good demand data and improved predictive models for the frequency and availability of renewable resources. The optimization sizing problem is non-convex and non-linear resulting in non-unique solutions. In addition, the sheer number of input variables (installed PV and radiation data, WT specifications and wind data, the battery capacity and state of charge (SOC) information, etc., makes the sizing procedure time-consuming [5]. As a result, novel optimization methods are being applied for faster and less costly computational methods. Maleki et al. adopted an artificial bee swarm optimization (ABSO) algorithm to design and size the lowest cost of the HRES that was composed of PV, WTs, FCs [5]. Shang et al. used an improved PSO technique to size batteries for a hybrid energy system composed of PV, WT, and distributed non-renewable generation (DG) systems. A cost-minimizing objective function was developed to analyze scenarios with different penetration levels of renewable energy in the hybrid system [6]. For a remote area in Iran, Maleki et al. designed a PV, WT, and battery-based hybrid energy system that optimized for the least cost by using a PSO-based Monte Carlo method [7]. Ramli et al. used a self-adaptive differential evolution algorithm with multiple optimization objectives such as minimization of cost, minimization of loss of power supply probability while maximizing the renewable fraction, for an HRES in Yanbu, Saudi Arabia. This system was composed of PV, WT, batteries, and DGs [8]. Using the HOMER software, Phurailatpam et al. sized a microgrid for rural and urban applications in India that had the lowest net present cost while also evaluating CO2 emissions [9]. To size a microgrid, another study performed a triple multi-objective optimization analysis using a genetic algorithm (GA) to minimize the electricity cost, CO2 emissions, and unmet loads [10]. Some other studies have also utilized multi-objective optimization techniques to size a microgrid while simultaneously minimizing LCOE (Levelized Cost of Energy) and CO2 emissions [11, 12]. These studies were mainly focused on finding an optimal size of a microgrid that had the lowest levelized cost of energy (LCOE) and LPSP value [5–9]. Some studies also evaluated the carbon footprint of the microgrid [10–16]. Researchers also used commercial software tools such as Hybrid Optimization of Multiple Energy Resources (HOMER), Distributed Energy Resources Customer Adoption Model (DER-CAM), and TRNSYS to perform a techno-economic and environmental assessment based on location, load, and renewable sources [9, 16]. 2.1 Load and Power Generation Forecasting The main purpose of load forecasting is to meet the energy demand. Precise forecasts can help in minimizing the operating and maintenance costs while maintaining the high reliability of the power supply. The power prediction can be adapted more accurately in real situations by applying linear predictive models. Predictive modelling uses statistics to predict future outcomes. Figure 1 shows a 72-h-ahead demand load, based on historical data for a medium-size office in Sacramento, California. The inherent intermittent nature of renewable resources raises the question of reliability. However, wind and solar power prediction models have smoothed the way for the integration of renewable energy generation into the grid. Based on the meteorological

Techno-Economic Feasibility Analysis and Optimal Design

Fig. 1. 72 h ahead PV production

789

Fig. 2. 72 h load forecast-mid-size office

Fig. 3. 72 h ahead wind power production forecast

data of Sacramento, California, and the historical database of the PV and wind power, a forecasted 72-h ahead of available solar and wind power is presented in Figs. 2 and 3. Figure 2 shows solar forecast output power for 3 days ahead of a 20-kW photovoltaic system whereas the plot in Fig. 3 demonstrates the output forecast power curve of a wind turbine with 3 square meters of swept area. 2.2 Hybrid System Model The hybrid behind-the-meter system architecture studied in this paper is shown in Fig. 4. Hourly raw meteorological data such as Global Horizontal Irradiance (GHI), temperature, and wind speed (WS) were collected for Sacramento, California (38.5816° N, 121.4944° W) as a case study from [18]. A year-long data is used to train the model considering all pertinent seasonal and daily patterns in the simulation. Also, a regression fill model is used for missing values in the data. Table 1 shows the model’s input, and the sources used for the validation sample. The initial step for the optimization of a HERS performance is the modeling of individual subsystems. As shown in Fig. 4, HRES

790

S. Cupples et al. Table 1. Hourly time-series of stochastic variables

Aspect

Challenges

Solar irradiation (W/m2 )

Vector of 8760 point (365 days × 24 h)

[15]

Temperature (°C)

Vector of 8760 point (365 days × 24 h)

[15]

Wind speed (m/s)

Vector of 8760 point (365 days × 24 h)

[15]

Electricity demand profile (kW)

Vector of 8760 point (365 days × 24 h)

[16]

Real time pricing (cent/kW)

Vector of 8760 point (365 days × 24 h)

[17]

Fig. 4. A schematic of the system model illustrating the important components and power flows. Because the inverter is assumed to be lossless it is not shown in this diagram

consists of four different elements: public grid, solar photovoltaics, wind turbine, and battery bank. In Fig. 4 shown above, the Central Energy Manager (CEM) controls the system. CEM in HERS can be integrated with a smart meter that can send and receive information about hourly electricity costs and use this information to proportionally distribute the energy between the different HRES components and the grid. The load is met in each time interval by a combination of grid purchases at real-time prices, PV and wind power production, and/or battery discharge. P load (t) = P Gr (t) + P dch (t) + P PV load (t) + P wind load (t)

(1)

Techno-Economic Feasibility Analysis and Optimal Design

791

Power provided from these sources can be optimized by CEM at any time from the forecasted loads’ data, solar, and wind power production. The solar energy output of a photovoltaic system depends on two primary variables, the solar irradiation, and the ambient air temperature, as shown by the following equation: P PV (t) = APV ξ PV R(t)

(2)

Where R is solar irradiance (W/m2 ), ξPV is the PV generation efficiency, and APV is the total solar panels’ area. Ambient air temperature affects the efficiency in the following manner: (m2 )

ξ PV = ξ r[1 − β(T c − T cref )]

(3)

Where ξ r is the reference module efficiency, β is the temperature coefficient, and Tcref is the reference cell temperature in degree Celsius. For this study, Tcref = 25 °C, and ξ r = 25% and β = 0.08%°C. At each instant, the solar energy output can be categorized into: PPVload which is considered part of the normal operation; PPVbat to charge the batteries; PPVlost for transmission losses or if battery bank’s maximum state of charge has been reached (SOCmax ). APV ξ PV R(t) = P PV load (t) + P PV bat (t) + P PV lost (t)

(4)

Fig. 5. Typical WTG power curve.

Wind Turbines Generator Systems (WTG) convert the wind kinetic energy to power as per Eq. (5). Figure 5 shows a typical power curve for a WTG. For speeds between cut-in rated speed, the power output Pwind is given by [12]: P wind (t) =

1 ρAwind V 3w (t)C p (λ) 2

(5)

Where ρ is the air density (kg/m3 ), Vw is the wind speed (m/s), Awind is the turbine’s blades sweep area (m2 ), and Cp is the power coefficient which is the function of the rotor tip-speed to wind speed ratio (λ):   −21 116 − 0.4β − 5 e λi + 0.006λ (6) Cp(λ) = 0.5176 λi

792

S. Cupples et al.

1 ωR = λ+0.08β − β0.035 3 +1 and where λ = V , and β is the pitch angle. The WTG output energy, like PV, can be divided into three parts: Pwindload to support the load; Pwindbat to charge the batteries; and lost energy due to transfer through wires, or if the battery maximum state of charge has been reached (SOC max ). 1 λi

1 ρAwind V 3w (t)C p (λ) = P wind load (t) + P wind bat (t) + P wind lost (t) 2

(7)

The total system lost as the result of battery size and line losses is: P lost (t) = P wind load (t) + P PV lost (t)

(8)

Battery State of Charge (SOC), is the indication of when a battery needs to recharge and can be described as the difference of the battery input and output energies up to the desired time. E0 is the initial condition, assumed as zero in this study. SOC(t) = SOC(t − 1) + ξ bat P PV bat (t − 1) + ξ bat P wind bat (t − 1) − P dch (t − 1) + E0 (9) SOC(t) =

t−1 i=1

(ξ bat P PV bat (i) + ξ bat P wind bat (i) − P dch (i)) + E0

(10)

3 Optimization Methodology The techno-economic feasibility analysis is simulated for a “behind the meter” HERS, through a two-step optimization strategy. The initial step in the optimization process is to minimize the cost of electricity by applying RTP in the objective function. For the second step, the results from the first step are used in an economic sensitivity analysis where the outcome is optimized by aggregating system parameters, assumptions, and predictions in a cost-benefit model. The system is simulated with MATLAB software. 3.1 Optimizing the Electricity Cost This section focuses on lowering electricity costs through demand charge reduction which can be achieved through the short term “demand shaving. Controlling energy usage spikes is the key to reducing electricity bills. The crucial question that should be answered in this step is: at what time, for how long, and by what fraction, should the batteries be charged or discharged. The outcome of this section is the most reasonable hourly energy price for the customer, given the specific solar panel area, wind turbine size, and battery capacity. Objective Function Considering the real-time price of the electricity, the end user’s electricity cost over the scheduling horizon of t will be: min C =

t i=1

(P Gr (t).C(t))

(11)

Techno-Economic Feasibility Analysis and Optimal Design

793

C(t) is the real-time pricing for the location of the study. Constraints: 1. Grid: The power supplied by the grid, PGr (t) is a positive value at any given time. This means that the energy generated by the hybrid system will support loads or charge the battery. From Eq. 1, 4 and 9 results in: P Gr (t) ≥ 0 ⇒

(12)

P load (t) − P PV load (t) + P PV bat (t) − P wind load (t) + P wind bat (t) + P lost (t) − P dch (t) ≥ 0

(13) 2. Battery State of Charge (SOC): The SOC is the main control function to perform with an HRES. Most rechargeable batteries are not meant to be fully discharged. The minimum state of charge is typically set to 30–50% to avoid damaging the storage bank by excessive discharge. Also, stored energy in the batteries cannot be greater than their maximum capacity. From Eq. 10, the following constraints result:

SOC min ≤ SOC(t) ≤ SOC max ⇒ SOC min − E0 ≤

t−1 i=1

(14)

(ξ bat P PV bat (i) + ξ bat P wind bat (i) − P dch (i)) ≤ SOC max − E0 (15)

3. Battery Discharge: The discharge energy at any time interval cannot be greater than available stored energy. Considering Eq. 10 the constraint will be: P dch (t) ≤ SOC(t) ⇒  t−1  P dch (t) ≤ i=1 ξ bat P PV bat (i) + ξ bat P wind bat (i) − P dch (i) + E0

(16)

4. Energy loss: Energy loss due to battery limitation and transmission at each period, should be less than the total energy generated by HRES. P lost (t) ≤ P PV (i) + P wind (i) ⇒ P lost (t) ≤ APV ξ PV R(t) + 21 ρAwind V 3w (t)C p (λ)

(17)

3.2 Optimizing the Efficiency of Investment The main objective of this section is to quantify the projected profit margins and feasibility of the HRES. In this work, three economic figures of merit (FOM) were used as the objective functions for economic evaluation, including: Net present value: NPV is a critical economic metric to validate the profitability of a project and it is primarily based on four major components: cash inflows (C t ), initial investment (C 0 ), discount rate (r) and the duration of the project (t): NPV =

T t=1

Ct − C0 (1 + r)t

(18)

794

S. Cupples et al.

Discount payback period: DPP, refers to the period required to recoup the expense of an investment for a fixed discount rate, it is defined as the years of operations such that NPV equals 0. T=DPP t=1

Ct = C0 (1 + r)t

(19)

Internal rate of return: IRR, also called the effective interest rate, is defined as, for a fixed T year of operations, the value of rIRR such that NPV equals 0. T t=1

Ct = C0 (1 + rIRR )t

(20)

4 Conclusions and Future Work 4.1 Minimizing the Electricity Bill An optimization exercise was performed for a small office building with a 20 kWh battery bank, a 90 square meter PV area, and 3 square meters of sweep area. Figure 6 shows the resulting optimized solution as compared to the initial monthly cost for January 2021. The initial electricity bill for this month was $1501, and the cost was reduced by 465 to $895.

Fig. 6. Daily initial vs. Optimized costs for january

4.2 Maximizing Return on Investment (ROI) Using the results of the first step, Fig. 7 shows the results for maximizing ROI. The system will recover its cost within 15 years. After the payback period, a profit of roughly $74,000 with a rate of return of 3.6%. The optimization exercise assumed a total of $5,000 capital, 20-year life for PV and WTG, a 5-year life for batteries at a capital cost of $24/kWh, O&Ms of 2.5% for WTG, and .66% for PV, and an inflation rate of 1.5%.

Techno-Economic Feasibility Analysis and Optimal Design

795

Fig. 7. Cash flow analysis represented as cumulative NPV for the HRES

The authors are working on expanding the utility of this model to a range of applications and power ratings, with different ratios of load to battery storage, as well as the ratio of PV to WTG. Hydrogen generated by stranded PV or WT will also be included alongside other storage units as either loads or generators.

References 1. Daniel A., Bohm B., Coughlin B.J., Layne M.: Evaluating implications of hydraulic fracturing in Shale gas reservoirs. In: SPE Americas E&P Environmental safety conference, San Antonio, TX, USA (2009) 2. Villaran, M.: Optimal planning and design of hybrid renewable energy systems for microgrids. Renew. Sust. Energy Rev. 70, 180–191 (2017). https://doi.org/10.1016/j.rser.2016.10.061 3. Billington, R.: Evaluation of different operating strategies in small standalone power systems. IEEE Trans. Energy Convers 20(3), 654–660 (2005). https://doi.org/10.1109/TEC.2005. 847996 4. De Bosio, F., Luna, A.C., Ribeiro, L., Graells, M., Saavedra, O.R., Guerrero, J.M.: Analysis and improvement of the energy management of an isolated microgrid in Lencois island based on a linear optimization approach. IEEE (2016). https://doi.org/10.1109/ECCE.2016.7854871 5. Maleki, A., Aksarzadeh, A.: Artificial bee swarm optimization for optimum sizing of a standalone PV/WT/FC hybrid system considering LPSP concept. Sol. Energy 107, 227–235 (2014). https://doi.org/10.1016/j.solener.2014.05.016C 6. Shang, C., Srinivasan, D., Reindl, T.: An improved particle swarm optimization algorithm applied to battery sizing for stand-alone hybrid energy power systems. Electr. Power Energy Syst. 74, 104–117 (2016). https://doi.org/10.1016/j.ijepes.2015.07.009 7. Maleki, A., Khajeh, M.G., Ameri, M.: Optimal sizing of a grid independent hybrid renewable energy system incorporating resource uncertainty. Int. J. Electr. Power Energy Syst. 83, 514– 524 (2016). https://doi.org/10.1016/j.ijepes.2016.04.008 8. Ramli, M.A.M., Bouchekara, H.R.E.H., Alghamdi, A.S.: Optimal sizing of PV/wind/diesel hybrid microgrid system using multi-objective self-adaptive differential evolution algorithm. Renew. Energy 121, 400–411 (2018). https://doi.org/10.1016/j.renene.2018.01.058 9. Phurailatpam, C., Rajpurohit, B.S., Wang, L.: Planning and optimization of autonomous DC microgrids for rural and urban applications in India. Renew. Sust. Energy Rev. 82, 194–204 (2018). https://doi.org/10.1016/j.rser.2017.09.022 10. Dufo-Lopez, R., Agustin, J.L.: Multi-objective design of PV-wind-diesel-hydrogen battery systems. Renew. Energy 33, 2559–2572 (2008). https://doi.org/10.1016/j.renene.2008.02.027

796

S. Cupples et al.

11. Katsigiannis, Y.A., Georgilakis, P.S., Karapidakis, E.S.: Multiobjective genetic algorithm solution to optimum economic and environmental performance problem of small autonomous hybrid power systems with renewables. IET Renew. Power Gener. 4(5), 404–419 (2010). https://doi.org/10.1049/iet-rpg.2009.0076 12. Pelet, X., Favrat, D., Leyland, G.: Multiobjective optimisation of integrated energy systems for remote communities considering economics and CO2 emissions. Int. J. Thermal Sci. 44(12), 1180–1189 (2005) 13. Elsied, M., Oukaour, A., Gualous, H., Brutto, O.A.L.: Optimal economic and environment operation of micro-grid power systems. Energy Convers. Manage. 122, 182–194 (2016). https://doi.org/10.1016/j.enconman.2016.05.074 14. Lyden, A., Pepper, R., Tuohy, P.: A modeling tool selection process for planning of community scale energy systems including storage and demand side management. Sust. Cities Soc. 39, 674–688 (2018). https://doi.org/10.1016/j.scs.2018.02.003 15. “The National Solar Radiation Data Base (NSRDB)”. https://nsrdb.nrel.gov/ 16. “Pacific Gas and Electric, PG and E.” https://www.pge.com/Utility 17. Rate Database (URDB). https://openei.org/ 18. Yang, H., Wei, Z., Chengzhi, L.: Optimal design and techno-economic analysis of a hybrid solar–wind power generation system. Appl. Energy 86(2), 163–216 (2009)

Information Technology Roles and Their Most-Used Programming Languages Oluwaseun Alexander Dada1,2(B) , Kehinde Aruleba3 , Abdullahi Abubakar Yunusa4,5,7 , Ismaila Temitayo Sanusi6 , and George Obaido8 1 Department of Computer Science, University of Helsinki, Helsinki, Finland

[email protected]

2 The School of Software, Lagos, Nigeria 3 Department of Information Technology, Walter Sisulu University, Mthatha, South Africa 4 Center for Instructional Technology and Multimedia, Universiti Sains Malaysia, USM,

Gelugor, Penang, Malaysia 5 Department of Curriculum Studies and Educational Technology, Usmanu Danfodiyo

University Sokoto, Sokoto, Nigeria 6 School of Computing, University of Eastern Finland, Joensuu, Finland 7 Centre for Multidisciplinary Research and Innovation (CEMRI), Abuja, Nigeria 8 Department of Computer Science and Engineering, University of California, San Diego, USA

Abstract. In this digital age, almost every area of our lives and businesses are controlled by Computer code. Consequently, programming skills are now in high demand across industries - not only for those in software development but also for almost every career role. According to a study conducted in the United States, over 7 million out of 26 million online job postings were reported to have required programming skills. In this study, we identified and ranked the ‘most-used’ programming languages for each of the 23 different job roles in IT teams. Knowing the ‘most-used’ programming languages for a given role will help stakeholders plan adequately. Our results show that JavaScript was the overall most-used language, followed by HTML/CSS and then SQL. However, in terms of specific roles, HTML/CSS was the most-used language for Marketing/Sales, JavaScript for those in software development, Python for Data Scientists, Academic researchers and Scientists. SQL for Database Administrators, Data Engineers and Data Analysts. The outcome of this study can be used by relevant stakeholders to make decisions regarding careers, professional development and curriculum designs. Keywords: Information technology · Software developer · IT job role · Skills · StackOverflow · JavaScript · Python · Programming languages

1 Introduction In this digital age, almost every area of our lives and businesses are controlled by Computer code. Consequently, programming skills are now in high demand across industries - not only for those in software development but also for almost every role. According © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 797–805, 2022. https://doi.org/10.1007/978-3-030-96299-9_75

798

O. A. Dada et al.

to a study conducted by [1], over 7 million out of 26 million online job postings in the United States were reported to have required programming skills. Recent studies have shown that the continuous advancement in technology will cause dramatic changes in the nature of work, resulting in the loss of many jobs and the creation of several new ones [2, 3]. Within the context of IT, 42% of the required technical skills are expected to shift between 2018 and 2022 [4]. Thus, recruiters are now targeting candidates with ever-changing technological skillsets that are the most relevant to the industry. On the other side, IT job seekers and professionals must understand the appropriate technical skills they are expected to have to achieve their career aspirations. Interestingly, the need to upgrade existing skillsets to meet and match changing technology requirements is not limited to job seekers but also to other stakeholders involved in the IT marketplace, including academic institutions, teachers, students and professionals. The most useful approach for job seekers is to focus on the skills most sought after in their industries – but how shall they find out what those skills are and how important they are? This information may be inferred from annually published lists of the mostused technologies, such as Gartner Top Strategic Technology Trends for 2021 [5], Top 10 Hot Artificial Intelligence (AI) Technologies [6] and Most Popular Technologies for 2020 [7]. However, the top skills required for a software developer may differ from those needed for other IT roles, such as project manager or data engineer. In the same vein, different industries may demand different IT skills. The banking industry requires data security skills and very strict discipline in deploying standards, whereas the entertainment industry requires media-handling skills and a degree of creativity. In this paper, we aim to identify the ‘most-used’ programming languages for different IT roles in organisations. Knowing the most-used technologies for a given IT role (e.g. mobile developer or database administrator) will help stakeholders understand the relevant skills required in that role. Conversely, this knowledge may also help individuals (who already have such skills) in different roles to make decisions regarding changing jobs/roles. The output of this research will benefit professionals already working in the IT industry and individuals planning to work in the IT field. As such, the outcome of this study can be used by various stakeholders (including students, educators, curriculum experts, IT professionals and career advisors) to make decisions regarding technical skills that are increasingly relevant in specific IT roles.

