Machine Learning for Cyber Security. 4th International Conference, ML4CS 2022 Guangzhou, China, December 2–4, 2022 Proceedings, Part I 9783031200953, 9783031200960

Citation preview

LNCS 13655

Yuan Xu · Hongyang Yan · Huang Teng · Jun Cai · Jin Li (Eds.)

Machine Learning for Cyber Security 4th International Conference, ML4CS 2022 Guangzhou, China, December 2–4, 2022 Proceedings, Part I

Lecture Notes in Computer Science Founding Editors Gerhard Goos Karlsruhe Institute of Technology, Karlsruhe, Germany Juris Hartmanis Cornell University, Ithaca, NY, USA

Editorial Board Members Elisa Bertino Purdue University, West Lafayette, IN, USA Wen Gao Peking University, Beijing, China Bernhard Steffen TU Dortmund University, Dortmund, Germany Moti Yung Columbia University, New York, NY, USA

13655

More information about this series at https://link.springer.com/bookseries/558

Yuan Xu · Hongyang Yan · Huang Teng · Jun Cai · Jin Li (Eds.)

Machine Learning for Cyber Security 4th International Conference, ML4CS 2022 Guangzhou, China, December 2–4, 2022 Proceedings, Part I

Editors Yuan Xu School of Computing and Informatics University of Louisiana at Lafayette Lafayette, IN, USA Huang Teng Institute of Artificial Intelligence and Blockchain Guangzhou University Guangzhou, China

Hongyang Yan Institute of Artificial Intelligence and Blockchain Guangzhou University Guangzhou, China Jun Cai Guangdong Polytechnic Normal University Guangzhou, China

Jin Li Institute of Artificial Intelligence and Blockchain Guangzhou University Guangzhou, China

ISSN 0302-9743 ISSN 1611-3349 (electronic) Lecture Notes in Computer Science ISBN 978-3-031-20095-3 ISBN 978-3-031-20096-0 (eBook) https://doi.org/10.1007/978-3-031-20096-0 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 This work is subject to copyright. All rights are reserved 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

The Fourth International Conference on Machine Learning for Cyber Security (ML4CS 2022) was held in Guangzhou, China, during December 2–4, 2022. ML4CS is a wellrecognized annual international forum for AI-driven security researchers to exchange ideas and present their works. The conference received 367 submissions. Committee accepted 100 regular papers and 46 short papers to be included in the conference program. It was single blind during the paper review process, and there are two reviews per paper at least. The proceedings contain revised versions of the accepted papers. While revisions are expected to take the referees comments into account, this was not enforced and the authors bear full responsibility for the content of their papers. ML4CS 2022 was organized by Guangdong Polytechnic Normal University, Pazhou Lab, and Sun Yat-sen University. The conference would not have been such a success without the support of these organizations, and we sincerely thank them for their continued assistance and support. We would also like to thank the authors who submitted their papers to ML4CS 2022, and the conference attendees for their interest and support. We thank the Organizing Committee for their time and effort dedicated to arranging the conference. This allowed us to focus on the paper selection and deal with the scientific program. We thank the Program Committee members and the external reviewers for their hard work in reviewing the submissions; the conference would not have been possible without their expert reviews. Finally, we thank the EasyChair system and its operators, for making the entire process of managing the conference convenient. September 2022

Xiaochun Cao Jin Li Jun Cai Huang Teng Yan Jia Min Yang Xu Yuan

Organization

General Chairs Xiaochun Cao Jin Li Jun Cai Teng Huang

Sun Yat-sen University, China Guangzhou University, China Guangdong Polytechnic Normal University, China Guangzhou University, China

Program Chairs Yan Jia Min Yang Xu Yuan

Peng Cheng Laboratory, China Fudan University, China University of Louisiana at Lafayette, USA

Track Chairs Machine Learning Based Cybersecurity Track Wei Wang Yu-an Tan

Beijing Jiaotong University, China Beijing Institute of Technology, China

Big Data Analytics for Cybersecurity Track Xuyun Zhang Wenchao Jiang

Macquaire University, Australia Guangdong University of Technology, China

Cryptography in Machine Learning Track Xinyi Huang Joseph K. Liu

Fujian Normal University, China Monash University, Australia

Differential Privacy Track Changyu Dong Tianqing Zhu

Newcastle University, UK University of Technology Sydney, Australia

viii

Organization

Data Security in Machine Learning Track Zheli Liu Zuoyong Li

Nankai University, China Minjiang University, China

Adversarial Attacks and Defenses Track Qian Wang Kai Chen

Wuhan University, China Institute of Information Engineering, Chinese Academy of Sciences, China