2 Literature Review Programming skills have become an important requirement in many for the high-paying jobs [8]. In the United States, [1] reported that jobs requiring programming skills paid over 20,000 dollars per year more than jobs not requiring programming skills. Another interesting observation from the research is that programming languages with higher applicability attracted higher demand more than those with less. For instance, comparing R programming and Python languages. Python would attract higher demand because R programming is mostly used for statistical analysis while Python can be used for statistical analysis as well as in many other domains such as web development (using Django and Flask), backend programming, scheduling/automation [9]. Taking a general look at trends in the popularity of programming languages as shown in Fig. 1, JavaScript, HTML/CSS and SQL maintained their positions as some of the most

Information Technology Roles and Their Most-Used Programming Languages

799

commonly used programming languages. Also, TypeScript programming language has gained more popularity over C language. While Ruby, which was in the top most popular programming language in 2017 has dropped, now outshined by Kotlin and Go – which are newer languages [7].

Fig. 1. The most popular programming languages based on the 2019 annual survey [10].

Awareness and understanding of the required programming languages for a given role will help individuals focus on how to acquire the specific skills most relevant to their career aspirations. As newer technologies emerge, programming languages sometimes lose or gain relevance for a given career role. To keep themselves up-to-date regarding technology changes and trends, it is essential that individuals should belong to relevant technology communities. One of such communities is StackOverflow. It is regarded as one of the most active communities for developers to share their programming knowledge [11]. StackOverflow [7] conducts a yearly survey and makes the anonymised results available for download to anyone, which makes it possible to perform further analysis. As a result, researchers have drawn several insights from the publicly available StackOverflow Developer Surveys (SODS) dataset [12–17]. The studies conducted using SODS data have been observed to adopt various analyses and drew conclusions that contribute to knowledge on IT skills. For example, [18] tested the correlations between skills and competencies in the job offers data set retrieved from

800

O. A. Dada et al.

StackOverflow. Similarly, [19] used the SODS data set to gain insights on the directions of programming languages, databases and job seeking status of the developers. Relatedly, [20] analysed 2018 data using data visualisation techniques to extract insights about how genders see their confidence in programming skills. Also, [21], using descriptive statistics (including correlation and Anova), analysed the 2019 StackOverflow annual survey and found top languages that are being actively used by software professionals considering gender and age differences in language usage.

3 Methodology We used the publicly available StackOverflow Developer Surveys (SODS) dataset. StackOverflow is an online community of technology enthusiasts, regarded as the biggest platform for Computer Science Q&A [7]. Since 2011, the platform has been conducting similar surveys on an annual basis. Nearly 65,111 technology enthusiasts comprising diverse representation from multiple races and ethnic groups took the survey. We started by analysing the different roles, followed by their programming languages they considered to be most relevant to their daily tasks. Table 1 shows the research questions to which the participants responded. Table 1. Focus areas and corresponding survey questions Focus areas

Survey questions

Role in the team

Which of the following describe you? Please select all that apply

Programming skill Which programming, scripting and markup languages have you done extensive development work in over the past year, and which do you want to work in next year? (If you both worked with the language and want to continue to do so, please check both boxes in that row)

Our data analysis was done using Python programming libraries such as pandas, collections, NumPy and matplotlib. The analysis revolved around the different roles present in the dataset. The survey question about IT roles made it possible for a respondent to specify multiple roles, e.g. “Designer; Developer, front-end; Developer, mobile”. In the example, the respondent has the following three different roles: (a) Designer (b) Developer, front-end (c) Developer, mobile. For the sake of simplicity, we split the response such that the different roles were grouped separately. To compute the top most used technical skills in each IT role, we constructed a pool of the relevant skills (programming languages) based on the survey responses. We then calculated the percentage frequency for each technology and ranked them in descending order.

4 Results and Discussion In total, 23 unique roles were identified, and the role frequencies were computed. Figure 2 shows the percentage composition of the respondents based on their roles. The ten

Information Technology Roles and Their Most-Used Programming Languages

801

most frequent roles were: Developer, back-end (N = 27,228; 15.8%), Developer, fullstack (N = 27,125; 15.8%), Developer, front-end (N = 18,296; 10.6%), Developer, desktop or enterprise applications (N = 11,784; 6.8%), Developer, mobile (N = 9,482; 5.5%), DevOps specialist (N = 5,969; 3.5%), Database administrator (N = 5,722; 3.3%), Designer (N = 5,321; 3.1%), System administrator (N = 5,242; 3.0%), Developer, embedded applications or devices (N = 4,750; 2.8%).

Fig. 2. Frequency of the IT roles in the dataset

Based on the frequency of occurrence, the top ten languages for a given role are plotted as a stacked bar chart as illustrated in Fig. 3. How to read the chart. In Fig. 3, every role is represented by a bar (rectangle) which is subdivided into ten segments. The length of each segment represents the percentage frequency of a specific programming language denoted by a specific colour - as indicated in the legend. The individual bars are constructed by stacking the top ten programming languages for each role vertically end-to-end. The bottom-most segment represents the percentage frequency of the tenth most-used language for a given role and that of the ninth mostused language is placed on top. The same is repeated for the eighth, seventh, sixth and so on until the percentage frequency of number-one most-used language for that role is added on top of the remaining nine. Consider the following examples. Front-end developer role: C++ (16.05%), Bash/Shell/PowerShell (29.33%), Python (31.3), PHP (32.6%), C# (33.54%), Java (33.73%), Typescript (39.02%), SQL (56.39%), HTML/CSS (78.31%) and JavaScript (84.01%).

802

O. A. Dada et al.

Fig. 3. Top programming skills used in various IT roles

DevOps specialist: Go (19.92%), PHP (26.55%), Typescript (32.17%), C# (35.13%), Java (38.43%), Python (51.77%), Bash/Shell/PowerShell (58.69%), HTML/CSS (62.71%), SQL (63.06%), JavaScript (70.3%). Database administrator: C++ (20.45%), Typescript (24.24%), Java (35.04%), C# (40.39%), PHP (41.51%), Python (42.24%), Bash/Shell/PowerShell (43.41%), HTML/CSS (73.07%), JavaScript (74.27%) and SQL (80.08%). Our results show that JavaScript was the overall most-used language, followed by HTML/CSS and then SQL. This finding aligns with the existing literature and the preliminary analysis of [7]. In terms of specific roles, HTML/CSS was the most-used language for Marketing/Sales, JavaScript for those in software development, Python for Data Scientists, Academic researchers and Scientists. SQL for Database Administrators, Data Engineers and Data Analysts. With the recent advances in technology and data usage, having programming skills will likely increase the chances of a candidate getting an IT job. JavaScript is the programming, scripting, and Markup language that most respondents have used extensively. This can be due to the flexibility of JavaScript, i.e., its ability to unify all of the web development around single programming rather than using a different language for the client-side and another for the server-side. According to [22], JavaScript is responsible for more than one-third of the popular applications on GitHub; the following five languages (Ruby, Objective-C, Python,

Information Technology Roles and Their Most-Used Programming Languages

803

Java, and PHP) are responsible for another third of the popular applications. Further, this finding can be connected to the nature of wide usage of the internet for daily communication and information sharing where most IT solutions are web-based. Hence, a flexible approach for developing the web can create more demand. Consequently, a huge opportunity to develop young students in JavaScript programming skill can be explored at the higher education level where educators can create more courses to facilitate the skill. Besides, HTML/CSS was rated very highly as a top programming language next to JavaScript on several occasions. This is undoubtedly due to its extensive use in designing and developing web interfaces and other presentation materials. Python, on the other hand, has also gained popularity in the scientific domain. It is frequently used for high-performance scientific applications and is widely used by scientists, academic researchers, and data engineers because it performs well and is easy to write. As shown in Fig. 3, several developers also showed interest in using python more than C++ and C, which are older programming languages. Although C and C++ are relatively old languages, they are still relevant in developing embedded applications and banking software. We also observed that SQL consistently ranked in the top three for 20 of the 23 roles, and it was unanimously selected as the most used language by all the IT roles belonging to the Data category. This lends credence to the importance of data management in almost every aspect of our daily tasks.

5 Conclusion This study analysed the SODS data for the year 2020. The data set contains 23 different IT roles. The majority of the respondents work in IT roles that fall under the software development category. Specifically, full-stack developers and backend developers are the most frequent. We then constructed a pool of the used technologies provided by the respondents. By computing the list of the ‘most-used’ technologies for a given IT role, we were able to infer the ‘most-used’ IT technical skills for that particular role. According to our findings, JavaScript was the most-used programming language for 16 of the 23 roles surveyed, especially those in the software development job related field. HTML/CSS was reported as the most-used programming language for people working in marketing and as the second-most used programming language for 11 other roles. The most-used programming language for database administrator, data engineer and data analyst was SQL and was cited as one of the three most-used languages for 17 other roles. Further, Python was considered the most-used language for data science professionals, academic researchers and scientists. The study is limited in that the data analysed were limited to one year (i.e. 2020), which limits the extent that this study can be generalised. It would be good to examine the annual trends of IT job roles and skillsets across multiple years to provide broader decision-making insights. The future direction of this study could seek to extend the scope and context by collecting and analysing data from educational settings and scholarly articles. In addition, future work should consider an extensive analysis of trends in the research domain.

804

O. A. Dada et al.

References 1. BurningGlass: Beyond Point and Click - The Expanding Demand for Coding Skills, 1 January 2015. https://www.burning-glass.com/research-project/coding-skills/ 2. Danuser, Y., Kendzia, M.J.: Technological advances and the changing nature of work: deriving a future skills set. Adv. Appl. Sociol. 9(10), 463–477 (2019) 3. Opute, J.E.: HRM in Africa: Understanding New Scenarios and Challenges in an Emerging Economy. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-47128-6 4. Watts, S., Raza, M.: The IT Skill Gap Explained, 7 May 2021. https://www.bmc.com/blogs/ the-it-skill-gap-explained/ 5. Panetta, K.: Gartner Top Strategic Technology Trends for 2021, 8 March 2020. https://www. gartner.com/smarterwithgartner/gartner-top-strategic-technology-trends-for-2021/ 6. Press, G.: Top 10 Hot Artificial Intelligence (AI) Technologies, 18 March 2021. https://www. forbes.com/sites/gilpress/2017/01/23/top-10-hot-artificial-intelligence-ai-technologies/#318 7299d1928 7. Stackoverflow: 2020 Developer Survey, 22 May 2021. https://insights.stackoverflow.com/sur vey/2020 8. Dishman, L.: Why Coding Is Still The Most Important Job Skill of the Future, 14 June 2016. https://www.fastcompany.com/3060883/why-coding-is-the-job-skill-of-thefuture-for-everyone 9. Figuière, N.: Top 10: Most In-Demand Programming Languages 2021, 10 February 2021. https://www.codingame.com/work/blog/hr-news-trends/top-10-in-demand-progra mming-languages/ 10. Codinginfinite: Stack Overflow Developers Survey 2019 vs 2018 Programming Technologies Comparison, 11 April 2019. https://codinginfinite.com/stack-overflow-developers-sur vey-2019-vs-2018-technology-comparison/ 11. Zhang, H., Wang, S., Chen, T.H., Hassan, A.E.: Reading answers on stack overflow: not enough! IEEE Trans. Softw. Eng. 47(11), 2520–2533 (2019) 12. Nivala, M., Seredko, A., Osborne, T., Hillman, T.: StackOverflow–informal learning and the global expansion of professional development and opportunities in programming? In: 2020 IEEE Global Engineering Education Conference (EDUCON), pp. 402–408. IEEE, April 2020 13. Setor, T., Joseph, D.: When agile means staying: the relationship between agile development usage and individual IT professional outcomes. In: Proceedings of the 2019 on Computers and People Research Conference, pp. 168–175, June 2019 14. Vadlamani, S.L., Baysal, O.: Studying software developer expertise and contributions in Stack Overflow and GitHub. In: 2020 IEEE International Conference on Software Maintenance and Evolution (ICSME), pp. 312–323. IEEE, September 2020 15. Rubei, R., Di Sipio, C., Nguyen, P.T., Di Rocco, J., Di Ruscio, D.: PostFinder: Mining Stack Overflow posts to support software developers. Inform. Softw. Technol. 127, 106367 (2020) 16. Venigalla, A.S.M., Chimalakonda, S.: StackEmo: towards enhancing user experience by augmenting stack overflow with Emojis. In: Proceedings of the 29th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering, pp. 1550–1554, August 2021 17. Zhang, Z., Mao, X., Lu, Y., Lu, J., Yu, Y., Li, Z.: Automatic voter recommendation method for closing questions in stack overflow. Int. J. Softw. Eng. Knowl. Eng. 30(11n12), 1707–1733 (2020) 18. Papoutsoglou, M., Mittas, N., Angelis, L.: Mining people analytics from StackOverflow job advertisements. In: Proceedings of the 43rd Euromicro Conference on Software Engineering and Advanced Applications (SEAA 2017), pp. 108–115. IEEE (2017)

Information Technology Roles and Their Most-Used Programming Languages

805

19. Beeharry, Y., Ganoo, M.: Analysis of data from the survey with developers on StackOverflow: A Case Study. ADBU J. Eng. Technol. 7(2), 2348–7305 (2018) 20. Silveira, K.K., Musse, S., Manssour, I.H., Vieira, R., Prikladnicki, R.: Confidence in programming skills: gender insights from StackOverflow developers survey. In: 2019 IEEE/ACM 41st International Conference on Software Engineering: Companion Proceedings (ICSE-Companion), pp. 234–235. IEEE (2019) 21. Peslak, A., Conforti, M.: Computer programming languages in 2020: what we use, who uses them, and how do they impact job satisfaction. Issues Inf. Syst. 21(2), 259–269 (2020) 22. Borges, H., Valente, M.T., Hora, A., Coelho, J.: On the popularity of GitHub applications: a preliminary note. arXiv preprint arXiv:1507.00604 (2015)

Automated Fingerprint Biometric System for Crime Record Management Muyideen AbdulRaheem1 , Sanjay Misra2(B) , Joseph Bamidele Awotunde1 Idowu Dauda Oladipo1 , and Jonathan Oluranti3

,

1 Department of Computer Science, University of Ilorin, Ilorin, Nigeria

{muyideen,awotunde.jb,odidowu}@unilorin.edu.ng

2 Department of Computer Science and Communication, Ostfold University College,

Halden, Norway [email protected] 3 Covenant University, Ota, Ogun State, Nigeria [email protected]

Abstract. Every society has laid down rules and regulations which are to be abide to by the citizens. Once the laws of the land are violated, then a crime is being committed and who break that law is called a criminal. A crime is an illegal conduct that is penalized by the government or another authority. Tracking and managing crimes committed by an individual whose conduct is extremely susceptible to a variety of framing situations is what crime management entails. The crime monitoring system can assist in the storage of records relating to criminals, cases, complaint records, and case histories, among other things. This process is usually done manually and it attracts a lot of issues. Low case tracking capability and a lack of searchable crime databases are among these challenges. In addition, there are issues with paper document management and filing, which can lead to data loss, unwanted access, and damage. Therefore, there is need to automate the system of crime record management. Some researchers have worked in this field but none of them have been able to proffer adequate solution in using fingerprint biometric system to identify criminals based on their unique identifiers. Hence, this study aims at developing an automated fingerprint biometric system for crime record management. The system would be developed using PHP and MYSQL and tested on some datasets and at the end would be able to manage crime records efficiently and effectively. Keywords: Fingerprint · Biometric · Crime record management · Criminal · Citizens

1 Introduction An ideal society is regulated by collectively agreed-upon laws and regulations, as well as measurable repercussions for any member of the community found guilty of breaking any specific law component [1]. We can observe that as a community or society forms, the people develop rules and regulations for and to the people. The goal of these rules © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 806–817, 2022. https://doi.org/10.1007/978-3-030-96299-9_76

Automated Fingerprint Biometric System

807

and laws is to provide a tranquil life for everyone in that society. When someone breaks a law, we call it a crime, and the person who does so is referred to as a criminal [2, 3]. In order for economic, social, and political activity to function correctly, crime is an unwelcomed evil in society that must be completely abolished [4, 5]. A crime is an illegal conduct that is penalized by the government or another authority [6, 7]. A crime, sometimes known as a criminal offense, is an act that causes harm to a community, society, or state in addition to an individual. Such behavior is prohibited and penalized under the law. According to the authors in [8], crime management is the challenge of tracking and managing crimes committed by a small group of persons whose behavior is highly responsive to a range of environmental conditions. The crime management system, among other things, can help with the storing of information on criminals, cases, complaint records, and case histories [9, 10]. The modern world is technology-driven, as it is used in many professions to carry out their tasks [10, 11]. The advancement of technology has opened up new avenues for utilizing the many benefits of information technology in crime reporting. Various technology systems have been developed to aid in the reporting, response, and investigation of crime incidents [12, 13]. Biometric systems are mostly utilized for security reasons. This identity method is mostly used for individual identification, such as scanning a finger, eye, iris, or voice. The biometric system is used to reduce criminal activities and terrorist assaults [14]. This study shows how fingerprints, a biometric technology, can be used to automate a crime record management system. Identity theft, recurrent offenses, authentication frauds, medical malpractices, and unlawful access & permissions are all on the rise these days, and fingerprint-based identification is becoming more useful and cost-effective to combat them [15]. The automated fingerprint is used to identify someone who is suspected of committing a crime [16]. The technology behind Automated Fingerprint Identification Systems is revolutionizing law enforcement’s capacity to catch criminals and investigate crimes. The use of manual mechanisms in law enforcement presents a variety of difficulties. Low case tracking capability and a lack of searchable criminal databases are among the issues. Managing papers and paper filing can sometimes be difficult, resulting in data loss, unwanted access, and destruction. Hence, a considerably number of works have been done in this field to automate the process of crime record management system in order to eliminate the problems of the existing manual system. However, the system is still not able to identify the main culprit as there is no technology put in place to uniquely identify offenders. This can result in punishing an innocent person. This study therefore proposes a fingerprint biometric system to record and manage crimes in order to uniquely identify individuals and be able to track criminals using their unique identifiers. The aim of the study is to develop an Automated Fingerprint Biometric System for Crime Record Management. The contributions of the study include: 1. to design a database to store and retrieve data, and this will manage criminal records. 2. to develop the system for managing crime record with fingerprint biometric system. 3. to test the developed system.

808

M. AbdulRaheem et al.

2 Related Work In the technological advancement bid, computers have made a significant contribution to the simplification of how things are done, as well as the automation of different tasks that occur in and around our daily lives. According to the way technology has advanced and improved, computer and internet technology is gradually replacing some components of tedious paper work that can be implemented using an online automated system [17]. The use of technology is thus viewed as one of the appropriate responses to deal with the threat posed by crime [18]. We can observe that as a community or society forms, the people develop rules and regulations for and to the people. The goal of these rules and laws is to provide a tranquil life for everyone in that society. When someone breaks a law, we call it a crime, and the person who does so is referred to as a criminal. We need a well-organized law enforcement system to live in peace [2]. In any society, anybody that goes against the law of the land is referred to as a criminal. When a crime is committed, such criminal act would be reported to the appropriate authority such as Nigeria Police Force (NPF). As society evolves and human motives and interactions become more diverse, the struggle to regulate crime and the collapse of law and order becomes more important. An ideal society is regulated by mutually decided policies and guidelines, as well as measurable repercussions for any member of the society proven guilty of flouting any specific component of the legal infrastructure [1]. According to authors in [4], crime is an unneeded evil in society, and it must be entirely eradicated from the community for any economic, social, or political activities to run smoothly. A crime is an illegal act that is punishable by the government or another authority. The most frequently accepted notion is that crime is a legal category; that is, something is a crime if it is declared so by the relevant and applicable legislation [19]. When a person commits a crime, the case is documented in a paper document called a criminal record. As long as crimes are committed, it is necessary to maintain track of the cases. A criminal record book is used to keep track of these criminal records. The authors proposed an online criminal record in [20], and the goal of this project was to create a website for criminal records. It is valuable for police to get data about the crooks quicker than expected which diminishes human endeavors. The paper incorporates three stages; In the principal stage, the client picks their undertakings freely online through Local Area Network (LAN). In stage two, the user/police can look, adjust, and update the criminal information at the server data set. The server is prepared to do autonomously executing fundamental customer demands. In stage three, the executive can save the last criminal record and the customer’s outcomes in the information base. In [21], the authors proposed a Real-Time Crime Records Management System for National Police Force (NPF). The exploration was meant to plan and execute an automated ongoing record the executives’ framework for the NPF. The framework was carried out utilizing Hypertext Mark-up Language (HTML) for a profoundly intelligent graphical UI, PHP, and MySQL for a strong data set. However, there is no method put in place to identify individual uniquely for appropriate tracking and managing of criminals.

Automated Fingerprint Biometric System

809

A paper on Integrating Biometrics into Police Information Management System was proposed by [22] which uses Zambia police as case study. A baseline research was conducted to determine the levels of formal education, information and communication technology (ICT) abilities, and ICT tool usage within the Zambia Police Service. Results showed that 47% have moved on from school, 32% finished secondary school and 21% had accomplished an alumni or postgraduate certification. Moreover, 24% had gotten fundamental PC preparing. The overview likewise uncovered that 39% of the respondents utilize their own email for business related interchanges. The concentrate additionally pointed toward building up the significant business processes. Utilizing the business interaction results from the pattern study, a model was created. The model was utilized to foster an electronic model by coordinating unique mark biometrics. The system was developed using Java Server Page (JSP) and MySQL as the database. The development procedure was carried out using the NetBeans Integrated Development Environment. The system is a three-tier application with a client-side user interface, business logic, and a server-side database. The technology, however, was solely used to manage information within the Zambian police force. There is no mechanism for handling criminal records in place. IOT based Online Police First Information system was developed by [14]. The system was aimed to design secured crime record management system, which would have efficient storage, quick retrieval, and exchange of information. The system was implemented using PHP and MYSQL while Visual basic was also used for the IOT. Still, the fingerprint approach was not tackled well to uniquely identify a criminal in order to help police manage criminal records and punish offenders. In their paper [9], the authors presented an online criminal management system. The study is a web-based system for online reporting and automated crime record administration. A person who desires to file a complaint or report an incident on this website must first register before logging in, and once the administrator has authenticated the user, he or she can log in and file a complaint. This protest will be gotten by police and police can communicate something specific with respect to the situation with the grumblings to the client who documented the grievance. The framework was created utilizing PHP and MYSQL. Nonetheless, a unique mark biometric framework was not utilized in overseeing criminal records. However, this study would eliminate the problems of the existing by developing a fingerprint biometric crime record management system which would be secured and makes crime record management easier and storage easy. The system would use fingerprint to uniquely identify criminals based on their unique character.

3 Methodology Having examined the manual process of managing criminal records, the proposed system is intended to simplify how to manage criminal records. The user of the system is the administrator (which can be a police officer in charge). The administrator would be tasked with registering individuals, capture fingerprints, save record and once record is being saved, manage records.

810

M. AbdulRaheem et al.

Fig. 1. Framework for the proposed system

Automated Fingerprint Biometric System

811

Also, the administrator would be tasked with registration of crimes, once crime is being committed, fingerprints would be collected at the crime scene, then the administrator will search the database for a match of the fingerprints collected. If there is a match, such person is the crime suspect, otherwise, no crime suspect. The system is a 3-tier application; the front end will be designed using languages and frameworks such as HTML, Bootstrap, JQuery and JavaScript. The second tier is the server side scripting language to interact with server resources (e.g. databases) using Hypertext Preprocessor (PHP) and the third tier is the Apache Server (XAMPP). The purpose of the system design is to create solution that satisfies the functional requirements for Crime Record Management using Fingerprint Biometric Technique. It is a software system designed with the aim of accepting individual’s fingerprints, and once crime is being committed, compare these fingerprints with the ones collected at the crime scene. If there is a match, then there is a crime suspect, else no crime suspect. These systems are designed with the aim of helping police identify criminal suspects as well as been able to carry out their investigations quickly, promptly and accurately. Figure 1 depicts the framework of the proposed system. The proposed system will authorize the admin (which can be a police officer in charge) to login their details to have access to the system. The system will validate login for the user and once the details are correct, it will load the interface for the software. Admin manages records of crimes and criminals. The admin is also tasked with the registration of peoples’ fingerprints and crimes as well as criminals. After a successful registration of individual fingerprints, these fingerprints will be saved in the database. Once a crime is being committed, fingerprints will be collected from the crime scene and be inputted into the system to check for a match. The database will be search for a match of the fingerprints, if there is a match, then there is a crime suspect and further investigation can take place. Otherwise, no crime suspect. The use case diagram for the administrator is presented in Fig. 2 which shows the activities that are required of the admin such as registering individuals’ fingerprints, registration of crimes, managing criminal records, searching through database for a match in identifying criminals. ERD provides developers with an overall grasp of the data requirements, modeling and database structures of the information system before the implementation phase. It contains three key elements: entities, attributes and relationships. Entities (often mapped into database tables) represent real world objects, relationships capture how two or more entities relate to one another, while attributes are properties of the entities. Figure 3 show the proposed database design of the system.

812

M. AbdulRaheem et al.

Fig. 2. Use case diagram for the proposed system

Fig. 3. Database design for the proposed system

Automated Fingerprint Biometric System

813

4 Results and Discussion Crime Record Management System using Fingerprints Biometric System is a web application that facilitates remote management of criminal records using fingerprint approach.

Fig. 4. Admin login page

Figure 4 is the admin login page for the crime record management system. The admin is the one in charge of the system. Once the login credentials are correct, the dashboard for the admin will be loaded.

Fig. 5. Administrator dashboard

Figure 5 is the administrator dashboard for the system. From the dashboard, the admin can add new admin, manage crimes, manage records, etc. It also includes the total registered Nigerians, Number of Nigerians with crimes and Nigerians with no crime as well as total crimes. From Fig. 6, the admin can add new administrator to oversee all the things that goes on in the system. The system is mainly controlled by the administrator. Figure 7 shows the list of registered Nigerians. At the point of registering each citizen, the fingerprints will be captured and after a successful registration, the admin can then view the details of registered Nigerians. From this place, the admin can add new record, see details of registered citizens. Admin can also delete records. From Fig. 8, the admin can add new crime for a citizen once the citizen commits crime. The system will require the admin to scan the citizen’s fingerprints, once there is a match in the database, then the admin can record the crime (Fig. 9).

814

M. AbdulRaheem et al.

Fig. 6. Adding new administrator

Fig. 7. Manage records

Fig. 8. Add new crime

Fig. 9. Add new record.

Automated Fingerprint Biometric System

815

At the point of registration, the fingerprint would be captured and there are 4 attempts. The required number of attempts is 4 times. Once the attempts are being exceeded, then the fingerprint has been captured, and the record can then be saved.

Fig. 10. Searching for criminals

In Fig. 10, once the fingerprint has been input into the system, the system will search the database for a match of criminal, once there is a match, then crime has been committed and the admin can then save the record.

Fig. 11. Saved crimes

Figure 11 shows the record of saved crimes. Each individual that have committed crimes, their records along with the details are saved. This paper has gone through series of activities to proffer solutions for managing criminal records. After analyzing the paper’s goal and research direction, a set of objectives were established. All activities performed during this paper were attempts to realize these objectives which is to examine how criminal records are been saved in Nigerian Police and elicit requirements from police officers in order to develop a requirement specification and to determine the most suitable software development methodology and technologies that will be used to develop the system with justifications. To sum up, the system’s objectives were to design, develop and test an online crime record management system using fingerprint biometric technique which comprises of a web application to address the problems faced by police officers in manual record management process and to proffer solutions to the problems.

816

M. AbdulRaheem et al.

5 Conclusions The aim of this study is to develop an online crime record management system to address and eliminate the problems of manual system. This paper successfully achieved its aim and objectives though there is still room for improvement. The application has been fully tested throughout the paper phases and it demonstrated acceptable performance. The has documented all the system’s relevant research details. If future extensions to the system are undertaken, this report will be helpful in assisting the completion of other requirements and future improvement that might be involved.

References 1. Nnadimma, C.E.: Design and implementation of an online crime reporting system. Dissertation, School of Postgraduate Studies, University of Lagos (2018) 2. Jamal, F., Barrow, M., Alam, M.J., Mustafa, N.: Unique model of criminal record management system in the perspective of Somalia. JOIV Int. J. Inform. Vis. 3, 332–336 (2019) 3. Macleod, R.: Crime and criminals in the north-west territories 1873–1905. In: The NorthWest Mounted Police and Law Enforcement, 1873–1905, pp. 114–130. University of Toronto Press (2019) 4. Emmanuel, A.: School of computing and informatics a business intelligence system to support crime management in law enforcement agencies: a case of Uganda police force by supervisor dr. elisha t. o. opiyo a research paper report submitted in partial fulfillment for the requirements of the award of degree of master of science in computational intelligence, School of Computing and Informatics, University of Nairobi (2017) 5. Turner, B.S.: Crimes against humanity. In: Vulnerability and Human Rights, pp. 1–24. Penn State University Press (2021) 6. Ganiron Jr, T.U., Chen, J.S., Dela Cruz, R.: Development of an online crime management & reporting system, June 2019 7. Ristroph, A.: The thin blue line from crime to punishment. J. Crim. Law Criminol. 108(2), 305–334 (2018) 8. Alrwisan, A., Ross, J., Williams, D.: Medication incidents reported to an online incident reporting system. Eur. J. Clin. Pharmacol. 67(5), 527–532 (2011). https://doi.org/10.1007/ s00228-010-0986-z 9. AbdulRaheem, M., et al.: An enhanced lightweight speck system for cloud-based smart healthcare. In: Communications in Computer and Information Science, 2021, 1455 CCIS, pp. 363–376 (2021) 10. Adesola, F., Misra, S., Omoregbe, N., Damasevicius, R., Maskeliunas, R.: An IOT-based architecture for crime management in Nigeria. In: Shukla, R.K., Agrawal, J., Sharma, S., Singh Tomer, G. (eds.) Data, Engineering and Applications, pp. 245–254. Springer, Singapore (2019). https://doi.org/10.1007/978-981-13-6351-1_19 11. Awotunde, J.B., Adewunmi-Olowabi, F.T., Owolabi, A.A., Akanbi, M.B.: Automated global system for mobile-based vehicle inspection using short-code: case study of Nigeria. Comput. Inf. Syst. Dev. Inform. Allied Res. J. 5(3), 45–50 (2014) 12. Afah, D., Gautam, A., Misra, S., Agrawal, A., Damaševiˇcius, R., Maskeli¯unas, R.: Smartphones verification and identification by the use of fingerprint. In: Mandal, J.K., De, D. (eds.) Advanced Techniques for IoT Applications. EAIT 2020. LNNS, vol. 292, pp. 365–373. Springer, Singapore (2022). https://doi.org/10.1007/978-981-16-4435-1_35

Automated Fingerprint Biometric System

817

13. Awotunde, J.B., Fatai, O.W., Akanbi, M.B., Abdulkadir, S.I., Idepefo, O.F.: A hybrid fingerprint identification system for immigration control using the Minutiae and correlation methods. J. Comput. Sci. Appl. 21(2), 97–108 (2014) 14. Bhosale, M.R., Paradeshi, K.P.: Iot Based Online Police First Information Report (FIR) Record System. 107–112 (2019) 15. Usmani, Z., Irum, S., Mahmud, S.: How to build an automated fingerprint identification system (2013). https://doi.org/10.1109/ISBAST.2013.9 16. Belmon, L., Kozik, R., Demestichas, K.: Why do law enforcement agencies need AI for analyzing big data?. In: Proceeding of the 20th International Conference, Computer Information Systems and Industrial Management, CISIM 2021, Ełk, Poland, 24–26 September 2021, vol. 12883, p. 331. Springer Nature (2021) 17. Gerstlauer, A., Haubelt, C., Pimentel, A.D., Stefanov, T.P., Gajski, D.D., Teich, J.: Electronic system-level synthesis methodologies. IEEE Trans. Comput.-Aided Des. Integr. Circuits Syst. 28(10), 1517–1530 (2009) 18. Richardson, R., Director, C.S.I.: CSI computer crime and security survey. Comput. Secur. Inst. 1, 1–30 (2008) 19. Ganiron, Jr., T.U., Chen, J.S., Dela Cruz, R., Pelacio, J.G.: Development of an online crime management & reporting system. 131(june), 164–180 (2019) 20. Nawale, S.D., Songra, M.P.C., Karnik, R.: Online criminal record. 12(8) (2012) 21. Awodele, O., Onuiri Ernest, E., Olaore Olufunmike, A., Anita, S.O.O.U.E.: A real-time crime records management system for national security agencies. European J. Comput. Sci. Infor. Tech. 3(2), 1–12 (2015) 22. Lyoko, G., Phiri, J., Phiri, A.: Integrating biometrics into police information management system: a case integrating biometrics into police information management system: a case of Zambia police (2016). https://doi.org/10.18178/ijfcc.2016.5.1.433

Estimation Techniques for Scrum: A Qualitative Systematic Study Diaz Jorge-Martinez1 , Sanjay Misra2(B) , Shariq Aziz Butt3 , and Foluso Ayeni4,5 1 Department of Computer Science and Electronics, Universidad de la Costa,

Barranquilla, Colombia [email protected] 2 Department of Computer Science and Communication, Ostfold University College, Halden, Norway [email protected] 3 University of Lahore, Lahore, Pakistan 4 Information Systems and Quantitative Analysis Department, University of Nebraska Omaha, Omaha, USA 5 Center of ICT/ICE, Covenant University, Ota, Nigeria

Abstract. Every competitive IT industry cannot avoid underestimating their projects’ effort, cost, and time. Some scrum project is completed delayed and undergoes difficulties due to over budgeting and a lack of needed functions. Software project failures are caused by incorrect and imprecise estimation; thus, it should be taken into account. A substantial change is required when Agile-based processes (e.g., Scrum) are introduced to the industry. The analysis is still difficult with Agile since requirements are constantly changing. Projects, individuals, and resistance issues, incorrect usage of cost factors, unawareness of regression testing work, readability of software requirements size as well as its related complexities, and so forth are all causes behind the difference in anticipated and real effort. This work analysis examined several publications and prospective researchers striving to narrow the actual and estimated effort gap. Decision-Based techniques significantly outperformed non-Decision Based and conventional estimating strategies by extensive literature analysis. We found that the regression test based estimation technique should be improved for accurate estimation of effort. However, scrum still needs a significant estimation technique to resolve the over budgeting issue. This study discussed the machine learning techniques, there proficiencies for estimation and flaws. The overall effort is the sum of all sprints components’ efforts, and it repeats after the prospective deliverable version. Keywords: Agile methodology · Software development · Cost estimation techniques

1 Introduction Agile estimating is still difficult for IT professionals all over the world, and numerous studies have previously addressed this subject in their writings. Depicts a common estimating architecture used by most Software industries, in which needs, or requested © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 818–829, 2022. https://doi.org/10.1007/978-3-030-96299-9_77

Estimation Techniques for Scrum

819

user stories, are layered in the product backlog and then marked by their lengths. The most common unit for sizing a product backlog is the story point used by industries. According to ISPA, two-thirds of software development projects fail to deliver within time and under budget. There are two main causes of software project failures: one is an inaccurate prediction of requirement size, budget, and human resources, and the other is the unpredictability of system and software specifications. Modification and sprint-wise estimate are two of the most difficult aspects of estimating Scrum-based projects. The majority of Software companies have embraced hybrid performance measures that are mostly guided with agile umbrella methods. According to [2], the transition of methodologies from heavyweight, such as iterative waterfall, to lightweight, such as Agile, has resulted in a change in effort estimation methods. All traditional estimation techniques [3], such as expert judgment, top-down estimation, and Delphi cost estimates, are well suitable for heavier process modeling in some form or another; however, they ultimately lack the ability to connect the expected and actual. As a result of the ephemeral nature of Agile-based project specifications, academics began looking for alternatives, ultimately settled on soft computing techniques [4]. Providing guidelines for unforeseen difficulties is one of the most well-known applications of neuro-fuzzy frameworks [5]. On the other hand, software developments are generally uncertain and complicated. The result is that the available data is insufficient at the start of the development, and the problem of effort estimation is totally unknown. In this case, fuzzy and neuro-fuzzy techniques can address the problem while increasing estimation precision. Additionally, significant results have been attributed to fuzzy-based techniques in the domain of software cost estimation. The optimization technique plays a crucial role now due to the inconsistency of the effort estimate difficulty and the complexity with project and team attribute relationship evaluation. The analysis [6] can be directly linked to effort estimation techniques, including a quality rating in an analogy-based estimate or indirectly to Decision-Based algorithms. The structure of the paper is as follows. A survey modeling adopted for this study is provided in Sect. 2. Section 3 contains the experimental design and methodology, and Sect. 4 presents the statistical results applied to the collected data set and reveals the findings of the study.