Security and Privacy in Federated Learning Track Lianyong Qi Tong Li

Qufu Normal University, China Nankai University, China

Explainable Machine Learning Track Sheng Hong

Beihang University, China

Security in Machine Learning Application Track Tao Xiang Yilei Wang

Chongqing University, China Qufu Normal University, China

AI/Machine Learning Security and Application Track Hao Peng

Zhejiang Normal University, China

Workshop Chair Wei Gao

Yunnan Normal University, China

Publication Chair Di Wu

Guangzhou University, China

Publicity Chair Zhuo Ma

Xidian University, China

Organization

ix

Steering Committee Xiaofeng Chen Iqbal Gondal Ryan Ko Jonathan Oliver Islam Rafiqul Vijay Varadharajan Ian Welch Yang Xiang (Chair) Jun Zhang (Chair) Wanlei Zhou

Xidian University, China Federation University, Australia Waikato University, New Zealand Trend Micro, USA Charles Sturt University, Australia University of Newcastle, Australia Victoria University of Wellington, New Zealand Swinburne University of Technology, Australia Swinburne University of Technology, Australia Deakin University, Australia

Program Committee Silvio Barra M. Z. Alam Bhuiyan Carlo Blundo Yiqiao Cai Luigi Catuogno Lorenzo Cavallaro Liang Chang Fei Chen Xiaofeng Chen Zhe Chen Frédéric Cuppens Changyu Dong Guangjie Dong Mohammed EI-Abd Wei Gao Dieter Gollmann Zheng Gong Zhitao Guan Zhaolu Guo Jinguang Han Saeid Hosseini Chingfang Hsu Haibo Hu Teng Huang Xinyi Huang

University of Salerno, Italy Guangzhou University, China University of Salerno, Italy Huaqiao University, China University of Salerno, Italy King’s College London, UK Guilin University of Electronic Technology, China Shenzhen University, China Xidian University, China Singapore Management University, Singapore IMT Atlantique, France Newcastle University, UK East China Jiaotong University, China American University of Kuwait, Kuwait Yunnan Normal University, China Hamburg University of Technology, Germany South China Normal University, China North China Electric Power University, China Chinese Academy of Sciences, China Queen’s University Belfast, UK Singapore University of Technology and Design, Singapore Huazhong University of Science and Technology, China The Hong Kong Polytechnic University, Hong Kong Guangzhou University, China Fujian Normal University, China

x

Organization

Wenchao Jiang Lutful Karim Hadis Karimipour Sokratis Katsikas Neeraj Kumar Kangshun Li Ping Li Tong Li Wei Li Xuejun Li Kaitai Liang Hui Liu Wei Lu Xiaobo Ma Fabio Martinelli Ficco Massimo Weizhi Meng Vincenzo Moscato Francesco Palmieri Fei Peng Hu Peng Lizhi Peng Umberto Petrillo Lianyong Qi Shahryar Rahnamayan Khaled Riad Yu Sun Yu-An Tan Zhiyuan Tan Ming Tao Donghai Tian Chundong Wang Ding Wang Feng Wang Hui Wang Jianfeng Wang Jin Wang

Guangdong University of Technology, China Seneca College of Applied Arts and Technology, Canada University of Guelph, Canada Open University of Cyprus, Cyprus Thapar Institute of Engineering and Technology, India South China Agricultural University, China South China Normal University, China Naikai University, China Jiangxi University of Science and Technology, China Anhui University, China University of Surrey, UK University of Calgary, Canada Sun Yat-sen University, China Xi’an Jiaotong University, China IIT-CNR, Italy Second University of Naples, Italy Technical University of Denmark, Denmark University of Naples, Italy University of Salerno, Italy Hunan University, China Wuhan University, China Jinan University, China Sapienza University of Rome, Italy Qufu Normal University, China University of Ontario Institute of Technology, Canada Guangzhou University, China Guangxi University, China Beijing Institute of Technology, China Edinburgh Napier University, UK Dongguan University of Technology, China Beijing Institute of Technology, China Tianjin University of Technology, China Peking University, China Wuhan University, China Nanchang Institute of Technology, China Xidian University, China Soochow University, China

Organization

Licheng Wang Lingyu Wang Tianyin Wang Wei Wang Wenle Wang Sheng Wen Yang Xiang Run Xie Xiaolong Xu Li Yang Shao-Jun Yang Zhe Yang Yanqing Yao Xu Yuan Qikun Zhang Xiao Zhang Xiaosong Zhang Xuyun Zhang Yuan Zhang Xianfeng Zhao Lei Zhu Tianqing Zhu

Beijing University of Posts and Telecommunications, China Concordia University, Canada Luoyang Normal University, China Beijing Jiaotong University, China Jiangxi Normal University, China Swinburne University of Technology, Australia Swinburne University of Technology, Australia Yibin University, China Nanjing University of Information Science & Technology, China Xidian University, China Fujian Normal University, China Northwestern Polytechnical University, China Beihang University, China University of Louisiana at Lafayette, USA Beijing Institute of Technology, China Beihang University, China Tangshan University, China Macquarie University, Australia Nanjing University, China Chinese Academy of Sciences, China Huazhong University of Science and Technology, China China University of Geosciences, China