2 Agile Software Development Agile software development resolves these issues and removes these obstacles from efficient software development. It facilitates the organizations to develop the software very fast and according to clients’ expectations and needs [7, 8]. Agile software development works in sprints (sprints) [9]. The agile model divides the project into a tiny scale of sprints. It supports the software development team to manage the project, get quick feedback from the client and develop the project swiftly. As agile supports change in requirements even in later stages, the team can also easily manage the change request from the client-side. Issues with Agile Software Development: There are a lot of issues with agile software development. Such as frequent change requests, cost increment, time exceeds from the

820

D. Jorge-Martinez et al.

deadline, and development team coordination about the project. These all issues are explained below one by one.

2.1 Frequent Change Request Issue The agile model breaks the project into small sub-programs and starts development accordingly until the completion of the whole project. The client’s change of request has the main effect on the project like cost, determination, time, and some upgradation resources. The change in one sub-program of the project affects the other sub-program of the project. The change in a specific sub-program just not only limited to that sub-program but also circulates to all other sub-programs in the software development [10]. All subprograms are joined with each other in this model. The successive difference in ask for influences one to more sub-program (s) ineffectively. The aggregate no of sub-programs is five, and every one of them requires some opportunity to finish, battle, and cost-related with it. These two sub-programs out of five, sub-programs 3 and 5, have changed ask for from the customer side. These two sub-programs take more struggle, cost, and time to complete. Finally, agile software development’s basic purpose is to frequent request change from the client-side and accomplish it. The client’s frequent change of request can disturb the entire project. The client’s change of recommendation can be developed on a priority basis in agile software development. The developer transforms the client’s priority-based requirements and starts application development accordingly. Due to these priority-based requirements, project complexity may increase because it might possible clients require complex changes during the application development. It may have major effects on the other parts of the projects [11, 12]. Hence the key feature of the ASD permits the client’s change in requirements in even later stages of a project, which is the biggest drawback of the agile model and for application development using an agile model in the software industry. 2.2 Cost Increase in ASD Issue Estimation is the procedure to assess the cost, exertion, and time taken for a product venture. The procedure of estimation begins from the arranging period of the SDLC and this is produced all through every one of the periods of the SDLC [13]. The product advancement has principle two parts: a season of the closing purpose of the project and the cost of the project. The product cost is an exceptionally critical phase of the product. To appraise the correct cost of the product and its all assignments are extremely intense [14, 15]. It very well may be the reason for an awful relationship of the organization with a customer because the customer has really an incentive in this Developing stage [14]. The primary reason for the expansion of the venture cost is the change asked for originating from the customer side. Because of progress asked of from the customer side, the cost of the product outperforms the pre-decided cost of the product. The ASD enables the customer to give change ask for whenever and at the level of the product. Because of these successive changes of demand from the customer side, increment the cost of the product. The ASD has separated the venture into different runs, and on these runs, the change asks for originates from the customer side. Multi change demands at a

Estimation Techniques for Scrum

821

single run can be originated from the customer side at different times. Because of this different time change, ask for the cost of that specific run at one run and different dashes in the venture rise. So, there is a need to estimation the correct cost of the product. There is an additional need to compute the cost of each sprint/iteration. ASD needs a cost estimation system or model to examine the correct cost of the undertaking with the goal that at whatever point and not makes any difference how much the successive progress of demand originate from the customer side they chose the cost of the product must not outperform [15]. 2.3 Time Increase in ASD Issue The estimation time is used to evaluate the completion time of the project. At the beginning of the project, the completion time of the project is decided. While software development life cycle first plans phase takes place, we start developing according to that plan [16]. The completion time of the project plays a vital role in software development. It is very difficult to evaluate the project completion time [17]. If the project’s completion time is estimated wrong, then it can become a failure of that project. The company’s goodwill may decline due to this aspect. The client is a greater asset for the company because he is investing his time and money for his project and the company does not want to spoil its relationship with the client. The client’s change of request is the main reason for the completion of the whole project. Because the change request for the time outperforms the time chosen toward the beginning of the product. Not fulfilling the undertaking on the given focus on time is a major issue in ASD. The deft permits the customer for the continuous difference in request. As coordinated partitions the undertakings into little subprograms. These subprograms are the little availabilities. Each subprogram sets aside some opportunity to finish. When the continuous difference in a request originated from the customer side at a specific subprogram, the time surpasses to finish that subprogram and fulfill the customer, toward the end, it influences the other sub program’s chance in the project the entire undertaking convey late. So, the agile software development needs training or model to appraise the privilege and correct time of the finishing of the undertaking with the goal that at whatever point and not makes any difference what amount visit change of demand originated from the customer side the time won’t outperform shape the focused-on time toward the start of the product [18].

3 Related Literature In a research published in [7], more than twelve estimation approaches for effort estimation were identified in the 1980s, with regression-based techniques exceeding the empirical estimation method. Despite a large number of detailed studies on Decision Based-based techniques in the estimate of software development projects, regardless of the procedure prediction method, conflicting results have been explained in terms of the estimating accuracy of such techniques. For example, when a comparable Decision Based model is constructed with different facts [19] or contexts, the estimation’s accuracy becomes varies. In terms of the relationship between the Decision Based model

822

D. Jorge-Martinez et al.

and the regression model, considers in [13] declared that the Decision Based method is preferable to the regression method, whereas considers in [15] concluded that the regression method outperforms the Decision Based method. When performed to various datasets, artificial neural networks and case-based reasoning strategies beat each other in and [18]. The variance in existing empirical evaluations of Decision-Based methods has still not been fully understood, and it may discourage experts from adopting Decision Based methods by comparing them to other domains where systems are being interconnected effectively. Furthermore, Decision Based methods’ theories are far more complicated than typical estimating processes. It is critical to consciously summarize the empirical proof on Decision Based models in the current study and experimentation to boost the adoption of Decision Based processes in the Software effort estimation field. Expert judgment and Delphi cost estimating methodologies are used by software industry specialists more than Decision Based. Case-based justification, Artificial neural networks, Support vector regression, and Function point analysis are some of the Decision Based methodologies which are understanding the issues for Software effort estimating; however, the majority of techniques are not yet being used for the agile estimate. The Decision Based techniques mentioned above can be used alone or in combination with other Decision Based or non-machine learning techniques. For example, genetic algorithm techniques have been used with Case-based thinking, artificial neural networks, and Support vector machine to emphasize weight and selection. The Bayesian network is displayed to have the worst performance of all the Decision-Based techniques. In contrary to Case-based logic (52%), Deep neural networks (35%), Decision trees (57%), Support vector machines (36%), and Genetic algorithm, the mean size of proportional improvement system for predicting projects contain both conventional and lightweight techniques, although not in all cases [20]. Although studies show that deep neural networks and support vector machines [21] outperformed other predictive techniques, this does not mean that we should use them without restriction. Increasing the size of hidden levels will increase preparation time and may cause over-fitting problems. Decision-Based approaches are investigated using regression models, COCOMO estimation, Expert judgment, and Function point analysis [22]. Case-based logic and deep neural networks are more accurate than regression models in research. Regression is more exact than genetic programming. Decision-Based approaches are investigated using regression models, COCOMO estimation, Expert judgment, and Function point analysis. Case-based logic and deep neural networks are more accurate than regression models in research. Regression is more exact than genetic programming. So, based on the data, we conclude that Decision Based is a good thing. In writing, distinct estimating parameters are established concerning the limited informational gathering, anomalies, absolute highlights, and missing attributes. Analysts indicate that determining the optimum estimate in a certain environment is more efficient than selecting the finest single model because estimate techniques differ from one input data to the next, making them susceptible. According to research on information gathering, collective strategies produce identical results when compared to single strategies because each approach has excellence and drawbacks, and merging them would reduce the deficiencies. Homogeneous (e.g., Case-based logic, deep neural networks, Decision trees, Support vector machines, and so on) and heterogeneous effort

Estimation Techniques for Scrum

823

estimation methods are distinguished by their base modeling techniques and a mixture of guidelines. The deep neural network is used most. According to studies, solitary processes are the most common way for estimation. It is revealed that homogeneous using Decision trees are the strongest efficient, followed by homogeneous using Case-based reasoning, and then Support vector machines homogeneous development.

4 Cost and Time/Effort Estimation Techniques There is a lot of cost and time estimation techniques to help the agile software development managerial head to find the better and best cost and time estimation of the software projects. 4.1 Algorithmic and Non-algorithmic Methods Cost estimation models are categorized into two types; algorithmic and non-algorithmic. The algorithmic models are repeatable and use formulas and equations to calculate the cost; it may be further divided into linear, non-linear, or quadratic categories, while the non-algorithmic cost modeling often depends on analysis heuristics and sometimes demand knowledge to find the cost of the software. 4.1.1 Non-algorithmic Methods Non-Algorithmic Methods are contrary to the Algorithmic methods; methods that belong to this group depend on the Analytical comparisons and the inferences based on them. Some information about the previous projects is required for using the Non-Algorithmic methods, which are quite similar to the underestimate project, and usually, estimation deduction in these methods is made according to the analysis of the previous datasets. Three methods have been selected here for the assessment because these methods are better liked than any other Non-Algorithmic methods, and much publishable and applicable research has already been done on them [10, 12]. 4.1.2 Algorithmic Methods In Algorithmic Models, costs usually are analyzed using mathematically deduced formulae whose input are the metrics that produce the estimation cost at the output. The formulae which are used in a formal model emerge from the analysis of historical data. The precision and, most importantly, the validity of the model can be improved by tuning the model to your specific development conditions and environment, which usually involves calibrating the parameters and weightings of the metrics [17].

5 Method Research selection methods have all been covered in this part. Research Questions The objective of this study is to explain the current state of Decision-Based techniques’ application in Scrum-based developments. In this respect, the following research questions are being formulated as follows:

824

D. Jorge-Martinez et al.

RQ1: Which Decision Based techniques are employed to estimate Scrum? RQ2: Do Decision Based models outperform other Decision Based modeling techniques when it comes to Scrum estimation? RQ3: How accurate are Decision Based techniques employed in Scrum-based projects in terms of total estimate correctness? RQ4: Does employing Meta-heuristic algorithms improve the estimating effectiveness of Scrum-based developments? RQ5: In Scrum projects, are ensemble estimating techniques stronger than the single estimate? RQ6: What are the numerous important aspects that influence the Scrum project effort? 5.1 Criteria to Inclusion and Exclusion This research includes publications that link several soft computing strategies for estimating in agile development. Papers have been compiled from several different sources, journals, conferences, and other sources that have been distributed to date. A few studies that aren’t clear about ASD are still included as they include important information. The assessment excludes papers and data that are not relevant to the study area. 5.2 Data and Literature Sources Description Publications from TOSEM (ACM), IEEE Transactions, Science Direct, Google Scholar, Springer, and other sources are included in the research. Some search strings are adopted for the searching of publications: “Scrum” OR “Agile” AND “software” AND “development” AND (effort OR cost) AND (estimation) AND (“Decision Based”) OR “casebased reasoning” OR “decision tree” OR “prediction based” OR “neural network” OR “Bayesian technique” OR “support vector machine” OR “support vector regression” OR “deep” OR “learning” OR “expert judgment” OR “neuro-fuzzy” OR “ANFIS” OR “meta-heuristic” OR OR “genetic algorithms” OR “planning poker. 5.3 Methodology of Study Selection Following the application of include and exclude criteria, the selection process is divided into two steps: • Selecting an abstract and title: The review technique is carried out using several research articles, many of which are selected based on their titles and improved outputs. • Selecting the entire article: A substantial number of publications and articles have been evaluated and extensively analyzed, as mentioned in Sect. 4 questions of the study.

Estimation Techniques for Scrum

825

6 Results and Discussion In this section, we are discussing our findings of the study: RQ1: Which Decision Based techniques are employed to estimate Scrum? Agile development and its accompanying practices have made substantial use of a wide range of Decision Based techniques. Table 1 shows the frequency and year of publication of Decision Based techniques applied within the Scrum estimate. Table 1. Decision based techniques for estimate. Decision based techniques

References

Optimized neural network

[1, 4, 5, 7]

Neural networks with general prediction

[7–9, 20]

Decision tree

[7–9, 20]

Case-based justification

[11, 12, 19, 20]

Support vector regression

[6, 9, 11, 21]

Function point analysis

[17–19, 22, 23]

Expert opinion

[9, 12, 13, 20]

Fuzzy based

[21–24]

Mean magnitude of relative error

[6, 13, 17]

From the table above, it can be observed that the majority of the contributors used various Decision Based techniques, and based on the year of publication, a pattern can be drawn that studies over the years using Decision Based techniques that develop an auto-estimate system. A comprehensive study is conducted out in the next section [20]. RQ2: Do Decision Based models outperform other Decision Based modeling techniques when it comes to Scrum estimation? The fact that AI techniques outperform non-AI techniques is stated in Sect. 2. Human bias is also present in expert-based judgments. Table 3 provides a comparative study of all the AI approaches used for Scrum-based project prediction in this study objectives. Multiple measures like Magnitude of relative error and Percentage relative error deviation are being presented as accuracy parameters based on the data available in the studies. When applied to a similar dataset or other datasets, certain AI techniques outperformed other AI techniques. Table 3 shows that, based on the accuracy parameter, the strongest existing AI technique is The fireworks algorithm optimal neural network has a comparative accuracy of 2.74% for 22 project data. It’s impossible to say with certainty because the projects/datasets used by different contributors vary and will have less or more predictability. Many researchers with different datasets have also reported improved prediction. RQ3: How accurate are Decision Based techniques employed in Scrum-based projects in terms of total estimate correctness?

826

D. Jorge-Martinez et al.

Several AI algorithms are being applied to Scrum-based project estimate to the greatest of the information. Table 2 shows that the average mean Magnitude of relative error for AI approaches on the identical data is 0.2822. RQ4: Does employing Metaheuristic algorithms improve the estimating effectiveness of Scrum-based developments? In the research, there is relatively insufficient actual data in support of using metaheuristic methods for agile software development projects. As shown in Table 2, only two such articles, namely the fireworks algorithm, have been peer-reviewed. In comparison to all AI techniques being used in Scrum, the fireworks algorithm provides high predictive performance. In this perspective, we can infer that including meta-heuristic techniques improve estimation. Table 2. Comparative accuracies of different ML estimation techniques Decision based techniques

References

Accuracy factors

Outperformed

Optimized neural network

[1, 4, 5, 7]

MMRE—0.0174

optimization depending on learning

Neural networks with general prediction

[7–9, 20]

MMRE—0.2

Actual estimations are compared

Decision tree

[7–9, 20]

MMRE—0.2710

Association rule

Case-based justification

[11, 12, 19, 20]

MMRE—0.1375

Association rule

Support vector regression

[6, 9, 11, 21]

MMRE—0.41

Logical method

Function point analysis

[17–19, 22, 23]

MMRE—0.1381

Decision-based on category

Expert opinion

[9, 12, 13, 20]

MMRE—0.1560

Delphi and planning poker

Fuzzy based

[21–24]

MMRE—0.1610

Empirical estimation technique like educated guess

Mean Magnitude of relative error

[6, 13, 17]

MMRE—0.1701

Neural network with a high probability

RQ5: In Scrum projects, are ensemble estimating techniques stronger than the single estimate? Yes, the research concludes that aggregate estimation techniques produce stronger outcomes than a single estimating method. Similarly, the community wins by a wide margin when we look back at the research for estimate strategies for weighty process equations. RQ6: What are the numerous important aspects that influence the Scrum project effort?

Estimation Techniques for Scrum

827

Different people, attitudes, and project variables have all had an impact on the Scrum project effort. In this regard, several authors have identified different factors, which are shown in Table 3. Table 3. Factors influencing the effort of Scrum-based projects Factors affecting the project

Factors relating to people

Factors of resistance

The realm of the project

Communicating capacities

Errors in third-party technologies and a suitable team configuration

Requirement for success

In the team, there is a lot of familiarities

Uncomfortable working conditions and stakeholder reactions

Hardware and software specifications are required

Management abilities

Drifting to Agile, a shortage of consistency within requirements, and the volatility of needs are all factors to consider

Operational simplicity

Safety

Team dynamics and changes in the workplace

Complexity

Working hours

Changes in the team that is expected, as well as other project duties

A data exchange

Experience in previous projects

Introduction to new technologies and the availability of resources are both required

Multiple sites

Technical aptitude

Usability

7 Conclusion The major study deficiencies have been found in this systematic review, which provides the potential for all future researchers worldwide. Because various accelerating and decelerating elements can impact the estimation of Agile-based projects, neglecting estimating components may lead to poor estimates. The overall effort is the sum of all sprints components’ efforts, and it repeats after the prospective deliverable version; therefore, regression testing effort should be improved to give it more accuracy. There is a multitude of Decision Based and optimization technologies that have still to be used to estimate project work in Scrum-based projects. In the literature, there is no suitable and adequate scale for user story size and complexity. Scrum estimate does not have a generic or single estimate technique.

828

D. Jorge-Martinez et al.

References 1. Steghöfer, J.P., Knauss, E., Alégroth, E., Hammouda, I., Burden, H., Ericsson, M.: Teaching agile-addressing the conflict between project delivery and application of agile methods. In: 2016 IEEE/ACM 38th International Conference on Software Engineering Companion (ICSEC), pp. 303–312. IEEE, May 2016 2. Meyer, B.: Making sense of agile methods. IEEE Softw. 35(2), 91–94 (2018) 3. Martin, A., Anslow, C., Johnson, D.: Teaching agile methods to software engineering professionals: 10 years, 1000 release plans. In: Baumeister, H., Lichter, H., Riebisch, M. (eds.) Agile Processes in Software Engineering and Extreme Programming. XP 2017. LNBIP, vol. 283, pp. 151–166. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-57633-6_10 4. Butt, S.A.: Study of agile methodology with the cloud. Pac. Sci. Rev. B Humanit. Soc. Sci. 2(1), 22–28 (2016) 5. Fuchs, C.: Adapting (to) agile methods: exploring the interplay of agile methods and organizational features (2019) 6. Przybyłek, A., Kotecka, D.: Making agile retrospectives more awesome. In: 2017 Federated Conference on Computer Science and Information Systems (FedCSIS), pp. 1211–1216. IEEE, September 2017 7. Tessem, B.: The customer effect in agile system development projects. A process tracing case study. Procedia Comput. Sci. 121, 244–251 (2017) 8. Butt, S.A., Abbas, S.A., Ahsan, M.: Software development life cycle & software quality measuring types. Asian J. Math. Comput. Res. 11(2), 112–122 (2016) 9. Kim, S.I., Lee, J.Y.: Walk-Through screening center for COVID-19: an accessible and efficient screening system in a pandemic situation. J. Korean Med. Sci. 35(15), e154 (2020) 10. Janssen, M., van der Voort, H.: Agile and adaptive governance in crisis response: lessons from the COVID-19 pandemic. Int. J. Inf. Manag. 55, 102180 (2020) 11. Asare, A.O., Addo, P.C., Sarpong, E.O., Kotei, D.: COVID-19: optimizing business performance through agile business intelligence and data analytics. Open J. Bus. Manag. 8(5), 2071–2080 (2020) 12. Mishra, A., Misra, S.: People management in the software industry: the key to success. ACM SIGSOFT Softw. Eng. Notes 35(6), 1–4 (2010) 13. Fernández-Sanz, L., Gómez-Pérez, J., Diez-Folledo, T.I., Misra, S.: Researching human and organizational factors impact for decisions on software quality. In: Proceedings of the11th International Conference on Software Engineering and Applications, pp. 283–289 (2016) 14. Fernández-Sanz, L., Misra, S.: Influence of human factors in software quality and productivity. In: Murgante, B., Gervasi, O., Iglesias, A., Taniar, D., Apduhan, B.O. (eds.) Computational Science and Its Applications - ICCSA 2011. LNCS, vol. 6786, pp. 257–269. Springer, Heidelberg (2011). https://doi.org/10.1007/978-3-642-21934-4_22 15. Butt, S.A., Misra, S., Anjum, M.W., Hassan, S.A.: Agile project development issues during COVID-19. In: Przybyłek, A., Miler, J., Poth, A., Riel, A. (eds.) Lean and Agile Software Development. LASD 2021. LNBIP, vol. 408, pp. 59–70. Springer, Cham (2021). https://doi. org/10.1007/978-3-030-67084-9_4 16. Butt, S.A.: Analysis of unfair means cases in computer-based examination systems. Pac. Sci. Rev. B Humanit. Soc. Sci. 2(2), 75–79 (2016) 17. Przybyłek, A., Zakrzewski, M.: Adopting collaborative games into agile requirements engineering (2018) 18. Al Asheeri, M.M., Hammad, M.: Machine learning models for software cost estimation. In: 2019 International Conference on Innovation and Intelligence for Informatics, Computing, and Technologies (3ICT), pp. 1–6. IEEE, September 2019

Estimation Techniques for Scrum

829

19. Rao, C.P., Siva Kumar, P., Rama Sree, S., Devi, J.: An agile effort estimation based on story points using machine learning techniques. In: Bhateja, V., Tavares, J., Rani, B., Prasad, V., Raju, K. (eds.) Proceedings of the Second International Conference on Computational Intelligence and Informatics. AISC, vol. 712, pp. 209–219. Springer, Singapore (2018). https:// doi.org/10.1007/978-981-10-8228-3_20 20. Periyasamy, K., Chianelli, J.: A project tracking tool for scrum projects with machine learning support for cost estimation. In: EPiC Series in Computing, vol. 76, pp. 86–94 (2021) 21. Adnan, M., Afzal, M.: Ontology based multiagent effort estimation system for scrum agile method. IEEE Access 5, 25993–26005 (2017) 22. Kokol, P., Zagoranski, S., Kokol, M.: Software development with scrum: a bibliometric analysis and profile fi (2020) 23. Sharma, A., Chaudhary, N.: Linear regression model for agile software development effort estimation. In: 2020 5th IEEE International Conference on Recent Advances and Innovations in Engineering (ICRAIE), pp. 1–4. IEEE, December 2020 24. Syahputri, I.W., Ferdiana, R., Kusumawardani, S.S.: Does system based on decision based need software engineering method? Systematic review. In: 2020 Fifth International Conference on Informatics and Computing (ICIC), pp. 1–6. IEEE, November 2020

Development of Students’ Results Help Desk System for First Tier Tertiary Institutions Abraham Ayegba Alfa1 , Sanjay Misra2(B) , Blessing Iganya Attah3 , Kharimah Bimbola Ahmed4 , Jonathan Oluranti5 , Robertas Damaševiˇcius6 , and Rytis Maskeli¯unas6 1 Confluence University of Science and Technology, Osara, Nigeria

[email protected]

2 Department of Computer Science and Communication, Ostfold University College, Halden,

Norway [email protected] 3 Federal University of Technology, Minna, Nigeria 4 Kogi State College of Education, Ankpa, Nigeria 5 Covenant University, Ota, Ogun State, Nigeria 6 Faculty of Applied Mathematics, Silesian University of Technology, Gliwice, Poland

Abstract. The paper attempts to develop an effective students’ results help desk system for third-tier higher institutions in Nigeria (such as Colleges of Education). This system offers help desk assistance or services, that is, technical and information assistance for students through enrolling, storing, unifying, tracking, and undertaking students’ results issues and challenges. These were previously handled and conducted manually using paper and pen approach. The target of this electronic-based results help system is to assist students to communicate common results computation and publication issues for effective service delivery, students’ satisfaction, and enhanced the reliability of formative and summative assessments records undertaken by the students. The new system was prototyped using PHP, CSS, HTML and MySQL Server Web development platform. The outcomes show that new system enables the educational needs of students to be timely catered for in more effective ways. This system enables results corrections fraud cases due to lack of audit trail and record keeping by the institutions’ staff in old system to be minimized. It encourages the use of ICT to automate and perform operations more effectively and speedily. Keywords: Help desk · Information system · Complaint · Results processing · Students · Management

1 Introduction In everyday lives of people, it is nearly impossible living in the present-day world short of information technologies as means of fostering serious and effective interaction. Through, it needs supplementary support and maintenance. Consequent upon this, the Help-Desk technical assistance or service is apt because it attempts to speed-up user © The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 830–841, 2022. https://doi.org/10.1007/978-3-030-96299-9_78

Development of Students’ Results Help Desk System

831

complaints and requests processing (Rusakov and Loseva [1]). The information and educational environment globally are embracing smart technologies to facilitate communication and offering services to students and faculty members. In effect, students utilize technology for the purpose of sharing and exchanging information with each other and carry out several tasks more quickly and accurately. This significantly assist students to conserve their time, and to make the institutions’ resources available speedily and simply with high access (Alboaneen et al. [2]). The tasks of improving the efficiency of the information and educational environment, involve the introduction of the Help Desk system. It will significantly increase the efficiency of the information and educational environment through the overall improvement of the University’s IT infrastructure. Help Desk is a technical support tool for users. For users, its implementation is characterized by an increase in the level of service provided and a reduction in periods of downtime due to problems (Nikiforov and Mukhametova [3]). Help desk system allows authorized access to information and educational environment services by students, teachers, and employees by enforcing the communication between members of staff in a typical bureaucratic setup, and students from various backgrounds can be performed more efficiently (Chen [4]). Online help desk system comes handy to automate the entire procedure of College activities through effective optimizations and speedups (Asha and Babu [5]). Several institutions of learning are evolving means to assist students through their educational journey with valuable information and services such as chatbot (Dibitonto et al. [6]). Presently, enterprises are faced with numerous difficulties in bids to resolve complaints posed by their customers. In fact, part of this challenge can be blamed on inability of enterprises to maintain appropriate complaints management solution or system. This is done manually in majority of organizations in which a customer request for service improvements or lays complaints concerning specific areas to the relevant party. The outcomes have led to further dissatisfaction on the part of the customer (Munasinghe [7]). Similarly, students of higher institutions are faced with diverse complaints relating to results publications and other educationally related issues. Help-desk services in public institutions as currently obtains are flooded with uncountable number of complaints exceeding their capability to undertake (Zaza, Junglas and Armstrong [8]). The process of complaint escalation is inappropriate. Problems have been ensued including: larger communication gaps, and longer periods of closing complaint cases. There is need to minimize discrepancy and raise the level of efficiency in process to serve customers better while enhancing their competitiveness (Justitia, Zaman, and Putra [9]). This paper designs and implements a computer-based students’ results help desk system for thirdtier tertiary institutions in Nigeria. The remaining portions of this paper are sectioned as follows: literature review is in second section; methodology is in third section; system implementation and results are in fourth section; conclusion is in the fifth section.

832

A. A. Alfa et al.

2 Literature Review Concept of Help Desk and Complaint Management The concept of service is largely variable whose quality is judged on the providers at a particular time and place (Nwokorie [10]). In practice, the assurance of quality service allows its providers to detect faults or issues and initiate resolution; thereby eliminating as much as possible service down time, and identifying possible degradation of quality of service. There are seemingly low incidences of prospective complainants owing to the lack of confidentiality of data generated from the complaints of the customers. Aside this, there are no mechanisms in place for speedily filing complaints about service failure from the point of origination. The quest for effective customer service delivery is borne out of previous happenings as related to failure of service in most organizations. The failure in service play a key role in unproductivity cases prevalent for the government or private institutions. Recently, establishments, institutions and organizations around the countries have put in place different measures to checkmate failure in service; though, majority of them are unsuccessful. The quality service assurance criteria largely rely on the model selected for operation. In fact, certain institutions affirm solely on providing the appropriate training at the various service point representatives, but, other hold the notion that, the optimal tools can be provided in rendering quality service (Omaka, Onwudebelu, and Okemiri [11]). Help desk or counter can be akin to a piece of furniture that as the entry/exit place in administrative, commercial or cultural affiliations whether in private or public spheres for the purpose of gaining access physically. It depicts the physical barrier where providers of service and consumers of service in organizations (such as tourism) relate and exchange information concerning restaurants, hotels, tourist offices, and others. Consequently, new technologies are capable of offering fresh ways of exchanges on means of delivering quality services, which enhances experience of customer or user experience (UX) (Rogers, Sharp and Preece [12]). The most conspicuous attributes of the available error complaints and tracking systems are itemized as follows (Munasinghe [7]): Reporting facility: User provides relevant information about bugs, severity, model, screenshots through appropriate records filling style. Assigning: Users can be allotted specific issues. History/comments/work logs. Reports: Charts and graphical representation can be adopted. Technology in Educational Services The Office of the Presidency and the Cabinet of Nigeria in 2004 signed the social compact for purpose of offering the essential services to Nigerians in a fair, timely, honest, transparent and effective manner (Omaka et al. [11]). The attempts to implement this customer loyalty compact led to the development of the customer complaint model for the assurance of quality service. It has been adjudged as the best performing and productive; because it puts the power of service rating in the hands of the customer after experience. The primary attention is on government higher institution of learning

Development of Students’ Results Help Desk System

833

in Nigeria and their services quality. Though, failure in service is mostly caused by sophistication of technologies utilized in the institutions. To speed-up the rate of the social compact enactment, mobile technology was involved into this system, for easier and quicker complaint collection, processing and improve process of making decisions. The SERVICOM administrative component ensures better quality services are assured to various customers. The computer-based application was incorporated to enables dissatisfied customers to press home their demands, which still retains the bedrock of the system evolution. It eliminates the in-person complaints filing with the use of a computer-based platform. This way, the anonymity of consumers is protected in cases where highly-sensitive failure in service is being reported. Then, feedback can be received almost instantaneously through the email services associated with registered complaints. Aside, education being considered as a veritable tool of reforms and socialization of the human minds, it is a key ingredient for developing a virile and strong a society as well as a nation. Often, it is a tool for growth and integration, which makes efforts towards ensuring survival non-negotiable by means of adequate supervision and monitoring (Ojo [14]). In the views of Babalola [10], the education quality could mean the value of education (with regards to its input, the process of teaching-learning and the outcomes). Several studies are devoted for developing tools to assist the students [13, 20, 21]. Related Studies Help desk systems serve as a sole platform for contact initiation among IT staff as well as the users according to Al-hawari and Barham [16]. The authors use an effective ticket categorizing based on machine learning approach to link a help desk ticket to its corresponding service point at the point of commencement whose goal is to minimize time taken to resolve a ticket while conserving resources and human efforts, and raising the satisfaction of users. In the system, an administrator component ensures services on offering are to be determined such as creation of roles for users, tricking handling of tickets, and reports management and generation. Again, it enables a user component to help employees to file complaints or issues, services requests, and sharing of information among the IT staff on the basis of the tickets assigned by help desk. Attanasio, Sotiropoulos, and Alami [17] presented an advanced structure of help desk in the field of tourism. The purpose of the study was to evolve a design for an interactive counter by applying the robotics techniques and the technologies in woodworks. It affords the customers a fresh experience through an intelligent and social interaction in a connected scenario with wood as the basic material. Omaka et al. [11] relied on the capability of mobile application to build a complaint solution that can be operated on android devices. It enables customers report service failure just at the point of failure. Data confidentiality, interactive feedback function, simplified complaint registration and issues resolution transparency are unapplicable to the existing system, which were undertaken for the SERVICOM office and the customers. It affords the administrative side of the SERVICOM office to collect and manage complaints arising from customers at various offices within the SERVICOM Nigeria. These create an effective customer complaints model for guaranteeing the quality of service of SERVICOM Nigeria.

834

A. A. Alfa et al.

Azahari et al. [18] investigated the ICT Complaint Management System of the tertiary education institutions in Brunei Darussalam and hampers on the knowledge to gain insights into the procedure of complaints resolution. The study diagnosed a number of problems requiring urgent solutions namely: poor process of making decisions, methods of communication, and manpower deficiency. Prasanthi et al. [19] undertook a study with sole goal of advancing a web-based system for placement and managing complaints of the selected university. It removes the work of the guide by substantially providing lesser desk functions and the time frame. Again, it built a Complaint Portal which enables in the new individual to enlist, and the subsequently authorized to make use of the portal for filing complaints such as hall of study hall, Inns and transportation arrangements.

3 Methodology Conceptual Design of the New System The new system conceptual layout is composed of the input, the process and the output components as depicted in Fig. 1.

INPUT

PROCESS

OUTPUT

Fig. 1. The conceptual design layout of the new system.

In Fig. 1, the major features and functions of the proposed students’ results help desk system are explained as follows: Input Component. It gathers the data and variables need from diverse data types usually in digital form by means of input devices such as keyboard and mouse. The students, examination officers, school and HODs at different points to generate and feed data into the students’ results help desk system for different processes to be consummated by other entities or components. Process Component. It is the logical and conceptual blend for all relationships and interconnections of the various data elements entered previously. The process component is concerned with the computing and processing of information about different students’ results complaints and requests at any given time for ease of monitoring, tracking, satisfaction, and services delivery. Output Component. It generates outcomes of the processing component using diverse supplied criteria such as complaints, status, complainant information, etc. The information realized from the system is useful to the students, staff and college management in planning, tracking, controlling, improving quality of educational services and students’ satisfaction. Use Case Diagram of the System. The system actors, cases and flow of activities are presented in the use diagram of the system are illustrated in Fig. 2.

Development of Students’ Results Help Desk System

835

Login Account

Approval

Lodge complaint

Student

User logging

Profile

Status

Feedback

Administrative staff

Fig. 2. The use case diagram of the system.

3.1 Experimental Setup Hardware Requirements. The minimal hardware specification needed for the prototyping of the new students’ results help desk system are presented in Table 1. Table 1. The new system prototype hardware required. Parameter

Value

Hard disk drive (HDD)

180 GB

Processor speed

3.0 GB

System processor type

64-bit

Processor name

AMD

Software Requirements. The minimum software specifications needed for the prototyping of the new students’ results help desk system are presented in Table 2. Table 2. The new system prototype software required hardware to software. Parameter

Value

Operating system

Windows 8

Database

MySQL

Application platform

WAMP Server 2.5

User interface

HTML/PHP

Server

Apache Web Server

Browser display

Internet Explorer 10

836

A. A. Alfa et al.

3.2 System Algorithm The step-by-step approach for utilizing the new system include: 1. The prototype application software for the students’ results help desk system (C: \wamp\www\www.resultshelp.net\) is copied to the HDD of target machine using specified hardware and software requirements. 2. The system users’ profiles and login details are generated for accessing the new system’s functionalities. 3. The student lodge all results complaints and issues into the system and unique Identification numbers are allocated accordingly. 4. The Administrator (school examination or help desk officer) receives complaints and requests concerning results of students. 5. The help desk officer verifies the complaints and initiates resolution processes. The actions to be performed include: a. b. c. d. e.

Users management and logging. Complaints management. Add items such as Category, Subcategory, School. View and update digital ledger for items added. Generate reports for students.

6. The digital records of all processes of results correction from students and administrator are transmitted to the server for safekeeping.

4 System Implementation and Results The prototype system is a web-based application software developed for students’ results help desk system. To utilize, first, it is installed on the target PCs by copying its files to the local disk drive, and thereafter creating database and its elements through: http://loc alhost/www.resultshelp.net/. The index page interface is shown in Fig. 3.

Fig. 3. The start-up page of the prototype system.

Development of Students’ Results Help Desk System

837

From Fig. 3, From Fig. 4, the start-up interface contains the menu bar, banner displaying the program information, footer and functions (such as users management, creation of items, complaints managements, reports generation, etc.). The various user logins credentials including emails and passwords are generated to in order to access the system functionalities and programming environment as shown in Fig. 4.

Fig. 4. The student user login account registration.

From Fig. 4, the students are required to create their system login accounts by supplying following information. The student user registers by entering Full Name, Email, Password and Contact No, then, the register button to commit. The login form for students and the administrator users are shown in Fig. 5.

Fig. 5. Student and administrator login forms.

From Fig. 5, the safety of the information on the system is granted by the encrypting the password information of system users, which avoid sniffing and unauthorized accesses of individuals. Therefore, the login form is used to verify the Email and Password entered for the student and administrator before granting new session to the application programming environment as shown in Fig. 6.

838

A. A. Alfa et al.

Fig. 6. Results complaint registration form.

From Fig. 6, the students are the second category of the system users because, the system intends to assist them in lodging complaints about results published. Similarly, the students can assess other functionalities on their dashboard such as complaint history, profile management, logout, and login credentials management. Similarly, the Administrator or designated staff after successful login have ability to perform several operations including manage complaint, manage users, add category, add subcategory, add school, user login log, and log out as shown in Fig. 7.

Fig. 7. The administrator/staff operations.

This paper introduced an online students’ results help desk system based on three mode of operation including: Not Process Yet Complaint, Pending Complaint, and Closed Complaints as shown in Fig. 8.

Development of Students’ Results Help Desk System

839

Fig. 8. Not Process Yet Complaint entry form.

From Fig. 8, the View Details button on Action column can be used to modify the status of the students’ complaints to in-process or closed item depending on the stage of processing as shown in Fig. 9.

Fig. 9. The interface of Take action functions of the administrator.

840

A. A. Alfa et al.

5 Conclusion This paper implemented a prototype system provides the simplified means for the thirdtier higher institutions to manage numerous challenges occasioned by students’ results publications. This electronic-based is recommended to replaces the manual approach of lodging and resolving complaints concerning students’ results through the appropriate generation of detailed trails of complaints from period of lodgment to final closure. This system enables students and staff to process and resolve results complaints and associated issues without the need for face-to-face contacts. It is available, secure, useable, and effective. It assists the College to track and monitor progress of students’ results complaints; thereby increasing the quality of educational services and satisfaction. The system facilitates the management of results complaints through the automation of tasks such as booking of complaints, online notifications to related parties, customer logs, etc. It helps to define and minimize the activities related to the reportage and processing of different students’ results complaints. However, the level of satisfaction and usability of the new system can be measured quantitatively in order to ascertain the influence on the learning and educational achievements learners in third-tier higher institutions in Nigeria.

References 1. Rusakov, D.V., Loseva, D.M.: Development of an information system for managing user requests using bitrix. In: 2021 IEEE Conference of Russian Young Researchers in Electrical and Electronic Engineering (ElConRus), pp. 1055–1058. IEEE (2021) 2. Alboaneen, D., et al.: Smart information desk system with voice assistant for universities. Int. J. Electr. Comput. Eng. 11(6), 5206–5215 (2021) 3. Nikiforov, O., Mukhametova, L.R.: Key aspects of implementing the Help Desk system in an educational institution. In: Proceedings of the III International Scientific and Practical Conference, pp. 1–3 (2020) 4. Chen, M.: Interaction and collaboration in international office’s help desk setting. In: Sun, Y., Li, L., Cai, H. (eds.) Asian Research on English for Specific Purposes, pp. 269–283. Springer, Singapore (2020). https://doi.org/10.1007/978-981-15-1037-3_16 5. Asha, V.G., Babu, K.R.M.: On-line help desk for college departmental activities. In: 2017 International Conference on Intelligent Computing and Control Systems (ICICCS), pp. 925– 930. IEEE (2017) 6. Dibitonto, M., Leszczynska, K., Tazzi, F., Medaglia, C.M.: Chatbot in a campus environment: design of LiSA, a virtual assistant to help students in their university life. In: Kurosu, M. (ed.) HCI 2018. LNCS, vol. 10903, pp. 103–116. Springer, Cham (2018). https://doi.org/10.1007/ 978-3-319-91250-9_9 7. Munasinghe, C.D.: Analytics-Driven Customer Complaint Management System. Master’s of Information Technology Dissertation. University of Colombo, Georgia, pp. 1–51 (2019) 8. Zaza, S., Junglas, I., Armstrong, D.J.: Who needs the help desk? Tackling one’s own technological problem via self IT service. Inf. Organization 31(3), 100367 (2021) 9. Justitia, A., Zaman, B., Putra, D.K.: Evaluating the quality of a help-desk complaint management service using six-sigma and COBIT 5 framework. In: AIP Conference Proceedings, vol. 2329, no. 1, p. 050009. AIP Publishing LLC (2021) 10. Nwokorie, E.C.: Service recovery strategies and customer in selected hotels in Lagos State, loyalty Nigeria. Net J. Bus. Manag. 4(1), 1–8 (2016)

Development of Students’ Results Help Desk System

841

11. Omaka, S., Onwudebelu, U., Okemiri, H.: Enhanced quality service assurance system: a better approach to service delivery. J. Sci. Rep. 3(1), 51–68 (2021). https://doi.org/10.5281/zenodo. 5256090 12. Rogers, Y., Sharp, H., Preece, J.: Interaction Design-Beyond Human-Computer Interaction, 3rd edn. John Wiley & Sons, New York (2011) 13. Erekata, O., Azeta, A., Misra, S., Odusami, M., Ahuja, R.: Development of a text and speech enabled conversational agent for students’ activities planning using dialog flow. In: Singh, P.K., Veselov, G., Vyatkin, V., Pljonkin, A., Dodero, J.M., Kumar, Y. (eds.) FTNCT 2020. CCIS, vol. 1395, pp. 486–499. Springer, Singapore (2021). https://doi.org/10.1007/978-98116-1480-4_44 14. Ojo, B.: Supervision and quality assurance strategies in education: implication for educational policy making. Afr. Res. Rev. 1(2), 1–14 (2008) 15. Babalola, J.B.: Management of Primary and Secondary Education in Nigeria. NAEP Publication, Ibadan (2004) 16. Al-Hawari, F., Barham, H.: A machine learning based help desk system for IT service management. J. King Saud Univ. Comput. Inf. Sci. 33(6), 702–718 (2019) 17. Attanasio, S.D., Sotiropoulos, T., Alami, R.: Design and development of the first prototype of a social, intelligent and connected help desk. In: 3rd International Conference on ComputerHuman Interaction Research and Applications, pp. 120–127 (2019) 18. Azahari, L.M.H., Ason, M.L.A., Rossiman, N.D., Sion, W.T., Idris, N.A.H..: ICT complaint management within a higher education institute in Brunei Darussalam: a case study. Int. J. ‘Umranic Stud. Jurnal Antarabangsa Kajian’ Umran 3(1), 13–22 (2020) 19. Prasanthi, S., Maganty, C.S., Mupparaju, M.P., Kilaru, M.: An complaint and placement management system using servicenow. J. Crit. Rev. 7(7), 201–204 (2020) 20. Ojajuni, O., et al.: Predicting student academic performance using machine learning. In: Gervasi, O., et al. (eds.) ICCSA 2021. LNCS, vol. 12957, pp. 481–491. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-87013-3_36 21. Azeeta, A., Misra, S., Odusami, M., Peter, O.U., Ahuja, R.: An intelligent student hostel allocation system based on web applications. In: Singh, P.K., Singh, Y., Kolekar, M.H., Kar, A.K., Chhabra, J.K., Sen, A. (eds.) ICRIC 2020. LNEE, vol. 701, pp. 779–791. Springer, Singapore (2021). https://doi.org/10.1007/978-981-15-8297-4_62 22. Adedeji, A., Ibukun, A., Rapheal, O., Misra, S., Ahuja, R.: Employability skills: a web-based employer appraisal system for construction students. In: Abraham, A., Siarry, P., Ma, K., Kaklauskas, A. (eds.) ISDA 2019. AISC, vol. 1181, pp. 612–621. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-49342-4_59

Performance Evaluation of Machine Learning Techniques for Prescription of Herbal Medicine for Obstetrics and Gynecology Problems Oluwasefunmi Arogundade1 , Adeniyi Akanni2 , Sanjay Misra3(B) , Temilade Opanuga1 , Oreoluwa Tinubu1 , Muhammad Akram4 , and Jonathan Oluranti5 1 Federal University of Agriculture, Abeokuta, Ogun State, Nigeria

{arogundadeot,tinubuco}@funaab.edu.ng 2 Anchor University Lagos, Agege, Nigeria [email protected] 3 Department of Computer Science and Communication, Ostfold University College, Halden, Norway [email protected] 4 Department of Eastern Medicine, Government College University, Faisalabad, Pakistan 5 Center of ICT/ICE Research, Covenant University, Ota, Nigeria

Abstract. Women especially low income earners opt for herbal medicine to maintain their health status, curative purposes as well taking care of their Obstetrics (OB) and Gynecology (GYN) problems. The cost of herbal medicines are low compared to pharmaceutical drugs, however, several potential risks arises from the use of incorrectly prescribed herbal therapies. These arouse our interest in this study by conducting a comparative analysis of machine learning techniques for the prescription of herbal solutions for OB-GYN issues. This research involves intensive study of local herbal remedies and survey of traditional health care delivery within the western part of Nigeria. Four machine learning algorithms, such as Multilayer Perceptron, J48 Decision Trees, Naïve-Bayes and IBK (Instance Based Learner) were employed on thirty (30) data features for the performance evaluation process. This is aimed at obtaining the most suitable machine learning algorithm for an efficient herbal medicine prescription model for OB-GYN diseases. In this work, assessment and comparison of the four machine learning algorithms, specifically Instance-Based Learner (IBK), Multi-Layer Perceptron (MLP), J48 decision tree, Naïve-Bayes were carried out. Results showed an achieved accuracy of 100% using the Naive Bayes, MLP, IBK classification algorithms. We can reduce mortality rate among less privileged women through accurate diagnosis and prescription of herbal remedies. Keywords: Obstetrics · Gyncology · Prescription · Herbal medicine · Machine learning

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 842–851, 2022. https://doi.org/10.1007/978-3-030-96299-9_79

Performance Evaluation of Machine Learning Techniques

843

1 Background The persistent increase in the death rate of women during pregnancy and at childbirth especially in under-developed and developing countries where the citizens can hardly afford conventional medical care is alarming. Poverty is a major factor in maternal mortality. In [2], it was emphasized that the ability to detect ectopic pregnancy in good time was hampered by lack of financial capacity. Poverty has led many women to seek medical attention in places rather than standard hospitals where basic facilities are found. The essentials of life are good food, clothing and shelter. In the absence of these essentials, the quality of life is affected, which has direct effects on an individual’s health status. In [3], a comparative study on the quality of life in the recurrent nature of miscarriages was conducted. Their study revealed that women with better quality of life had lesser miscarriages as compared with those with poor quality of life. Authors [4] listed low income which is akin to poverty as being associated to postnatal health care utilization. Several women seeking Obstetrics (OB) and Gynecology (GYN) care are plagued with many challenges. In recent times, women who have little or no access to public healthcare have resorted to the use of herbal medicines to treat their OB and GYN issues. Lack of adequate medical services contributes to the high maternal mortality rate, especially in Nigeria. The common OB and GYN problems include uterine fibroids, polycystic ovarian syndrome, cervical dysplasia, urinary incontinence, menstrual disorders, pelvic floor prolapse, pre-eclampsia etc. The use of herbal treatments for managing complications of the female reproductive system have increased significantly over the last decade, without data on safety, potency or rates of usage. Herbal medical practices are often based on mystical concepts [13]. This differentiates them from western medicines that are based on scientifically proven evidences [19]. Often times, herbal medicines are self-administered, with no dosing recommendations. Dangerous interactions of herbal medications with other pharmaceutical medications are common and can lead to devastating effects. Some other challenges with herbal medications include lack of regulations, poisoning and potentially serious health problems. Improving patient’s safety by reducing medication errors is a top priority [11]. Prescription and administration of drugs are the most frequent sources of medication errors. The conventional herbal medicine prescription systems have been observed to have a low-quality process while lacking accuracy. Herbal medicines usually contain a range of pharmacologically active compounds, in some cases it is unknown which ingredients are important for the therapeutic effect [10]. As such, herbal remedies require accurate prescriptions for treating temporary and permanent conditions and for the improvement of the overall well-being of an individual. Machine learning is a powerful tool for identification, prediction and evaluation in various field [20–24] and in particular for healthcare systems. It is a fundamental technology needed to coherently process data that surpasses the ability of the human brain. Machine learning approaches automatically learns to recognise complex patterns and make intelligent decisions based on data [6]. Machine learning provides a wide range of techniques for prescription of medications. The large data sets generated from patients can be processed for an improved and automated prescription system. The quickening intensity of machine learning techniques in diagnosing diseases, arranging and grouping health data helps medical practitioners to accelerate clinical dynamics.

844

O. Arogundade et al.

Patients diagnosed with OB-GYN diseases can therefore get accurate prescription of herbal medicines. In this study, we present an approach to herbal medicine prescriptions for Obstetrics and Gynecology problems using machine learning algorithms. The performance of the selected machine learning classifiers will be evaluated for effective prescription of herbal medicine for various OB-GYN problems. The remaining section of this paper is organized as follows; Sect. 2 provides the review of literature related to this research work. The research methodology is presented in Sect. 3. Results and discussion are presented in Sect. 4, while Sect. 5 concludes the paper and presents future intentions on the research.

2 Literature Review Researches on herbal medicine, herbal drugs and herbal medicaments are being done by several researchers [1, 14–16]. Since alternate therapy is being so much sought after especially in developing countries, it becomes highly imperative to derive better means of managing the process of this class of health care. Ref [5] emphasized that proper diagnostic evaluation should be carried out. In the absence of this, prescription of drugs and its administration can be difficult for effectiveness. After a proper diagnosis is done, prescription of the appropriate drugs for curative or maintenance purposes follows. When diagnosis fails and diseases linger, it may have adverse effects on the female reproductive organ [5]. Ref [17] analysed the use of traditional medication for the treatment of vaginal infections among women of reproductive age in Nigeria. They proved that herbal remedies are effective in the treatment of different types of vaginal infections as vaginal infection is one of the major threats faced by women generally, in reproductive ages in the area. Their investigation showed that most women in the study area utilised traditional medication because of the failure of the orthodox therapy to properly treat the infection, high cost of treatment in the conventional therapies and perceived effectiveness of traditional therapy. Ref [12] systematically analyzed the herbal therapy in to cure the COVID-19 patients. In [1], a mobile application was developed for herbal medicine prescription for obstetrics and gynecology problems. Machine learning technology is well suited for analysing data. In the medical domain, diagnosis, classification and treatment are the main tasks for a physician. These tasks can be effectively accomplished with machine learning techniques. Ref [11] proposed a machine learning-based clinical decision support system for the identification of prescriptions with a high risk of medication errors. This was aimed at improving the safety of patients. Ref [8] proposed a medicine prescription system using Naive Bayes classifier. Ref [7] devised a machine learning-based system capable of predicting pregnancy outcomes form a patient’s data. In [9], the study evaluated the performances of various ML techniques for predicting the risk of developing hypertension at individuals. Ref [18] analyzed the use of machine learning techniques to assist pharmacists review medication orders.

Performance Evaluation of Machine Learning Techniques

845

3 Methodology The process flow of the proposed system for evaluating machine learning techniques for the prescription of herbal medicine for Obstetrics and Gynecology problems is presented in Fig. 1. This research was carried out through an intensive study of local herbal remedies and survey of traditional health care delivery within the western region of Nigeria. Four classification algorithms such as Multilayer Perceptron, J48 Decision Trees, Knearest neighbor called IBK (Instance Based Learner) and Naive Bayes were deployed in a machine learning data analysis tool (WEKA 3.8.3) for the experimental evaluation process to obtain the most suitable technique for an efficient prescription model.

Fig. 1. A sequence diagram showing the process flow of the proposed system

3.1 Description of Data Set Data on OB-GYN diseases were collected through a series of oral consultation and interviews with some traditional herbal practitioners in Ogbomosho, Ibadan and Abeokuta in Nigeria. Information on the names of some herbal drugs and its specifications were also obtained from the website of a popular herbal medicine producing company in Nigeria. A symptomatic analysis of some common OB-GYN diseases was done, and it was observed that several of the diseases share common symptoms. However, the mixture of some symptoms may well be accustomed to distinguish a disease. In Table 1, the obtained symptoms are summarized; this constitutes the features and attributes of both the train and test dataset. Table 2 shows a list of some OB-GYN diseases obtained and a list of manufactured herbal drugs that can be prescribed for patients. Table 3 shows the output class labeling with respect to prescribed drugs as a function of the symptoms exhibited in each class instance. 3.2 Data Preprocessing Data processing is an important aspect of the machine learning process owing to the fact that raw data is often dirty, noisy, inconsistent, incomplete etc. The set of data processing activities includes data cleaning, data integration, data transformation, data reduction and data discretization.

846

O. Arogundade et al. Table 1. Summarized symptoms of OB-GYN diseases

Symptoms

Symptoms_ID

Symptoms

Symptoms_ID

Chest pain

S1

Misconception

S18

Dizziness

S2

Frequent urination

S19

Pulpitation

S3

Constipation

S20

Breath shortness

S4

Prolonged menstrual cycle

S21

High fever

S5

Sexual pain

S22

Breast inflammation

S6

Menstrual cramps

S23

Boils

S7

Breast aching

S24

Bowel bleeding

S8

Menstrual pain

S25

Anal itching

S9

Urinary difficult

S26

Anal swelling

S10

Vaginal itching

S27

Feaces leakage

S11

Vaginal discharge

S28

Headache

S12

Extreme anger/anxiety

S29

Blur vision

S13

Depression

S30

Rib pains

S14

Backache

S31

Vomiting

S15

Pelvic pain

S32

Abdominal pains

S16

Insomnia

S33

Irregular menstruation

S17

Table 2. Summarized ObGyn diseases and prescribed drugs S/N 1. 2. 3.

9.

Disease Cardiac Infertility Pre-menstrual disorder Endometriosis Pre-menstrual Syndrome Uterine fibroid Chronic pelvic pain Malposition of Uterus Cervicitis

10.

Mastodynia

4. 5. 6. 7. 8.

Disease_ID A1 A2 A3

S/N 1. 2. 3.

Prescribed Drugs HBP capsule FIJK (liquid & capsule) Formular YK 35

Drug_ID D1 D2 D3

A4 A5

4. 5.

Multi vitamin energy plus FORMULAR YK 15

D4 D5

A6 A7

6. 7.

Cag-Mag-100 tablets Quercetin plus

D6 D7

A8

8.

Ruzu bitters

D8

A9

9.

Sappiro Lemon GinsengLiquor

D9

A10

10.

D10 Chelated Zinc-100 tablets

11.

Staphylococus

A11

11.

D11 FORMULA YK 800

12.

Leucorrhea

A12

12.

D12 KOMPRESSOR PILLS

13.

Hermorrhoid

A13

Performance Evaluation of Machine Learning Techniques

847

Table 3. Mapping of symptoms of OB-GYN diseases with prescribed drugs Diseases

Symptoms pairing

Drugs

Production Company

A1

S 1 , S2 , S3 , S4

D1

YEMKEM

A2

S16 , S17 , S18

D3

YEMKEM

A3

D4

Tianshi

A4

S4, S14 , S29 , S30 S16 , S22 , S23

D6

Tianshi

A5

S12 , S15 , S16 , S24 , S31

D9

Abllat nig. Limited

A6

S17 , S19 , S20 , S21 , S32

D3

Universal Services

A7

S12 , S16 , S22 , S31

D3

Universal Services

A8

S22 , S28 , S32

D10

GNLD

A9

S22 , S28 , S32

D8

RUZU Herbal Bitters

A10

S1 , S6 , S24

D7

Tianshi

A11

S 5 , S6 , S7

D11

YEMKEM

A12

S16 , S26 , S27

D5

YEMKEM

A13

S8 , S9 , S10 , S11

D2

OKO OLOYUN

3.3 Feature Selection Feature selection is a very important part of constructing an exceptional prescription model. It identifies and removes the maximum amount of insignificant and inessential features as possible. It reduces the scope of the data and allows learning algorithms to have dominance faster and more efficiently. Feature selection algorithms have two constituents; a selection algorithm that produces recommended subdivisions of features and aims at seeking out an ideal subset; and an evaluation algorithm that regulates how good a projected feature subset is, recurring some quantity of goodness to the selection algorithm. In this work, a spontaneous R-based Machine Learning feature selection algorithm is engaged to rank the attributes utilizing Information Gain, Correlation and Relief-F attribute at that point and assess the significance of quality by deciding the entropy gain with an application to the outcome, and positions the properties by their distinct evaluations. Attributes that have information gain > 0 are subsequently utilized in the development of the machine learning models.

4 Results and Discussion The Multilayer Perceptron, IBK, J48 Decision Trees and Naive Bayes algorithms were all trained in phase one using the training dataset. The comparative analysis of the algorithms is done through the lenses of Accuracy, MEA, Duration of learning and ROC and are presented in Table 4. The performance of the classifiers for prescription are evaluated based on three metrics-Precision, Recall and ROC area as presented in Fig. 2. The output models generated in phase one on each of the four classifiers are loaded for

848

O. Arogundade et al.

the second phase and are all tested with the test dataset for prescription analysis. The prescription is carried out on thirty (30) data features based on the underlying attributes comprising of symptoms and diseases and the accuracy of the classifiers are as presented in Fig. 3. Table 5 and Fig. 4 shows the weighted averages of the attribute ranking phase and its comparative graph respectively. Table 4. Performance evaluation of the four prescriptive models S/N

Algorithm

Comparative analysis Accuracy

MEA

Duration (s)

ROC

1

Naïve Bayes

100%

0.0556

0

1.000

2

J48

84.6%

0.0256

0

0.980

3

IBK

100%

0.0733

0

1.000

4

Multilayer perceptron

100%

0.0162

0

1.000

Fig. 2. Performance evaluation of the classifiers

The training phase with its performance matrix presented in Fig. 2 shows that MP, IBK and Naïve-Bayes are the best classifiers in herbal medicine prescriptive analytics for obstetrics and gynecology problems. The attribute ranking analysis on Table 5 clearly shows diseases as the attribute with the highest information gain in Relief-f score and the major determinant of the class of prescription. Symptom 4 comes next as the most ranked independent attribute with significant correlation with the dependent class attribute. The accuracy of the prescriptive model with respect to the four algorithms, as shown on Fig. 3, carried out on thirty (30) data features shows high performance during the test phase of the model. As may be observed on the performance evaluation Table 4, the MLP, IBK and Naïve-Bayes classifiers prescribed correctly for all 30 features, while the J48 classifier was inaccurate for instances 7, 13 21 and 22, hence the lower accuracy level recorded and ROC area decline. The implication is that the classifier is not suitable for the OB-GYN herbal medicine prescription system because with the symptoms and diseases attribute

Performance Evaluation of Machine Learning Techniques

Fig. 3. ML algorithms for the OB-GYN herbal medicine prescription

Table 5. Weighted averages of the attribute ranking I.G

Correlation

Relief-F

Symptom 1

2.5654

0.2150

0.6140

Symptom 2

3.0851

0.1850

0.6660

Symptom 3

3.2389

0.1780

0.6790

Symptom 4

2.0349

0.2570

0.5840

Diseases

3.3927

0.1710

0.6920

Fig. 4. Attribute ranking of OB-GYN dataset

849

850

O. Arogundade et al.

supplied, a corresponding drug prescription applies. From the result of the prescription and consequent comparative analysis aforementioned, MLP, Naïve-Bayes and IBK are the appropriate models for the herbal medicine prescription of OB-GYN diseases which is in tandem with the performance metrics presented on Fig. 3.

5 Conclusion Machine learning techniques have been delineated to give generous forecast achievement in different application spaces, including medication and medicinal services. In this work, assessment and comparison of the four machine learning algorithms, specifically Instance-Based Learner (IBK), Multi-Layer Perceptron (MLP), J48 decision tree, NaïveBayes were carried out. Experimental results determine the capability of the system with an achieved accuracy of 100% using the Naive Bayes, MLP, IBK classification algorithms which will invariably reduce mortality rate among less priviledged women through accurate diagnosis and prescription of herbal remedies. In future work, a larger dataset can also be employed to further establish the usability and accuracy of the system.

References 1. Arogundade, O., Usman Owoade, A.M., Ikotun, M., Shukla, M.: Design and implementation of mobile application for herbal medicine prescription for obstetrics and gynecology problems. Covenant J. Inform. Commun. Technol. 7(2), 1–17 (2019) 2. Mooij, R., Mgalega, G.C., Mwampagatwa, I.H., van Dillen, J., Stekelenburg, J.: A cohort of women with ectopic pregnancy: challenges in diagnosis and management in a rural hospital in a low-income country. BMC Pregnancy Childbirth 18(1), 1–7 (2018) 3. Wudineh, K.G., Nigusie, A.A., Gesese, S.S., Tesu, A.A., Beyene, F.Y.: Postnatal care service utilization and associated factors among women who gave birth in Debretabour town, North West Ethiopia: a community-based cross-sectional study. BMC Pregnancy Childbirth 18(1), 1–9 (2018) 4. Taran, F.A., Kagan, K.O., Hübner, M., Hoopmann, M., Wallwiener, D., Brucker, S.: The diagnosis and treatment of ectopic pregnancy. Dtsch. Arztebl. Int. 112(41), 693 (2015) 5. Rajkomar, A., Dean, J., Kohane, I.: Machine learning in medicine. N. Engl. J. Med. 380(14), 1347–1358 (2019) 6. Oguntimilehin, A., Adetunmbi, A.O., Abiola, O.B.: A Machine learning approach to clinical diagnosis of typhoid fever. Mach. Learn. Approach Clin. Diagn. Typhoid Fever 2(4), 1–6 (2013) 7. Hassan, M.R., Al-Insaif, S., Hossain, M.I., Kamruzzaman, J.: A machine learning approach for prediction of pregnancy outcome following IVF treatment. Neural Comput. Appl. 32(7), 2283–2297 (2018). https://doi.org/10.1007/s00521-018-3693-9 8. Suma, S., et al.: Medicine prescription system by using machine learning. Glob. Res. Dev. J. Eng. 3(5), 1–6 (2018) 9. Sakr, S., et al.: Using machine learning on cardiorespiratory fitness data for predicting hypertension: The Henry Ford ExercIse Testing (FIT) Project. PLoS One 13(4), e0195344 (2018) 10. Schulz, V., Hänsel, R., Tyler, V.E.: Rational Phytotherapy: A Physician’s Guide to Herbal Medicine. Psychology Press, Hove (2001)

Performance Evaluation of Machine Learning Techniques

851

11. Corny, J., et al.: A machine learning–based clinical decision support system to identify prescriptions with a high risk of medication error. J. Am. Med. Inform. Assoc. 27(11), 1688–1694 (2020) 12. Ang, L., Lee, H.W., Choi, J.Y., Zhang, J., Lee, M.S.: Herbal medicine and pattern identification for treating COVID-19: a rapid review of guidelines. Integrative Med. Res. 9(2), 100407 (2020) 13. Micozzi, M.S.: Fundamentals of Complementary, Alternative, and Integrative Medicine-Ebook. Elsevier Health Sciences, New York (2018) 14. Mukherjee, P.K.: Quality Control and Evaluation of Herbal Drugs: Evaluating Natural Products and Traditional Medicine. Elsevier, New York (2019) 15. Lam, D.H., Bell, S.M., Hira, R.S.: Concomitant use of antiplatelets and anticoagulants in patients with coronary heart disease and atrial fibrillation: what do recent clinical trials teach us? Curr. Atheroscler. Rep. 20(1), 1–10 (2018) 16. Khan, M.S.A., Ahmad, I.: Herbal medicine: current trends and future prospects. In: New Look to Phytomedicine, pp. 3–13. Academic Press, Cambridge (2019) 17. Iorkosu, T.S., Emmanuel, V.O., Benjamin, G.A., Tsembe, D.D., Aluka, T.M., Ajai, F.: Utilisation of Traditional Medication for the Treatment of Vaginal Infection among Women of Reproductive Age in Makurdi Metropolis of Benue State, Nigeria (2020) 18. Thibault, M., Lebel, D.: An application of machine learning to assist medication order review by pharmacists in a health care center. medRxiv 19013029 (2019) 19. Oladosu, J.B., Adigun, M.O., Mbarika, V.: Towards a pharmaceutical ontology for African traditional herbs. In: Proceedings of the World Congress on Engineering and Computer Science (WCECS 2012), vol. 1 (2012) 20. Ojajuni, O., et al.: Predicting student academic performance using machine learning. In: Gervasi, O., et al. (ed.) ICCSA 2021. LNCS, vol. 12957, pp. 481–491. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-87013-3_36 21. Ogundokun, R.O., Misra, S., Ogundokun, O.E., Oluranti, J., Maskeliunas, R.: Machine learning classification based techniques for fraud discovery in credit card datasets. In: Florez, H., Pollo-Cattaneo, M.F. (eds.) ICAI 2021. CCIS, vol. 1455, pp. 26–38. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-89654-6_3 22. Azeez, N.A., Lawal, A.O., Misra, S., Oluranti, J.: Machine learning approach for identifying suspicious uniform resource locators (URLs) on Reddit social network. Afr. J. Sci. Technol. Innov. Dev. 1–9 (2021) 23. Awotunde, J.B., Ogundokun, R.O., Jimoh, R.G., Misra, S., Aro, T.O.: Machine learning algorithm for cryptocurrencies price prediction. In: Misra, S., Kumar Tyagi, A. (eds.) Artificial Intelligence for Cyber Security: Methods, Issues and Possible Horizons or Opportunities. SCI, vol. 972, pp. 421–447. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-722364_17 24. Ogundokun, R.O., Awotunde, J.B., Misra, S., Abikoye, O.C., Folarin, O.: Application of machine learning for ransomware detection in IoT devices. In: Misra, S., Kumar Tyagi, A. (eds.) Artificial Intelligence for Cyber Security: Methods, Issues and Possible Horizons or Opportunities. SCI, vol. 972, pp. 393–420. Springer, Cham (2021). https://doi.org/10.1007/ 978-3-030-72236-4_16

Data Science for COVID-19 Vaccination Management Elham Rezaei1 , Kajal Ghoreyshi1 , and Kazi Masum Sadique2(B) 1

2

Faculty of Technology, Department of Informatics, Linnaeus University, Kalmar, Sweden {er222rv,kg222jf}@student.lnu.se Department of Computer and Systems Sciences (DSV), Stockholm University (SU), Stockholm, Sweden [email protected]

Abstract. Medicine and health are essential sectors in industrial societies. Extracting knowledge from the vast amount of data related to disease records and medical records of individuals using the data mining process can identify the laws governing the creation, growth, and spread of disease and provide valuable information to identify the causes of disease. To diagnose, predict and treat diseases according to the prevailing environmental factors to provide health professionals and practitioners. The result of this issue is to increase life expectancy and create peace for the people of the society. With the spread of the COVID-19 virus in recent months worldwide, various organizations are working to find ways to combat the virus. By using data mining technology, intelligent systems can be developed that can automatically understand and interpret the medical characteristics of individuals and extract useful information that can play an effective role in the process of vaccine supply chain management. In this article, we have proposed a solution for the efficient COVID-19 vaccination management. Keywords: Data science · COVID-19 Vaccination · Patient data

1

· Supply chain management ·

Introduction

The emergence of disease with the potential to spread in a vast geographic area occurs every few years around the world, such as Cholera, Ebola, Flu, SARS, etc. These are defined in different categories depending on their predicted rates and spread in the geographical area: an endemic refers to diseases spread among a population at a predicted rate. An epidemic is an unpredicted increase of disease in a large geographical area. A pandemic is an epidemic that spreads globally. More vast disease spreads, more challenging will be its prevention and control and supplying the medical tools in spreading areas and giving them access to aids. c The Author(s), under exclusive license to Springer Nature Switzerland AG 2022  A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 852–861, 2022. https://doi.org/10.1007/978-3-030-96299-9_80

Data Science for COVID-19 Vaccination Management

853

In late 2019, a virus infection started to spread from China that the World Health Organization (WHO) categorized as a pandemic, named Covid-19 [1]. Within about one year, in the absence of pharmaceutical intervention, some containment methods such as isolation, quarantine, and borders control were used to control the speed and extent of Covid-19 spreading [2]. At the same time, pharmaceutical manufacturing companies started a race to develop effective vaccines. After passing one year, as of February 2021, about seven approved vaccines began to be distributed in countries, and vulnerable populations are supposed to be the highest priority for vaccination. Now the challenge turns to global equitable access to vaccines that are the primary responsibility of COVAX [3,4]. COVAX is the vaccine pillar of ACT (Access to Covid-19 Tools) that is a global collaboration to accelerate equitable access to Covid-19 tools, including vaccines. According to WHO 2021 [4], one of the responsibilities of COVAX is: “working with governments and manufacturers to ensure fair and equitable allocation of the vaccines for all countries”. The world has stepped in new phase and now the challenge in front of COVAX is not vaccine but it turns to “vaccination” referring to problems in infrastructures, administration level and accessories including and not limited to: – Adaptation in the cold chain such as shortage of refrigerators in distribution systems. – Supplying the vaccination accessories such as syringes, waste bags etc. – Tracking of vaccinated people, the source of the vaccine they’ve gotten and if both doses are from the same source. These millions of most advanced manufactured vaccines will mean nothing to save the lives or economy’s recovery if they don’t reach the right consumers who need them the most at the right time and in the right way. One of the factors of on-time vaccination is the analytical prediction of demand. COVAX and likewise governments need to make decisions to determine the highest prioritized groups based on the criteria for vaccine recipients and as a result to predict the demand of each country and to estimate how long it will take to vaccinate the standard population throughout the world in order to reach herd immunity. Since the beginning of the pandemic, a big amount of data has been generated, which is the basis of COVAX’s decision to distribute the vaccine fairly, as well as to predict the demand for more doses to be for vaccine manufacturers. The more accurate and fast this data is generated and made available to COVAX, the more accurate and faster the management of this data and the macro decision-making process will be. Since the data has a down-to-up flow to make information, it is motivating that small segments in each area make its own data processed and visualized. For example, if each municipality handled data required for COVAX in its area, it will help the government to make decisions in real time and to exchange it to COVAX with no delay. In this project (as a fictive case and at a conceptual level) we want to propose to any municipality to make faster and more accurate decisions in the following cases by using big data technologies using data from Electronic Health Records

854

E. Rezaei et al.

(EHR): (1) Counting the number of the most vulnerable population for vaccination according to priorities set by the government as age classifications, the elderly, medical staff, people at risk of underlying diseases or (2) Estimating the exact number of vaccines in demand in the coming roll-out (3) Tracking vaccinated individuals and planning for a second doses of the same source (4) Appropriate ordering time to supply inventory. Although this project is presented at a conceptual level, it can be kind of the help for governments to make decisions based on integrated information, if implemented systematically by all regions.

2

Background

A fast worldwide vaccination is the key to reach herd immunity and put the end to the pandemic. It is estimated that if 20% of the world’s population is vaccinated, the herd immunity may happen. But this amount of demand is beyond seven manufacturers’ scale of production. On the other hand, some rich countries such as the United Kingdom and Canada have secured many doses for vaccinating all their populations that the act was criticized as “wildly uneven and unfair” by the UN [3]. If distribution goes unbalanced like this the world will fall far from the herd immunity and consequently those rich countries with fully vaccinated populations will have no choice to keep their borders closed toward the poor countries who didn’t succeed to be vaccinated. It means the collapse in the international trade and economy will remain in force. The solution is that all countries follow a layer-to-layer vaccination almost simultaneously and in parallel, which requires precise planning backed by accurate data at an accelerated pace. The more the demand for each vaccination is communicated in a timely manner, the more factories can have accurate production planning, and consequently the dosing inventory of each country is properly managed. 2.1

Purpose and Aims of the Project

Although various vaccination-related organizations have displayed publicly the number of people vaccinated in different countries on their platforms, none of them show the number of vulnerable people who are prioritized for vaccination. In other words, these data are inclined to the past, what happened and ended. There is no future-oriented data that leads to a decision on the number of vaccines in demand. The purpose of this project is to use the CRISP-DM data mining technique to help a municipality to build a dashboard in a small scale through which all the information about vaccination in this municipality can be seen. Health status data that can be used in vaccination-related decisions will be entered into the system through public and private health centers in the municipality, then converted into information by being aggregated or summarized. This information is provided to decision makers, including COVAX, in a few hours update intervals

Data Science for COVID-19 Vaccination Management

855

so that they can plan for the production and distribution of more doses. At present, such agility does not exist in the current infrastructure, and it is done carelessly in some parts of the world. The aim of this project is to enable a municipality to receive the demanded doses of vaccines in a timely manner, by providing timely information on the number of vulnerable people, age groups and the number of people receiving the first dose in real time, via a dashboard which is also accessible for other decision-making centers at the governmental level. Since health records HR are not understandable to everyone and decision makers do not have much time to interpret the relationship between the various data, the information should be presented in a very simple way to be perceivable briefly. For example, in the form of bar, line, cake charts and the tools that empowers decision makers to make quick and agile decisions. 2.2

The Specific Goals

The specific goal of the project is offering a Dashboard that displays a visual format of vaccination-related information regarding vulnerable groups to present a range of different performance indicators on one page, like a dashboard in a car, as described in [5], it will have fixed structure by static elements but at the meantime the customization of the dashboard widgets to set other targets for various metrics will be allowed. Through which the decision makers can easily retrieve well-organized information ready to digest and interpret. In the next section, we have presented our proposed solution in detail.

3

Related Works

As a related work, we found that Shankar Sh. et al. [6] identifies challenges of vaccine distribution among which are logistic and communication. They highlight that there are needed digital tools to address global equitable vaccine distribution and proper communication among governments, companies and general public and propose a comprehensive, uniform digital framework for constant tracking and monitoring of vaccine shipments. They also propose a record-keeping system as a vaccination monitoring system that needs to be real-time, transparent and from trusted sources.

4

Modules for Development of the System

To develop the different parts of the project, we divided it into three main models, including pre-implementation, implementation and post-implementation. – In the first module, the focus should be on the data, because the proposed solution is a dashboard, and the data in the design of the dashboard stands as the most important attribute, and the data in the dashboard is as important as the location in the real estate property [7]. It seems that data preparation

856

E. Rezaei et al.

will take the most time. However, since the main goal of this project is to speed up decision-making within the pandemic management process, it is recommended that more data experts join the project team, so that we can do this step as short and fast as possible – In the second module, we enter the design phase of the dashboard, and we have to focus on the visual aspects so that it is perceptible at a glance. A guide will also be needed for non-professional users. – In the third module, we have to focus on support and monitoring, which is an ongoing process because the project data, as described earlier, is constantly changing, which is added to the dashboard with business intelligence tools. Each module refers to a work package including main tasks and sub-task, listed in Fig. 1:

Fig. 1. Project’s work packages of CRISP-DM [13, 14].

5

Related Theory

Different industrial projects have their own methods and processes. Various methods have been developed to perform data mining processes, some popular

Data Science for COVID-19 Vaccination Management

857

models guiding the implementation of data mining projects include Knowledge Discovery Databases (KDD) process model, CRISP-DM and SEMMA but one of the most popular of which is the “cross-industry standard process” method, which is abbreviated to CRISP. An open standard process model that describes the general approaches of data mining professionals, this methodology is the most widely used analytical model. CRISP is a data mining process model as Fig. 2 that describes common strategies used by expert data miners to overcome data mining problems [15].

Fig. 2. The Six-Step CRISP-DM data mining process [14].

The six steps of CRISP-DM are as follows: – – – – – –

6

Business Understanding Data understanding Data preparation Modelling Evaluation Deployment

Proposed System Description

In this section, we have presented our proposed solution for COVID-19 vaccination management using data science.

858

6.1

E. Rezaei et al.

Dashboard

Dashboard is a way of data visualization. A dashboard allows graphical representation of results from any business Intelligence (BI) system. Dashboard allows three common operations: monitoring, analysis and management. A dashboard displays different sorts of data in a single screen. For many cases, a dashboard is used for representation of performance of different systems or analysis results [10]. In our project we are using a dashboard to visualize the data for monitoring of the vaccination process, analysis of current vaccination situation of vaccination in the municipality and allow vaccination related data management. 6.2

User Groups and Roles

As we have mentioned earlier, in our proposed model the dashboard will visualize information to different stakeholders who are the consumer of the information. In our system, we have considered the following stakeholders: – – – –

Healthcare service provider Municipality decision maker Vaccine producer Vaccine distributor/pharmacy

The above stakeholders should be using our proposed system with different roles. As we are handling human-sensitive private data, the dashboard should be different for different user groups. Below, we have presented details of the roles of the different user groups. 1. Healthcare service provider: The healthcare service providers will have more access rights to the patient data. A doctor or nurse will have a dashboard with a summary of the total patients within a certain healthcare clinic. The healthcare service provider will be able to access detailed information for a certain patient by clicking different links from the statistical representation (graphs) of total patients. For example, from the total patient statistics for high-risk groups, a doctor will be able to access the tabular representation of information for a certain group of people or for an exact person registered to that specific clinic. In case a patient moves to another clinic, the doctor at the new clinic may also access the data, but with permission from the data owner who is the patient himself/herself. 2. Municipality decision-maker: This group of users will also have grapes and statistics on their dashboard, from which they will be able to quickly make decisions on the future need of ordering vaccines. But these users will not have access to any specific patient data, as we want to keep the privacy of patients. 3. Vaccine producer: This is an extra facility that is only for producers/importers of vaccines. They will also have limited a dashboard with limited access rights. The main data that will be available via the graphs showing the current and future demand of vaccines for the mentioned municipality.

Data Science for COVID-19 Vaccination Management

859

4. Vaccine distributor/pharmacy: This community of users will also be able to see the demand for vaccines. They actually work as a middleman between the vaccine producer and different clinics within the municipality. 6.3

Tools and Packages

In this section we have described the tools and modules that should be used in our proposed system. Python Programming Language: Python is a high-level programming language (Python) with a rich number of packages and is very popular for its huge machine learning library [10]. Scikit-learn: Scikit-learn is a popular python package which includes implementation of different popular machine learning algorithms [9]. We have considered this as it is suitable for middle scaled data. Django Web Framework: Django is a python based open-source web framework (Django) which is easy to use and content many pre-defined facilities. For example, it allows easy user registration and grouping of users based on their roles in the system [8]. Plotly, Pandas, Matplotlib: Plotly, Pandas, Matplotlib: Plotly, Pandas, Matplotlib are also python packages [12] used for data visualization [11]. We will possibly use a combination of all these packages based on our need for a nice representation of results. 6.4

System Architecture

Our proposed system architecture is presented in Fig. 3. On the left side of Fig. 3, we have the data sources: electronic health records (EHR) of the people living in the municipality, and predefined risk group information (prepared by the medical experts). The data from the EHR are retrieved using application programming interfaces (API) to avoid unnecessary problems related to inter system data access. At the second stage we have python modules for preprocessing of the data which include selecting of certain attributes from the detailed table for further processing at the next step. At the third step, we have the scikit-learn python codes for both clustering and classification of data. The data modeling and evaluation is performed at this step. The results from these two data analysis modules are fed to the next step, which is the deployment of the result via a Django web application. The Django web application will process the authentication of the user group described in the previous section. It will provide role-based authentication to its user. Also, the web server will have its own database for tracking of individuals for the second dose of vaccine. The

860

E. Rezaei et al.

data visualization module is used for the graphical representation of data at the dashboard, which is the final step of our proposed system. As we have already described, the dashboard will show the results of our data mining application for vaccination decision-making system.

Fig. 3. Proposed system architecture.

We have proposed both clustering and classification algorithms to get a comparative result from the system. The users will be able to configure their dashboard to see results produced from the backend of the proposed system.

7

Conclusions and Futureworks

The specific goals of this project were to identify the most vulnerable population according to priorities set by the government as age classifications, the elderly, medical staff, people at risk of underlying diseases. In our proposed model in Fig. 1, we have a set of risk groups that is applied before the classification of data. Our proposed system helps any municipality estimate the exact number of vaccines needed for different healthcare clinics within the municipality. The database connected to the Django web server will provide the facility to keep track of vaccinated individuals. It will support the planning for a second dose for everyone via the dashboard. Also, the system provides information via a dashboard for appropriate ordering time for vaccine supply chain management. This project does not address issues such as the fact that people in some countries are reluctant to get vaccinated. According to a Gallup poll conducted in August 2020, thirty-five percent of respondents are dissatisfied with the vaccine. In this project, we have not considered the preferences of consumers. It is assumed that all people in the eligible groups or who have underlying diseases will be queued to receive the vaccine according to priority. The government is obliged to provide the vaccine for all these people, whether those who want to inject or those who do not want to be vaccinated. Other limitations of this project include limited infrastructure in different countries. In Sweden, for example, because the EHR is used, the project can be used, but in countries with weaker infrastructure, the project does not work. Also, due to political considerations, some of these countries do not provide real information to the World Health Organization and COVAX, which can be considered as another

Data Science for COVID-19 Vaccination Management

861

challenge of this project. But in general, the use of data mining can bring justice and equality to developing and developing countries. The actual implementation project should consider the security issues and issues and rules of GDPR that are implemented in Europe and implement them well. This work can further extend with the implementation of the proposed system.

References 1. WHO Director-General’s opening remarks at the media briefing on COVID19, 11 March 2020. https://www.who.int/director-general/speeches/detail/ who-director-general-s-opening-remarks-at-the-media-briefing-on-covid-19---11march-2020. Accessed 4 Oct 2021 2. Piret, J., Boivin, G.: Pandemic throughout history. Front. Microbiol. 11, 631736 (2021). https://doi.org/10.3389/fmicb.2020.631736 3. What is COVAX? All you need to know in 500 words. https://www.aljazeera.com/ news/2021/2/24/what-is-covax-all-you-need-to-know-in-500. Accessed 5 Oct 2021 4. COVID-19 vaccines. https://www.who.int/emergencies/diseases/novelcoronavirus-2019/covid-19-vaccines. Accessed 5 Oct 2021 5. Wallentin, G., Kaziyeva, D., Reibersdorfer-Adelsberger, E.: COVID-19 intervention scenarios for long-term disease management. Int. J. Health Policy Manag. 1(9), 508–516 (2020) 6. Shankar, S., Sukumaran, R., Patwa, P., Saxena, A.: Challenges of Equitable Vaccine Distribution in the COVID-19 Pandemic. ResearchGate. https://www. researchgate.net/publication/346356495. Accessed 5 Oct 2021 7. Ladu, M.: The role of city dashboards in managing public real estate in Italy: proposals for a conceptual framework. J. Urban Plan. Dev. 146(4), 04020047 (2020) 8. Django Web Framework. https://www.djangoproject.com/. Accessed 5 Oct 2021 9. Pedregosa, F., et al.: Scikit-learn: machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830 (2011) 10. Python: Applications for Python. https://www.python.org/about/apps/. Accessed 5 Oct 2021 11. Top 6 Python Libraries for Visualization: Which one to Use? https:// towardsdatascience.com/top-6-python-libraries-for-visualization-which-oneto-use-fe43381cd658. Accessed 6 Oct 2021 12. Introduction to Data Visualization in Python. https://towardsdatascience.com/ introduction-to-data-visualization-in-python-89a54c97fbed. Accessed 6 Oct 2021 13. Chapman, P., et al.: CRISP-DM 1.0: step-by-step data mining guide. SPSS Inc. 9, 13 (2000) 14. Sharda, R., Delen, D., Turban, E., Aronson, J.E., Liang, T.-P., King, D.: Business Intelligence and Analytics?: Systems for Decision Support. Pearson, Boston (2014) 15. Shafique, U., Qaiser, H.: A comparative study of data mining process models (KDD, CRISP-DM and SEMMA). Int. J. Innov. Sci. Res. 12(1), 217–222 (2014)

Author Index

A Abayomi, Abdultaofeek, 700 AbdulRaheem, Muyideen, 806 Abidi, Amna, 170 Abraham, Ajith, 130, 150, 226, 332, 491, 505 Abtahi, Amir, 786 Adam, Shuaibu Musa, 362 Adetiba, Emmanuel, 700 Afolabi, David, 362 Aftab, Muhammad, 675 Ahmed, Kharimah Bimbola, 830 Ajesh, F., 226 Akanni, Adeniyi, 842 Aklouf, Youcef, 412 Akram, Muhammad, 842 Alfa, Abraham Ayegba, 830 Amin, Rashid, 675 Amzil, Abdellah, 110 Anandamurugan, S., 261, 320 Anitha, N., 286, 307 Arogundade, Oluwasefunmi, 842 Aruleba, Kehinde, 467, 797 Aswanth, P., 278 Aswathy, S. U., 32, 55 Athithya, A., 278 Attah, Blessing Iganya, 830 Augusto, Maria Fernanda, 557 Avitabile, Gianfranco, 665 Awotunde, Joseph Bamidele, 806 Ayeni, Foluso, 818 Ayesha Howla, J., 320 Azevedo, Beatriz Flamia, 120

B B. Elvas, Luis, 251 Badhe, Tanishka, 65 Bajaj, Anu, 774 Baptista, José, 723 Barroso, João, 447, 457 Bastos, João A., 78 Begum, Arju Manara, 654 Bellouch, Abdessamad, 88, 110 Ben Mabrouk, Mouna, 98 Benammar, Nassima, 170 Bendimerad, Lydia Sonia, 402 Bernardino, Jorge, 205 Bharati, Subrato, 654 Bhattacherjee, Amrita, 332 Bnouni Rhim, Nesrine, 98 Boaventura, José, 723, 733 Boaventura-Cunha, José, 743 Borde, Janhavi, 65 Boujnoui, Ahmed, 88, 110 Brito, Marlene F., 78, 193 Butt, Shariq Aziz, 818

C Calcagno, Facundo M., 170 Cale, Daniel, 251 Canizes, Bruno, 751 Carvalho, Diana, 447, 457 Chandana, V., 24 Chaudhari, Anagha, 65 Chavez, Juliana, 751

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 A. Abraham et al. (Eds.): IBICA 2021, LNNS 419, pp. 863–866, 2022. https://doi.org/10.1007/978-3-030-96299-9

864

Author Index

Coelho, Duarte, 216, 371 Correia, Filipe, 205 Costa, Diana, 78 Costa, Domingos, 141 Coviello, Giuseppe, 665 Cupples, Shirin, 786

Helgheim, Berit Irene, 761 Honório, Tomás, 517, 527 Houacine, Naila Aziza, 402

D Dada, Oluwaseun Alexander, 467, 797 Damaševiˇcius, Robertas, 830 Dani, Virendra, 160 Das, Tanaya, 579 Deva Dharshini, B., 320 Devi, Kapila, 774 Dhlamini, Thabani, 425 Dias, João, 447 Dobrynin, Vladimir, 435 Dornberger, Rolf, 15 Drias, Habiba, 402

J Jain, Khushboo, 130 Janardhanan, K. A., 623 Jauhar, Sunil Kumar, 24 Jorge-Martinez, Diaz, 818

F Faizi, Rdouan, 549 Fernandes, Fábio, 371 Ferreira, Joao C., 251, 761 Ferreira, Judite, 723, 733, 743 Florio, Antonello, 665 Fonte, Bruno, 183 Fozooni, Nazanin, 505 G Gafeira, T., 590 Garjan, Hossein Shokri, 491, 505 Gatto, Bernardo Bentes, 239 Ghoreyshi, Kajal, 852 Goel, Neelam, 567 Gonçalves, Ramiro, 216 Goodarzian, Fariba, 491 Gopika, T., 278 Gowtham, S., 307 Graça, P., 590 Guarda, Teresa, 557 Gupta, Manali, 130 H Halmaoui, Houssam, 537 Hanini, Mohamed, 110 Hanne, Thomas, 15 Haqiq, Abdelkrim, 88, 537 Hassanien, Aboul Ella, 88 Hazarika, Ruhul Amin, 381

I Idris, Adamu, 362

K Kaarniha Shri, M., 307 Kahil, Haithem, 170 Kalaiyarasi, T., 307 Kamboj, Minakshi, 42 Kandar, Debdatta, 381 Kashyap, Aditya Singh, 685 Khoi, Bui Huy, 3 Koundal, Deepika, 675 Krähemann, Marc, 15 Krishnan, H. Muthu, 314 Kumar, M. Saravana, 261 L Lamacchia, Claudio Maria, 633 Leena Sri, R., 351 León, Marcelo, 713 León, Paulina, 713 Liu, Hongbo, 150 Loaiza, Vinicio, 713 Lufiya, George C., 55 M Mabrouk, Mouna Ben, 170 Madureira, Ana M., 78 Madureira, Ana, 205, 216, 251, 723, 733, 786 Maji, Arnab Kumar, 381 Majumdar, Arun Kumar, 579 Man, Ka Lok, 362 Martins, Ana Lucia, 761 Martins, Paulo, 447 Martins, S., 590 Maskeli¯unas, Rytis, 830 Mastroianni, Michele, 435 Mawela, Tendani, 425, 599 Maximiano, Marisa, 517, 527 Mescia, Luciano, 633 Mevoli, Gianvito, 633 Mienye, Ibomoiye Domor, 467

Author Index Misra, Sanjay, 806, 818, 830, 842 Mkhatshwa, Bonginkosi, 599 Molaei, Alireza Abbaszadeh, 491, 505 Mollinetti, Marco Antônio Florenzano, 239 Mondal, M. Rubaiyat Hossain, 654 Moreira, Joaquim, 183 Moyo, Sibusiso, 700 Mugde, Sareeta, 685 N Nagar, Sneha, 160 Nanjappan, Vijayakumar, 362 Neto, Aline, 183 Ngan, Nguyen Thi, 3 Nicola, S., 590 Nicola, Susana, 183 Nithin, K., 261 O Obaido, George, 467, 797 Obiyemi, Obiseye O., 700 Okoye, Kingsley, 475 Oladipo, Idowu Dauda, 806 Olasina, Jamiu R., 700 Oliveira, J., 590 Oliveira, Marco, 517, 527 Oluranti, Jonathan, 806, 830, 842 Opanuga, Temilade, 842 P Paredes, Hugo, 447 Pawar, Vishal, 160 Pereira, Ana I., 120 Pereira, Ivo, 216, 371 Podder, Prajoy, 654 Prajoon, P., 32 Prasanna Kumar, K. R., 270, 278 Prashanth, E. G., 261 Pravin, D., 270 Preetha, J., 314 Puga, Ricardo, 723, 733 Purkayastha, Roneeta, 643 Q Quadrado, José, 786 R Ramos, Ana L., 193 Ramos, Sérgio, 751 Rana, Sanjeev, 42 Rani, Shalli, 612

865 Ranjith, T., 320 Ratnoo, Saroj, 774 Rebelo, Miguel Ângelo, 371 Redroban, Carlos, 713 Reis, Catarina I., 517, 527 Rezaei, Elham, 852 Ribeiro, Pedro C., 193 Rocha, Tânia, 447, 457 Rokith Dhayal, N., 270 Roy, Abhishek, 579, 643 Roy, V. I., 623 S Sadique, Kazi Masum, 852 Saeed, Muhammad, 675 Salankar, Nilima, 391 Santos, André S., 78 Sanusi, Ismaila Temitayo, 797 Sanyal, Sugata, 332 Sathya, S., 270 Sentamilselvan, K., 296 Sequeira, A. H., 24 Sharma, Garima, 685 Sharma, Vagisha, 567 Sheveleva, Olga, 435 Shona, S. P., 314 Shruti, 612 Sivakami, A., 314 Soares, João, 751 Soundarajan, C., 286 Stephen, Divya, 32 Subaskaran, Arcsuta, 15 Sujitha, M., 296 Swastik, Swastika, 685 Swathi, V., 286 Swetha, J., 296 T Tamilselvan, M., 286 Teixeira, Otávio Noura, 239 Teixeira, Senhorinha, 141 Thakur, Surendra, 700 Thakur, Vaishnavi, 65 Thomas, Jyothi, 55 Tinubu, Oreoluwa, 842 Titah, Anas, 170 Tokkozhina, Ulpan, 761 V Vadivel, S. M., 24 Vale, Zita, 751 Varela, Leonilde R., 141 Varela, Maria Leonilde R., 120 Verma, Amandeep, 567 Vetriveeran, Divya, 351

866

Author Index

Vigasini, R., 296 Vincent, Bibin, 32

Y Yunusa, Abdullahi Abubakar, 797 Yusuf, Shamsuddeen, 362

W Waghmare, Bhagyashree, 65 Wang, Naiyao, 150 Wang, Yukun, 150

Z Zaaloul, Abdellah, 88, 110 Zebouchi, Ahmed, 412 Zhou, Changdong, 150