Track Program Committee - AI/Machine Learning Security and Application Hao Peng (Chair) Meng Cai Jianting Ning Hui Tian Fushao Jing Guangquan Xu Jun Shao

xi

Zhejiang Normal University, China Xi’an Jiaotong University, China Singapore Management University, Singapore Huaqiao University, China National University of Defense Technology, China Tianjin University, China Zhejiang Gongshang University, China

Contents – Part I

Traditional Chinese Medicine Health Status Identification with Graph Attention Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Amin Fu, Jishun Ma, Chuansheng Wang, Changen Zhou, Zuoyong Li, and Shenghua Teng Flexible Task Splitting Strategy in Aircraft Maintenance Technician Scheduling Based on Swarm Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Bowen Xue, Huifen Zhong, Junrui Lu, Tianwei Zhou, and Ben Niu Privacy Preserving CSI Fingerprint Device-Free Localization . . . . . . . . . . . . . . . . Tianxin Huang, Lingjun Zhao, Zeyang Dai, Liang Lin, and Huakun Huang

1

15

27

A Novel Blockchain-MEC-Based Near-Domain Medical Resource Sharing Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Haichao Wu, Xiaoming Liu, and Wei Ou

40

Pairwise Decomposition of Directed Graphic Models for Performing Amortized Approximate Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Peng Lin, Changsheng Dou, Nannan Gu, Zhiyuan Shi, and Lili Ma

57

VDDL: A Deep Learning-Based Vulnerability Detection Model for Smart Contracts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fan Jiang, Yuanlong Cao, Jianmao Xiao, Hui Yi, Gang Lei, Min Liu, Shuiguang Deng, and Hao Wang

72

Robust Remote Sensing Scene Classification with Multi-view Voting and Entropy Ranking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jinyang Wang, Tao Wang, Min Gan, and George Hadjichristofi

87

Visualized Analysis of the Emerging Trends of Automated Audio Description Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lingqian Zheng and Xinrong Cao

99

Anomaly Detection for Multi-time Series with Normalizing Flow . . . . . . . . . . . . 109 Weiye Ning, Xin Xie, Yuhui Huang, Si Yu, Zhao Li, and Hao Yang Encrypted Transmission Method of Network Speech Recognition Information Based on Big Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 Yanning Zhang

xiv

Contents – Part I

A Lightweight NFT Auction Protocol for Cross-chain Environment . . . . . . . . . . . 133 Hongyu Guo, Mao Chen, and Wei Ou A Multi-scale Framework for Out-of-Distribution Detection in Dermoscopic Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 Zhongzheng Huang, Tao Wang, Yuanzheng Cai, and Lingyu Liang Swarm Intelligence for Multi-objective Portfolio Optimization . . . . . . . . . . . . . . . 160 Li Chen, Yongjin Wang, Jia Liu, and Lijing Tan Research on Secure Cloud Storage of Regional Economic Data Network Based on Blockchain Technology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 Huiling Liu Data Leakage with Label Reconstruction in Distributed Learning Environments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 Xiaoxue Zhang, Xiuhua Zhou, and Kongyang Chen Analysis Method of Abnormal Traffic of Teaching Network in Higher Vocational Massive Open Online Course Based on Deep Convolutional Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198 Haiying Chen and Jiahui Zou Spatio-Temporal Context Modeling for Road Obstacle Detection . . . . . . . . . . . . . 213 Xiuen Wu, Tao Wang, Lingyu Liang, Zuoyong Li, and Fum Yew Ching A Survey of Android Malware Detection Based on Deep Learning . . . . . . . . . . . . 228 Dianxin Wang, Tian Chen, Zheng Zhang, and Nan Zhang Information Encryption Transmission Method of Automobile Communication Network Based on Neural Network . . . . . . . . . . . . . . . . . . . . . . . . 243 Chunhua Kong, Kai Ma, and Jiatong Wei Explanation-Guided Minimum Adversarial Attack . . . . . . . . . . . . . . . . . . . . . . . . . 257 Mingting Liu, Xiaozhang Liu, Anli Yan, Yuan Qi, and Wei Li CIFD: A Distance for Complex Intuitionistic Fuzzy Set . . . . . . . . . . . . . . . . . . . . . 271 Yangyang Zhao and Fuyuan Xiao Security Evaluation Method of Distance Education Network Nodes Based on Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 Jiajuan Fang MUEBA: A Multi-model System for Insider Threat Detection . . . . . . . . . . . . . . . 296 Jing Liu, Jingci Zhang, Changcun Du, and Dianxin Wang

Contents – Part I

xv

Bayesian Based Security Detection Method for Vehicle CAN Bus Network . . . . 311 Shen Jiang and Hailan Zhang Discrete Wavelet Transform-Based CNN for Breast Cancer Classification from Histopathology Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326 Yuan Qi, Xiaozhang Liu, Hua Li, Mingting Liu, and Wei Li Machine Learning Based Security Situation Awareness Method for Network Data Transmission Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341 Hui Du Multi-objective Hydrologic Cycle Optimization for Integrated Container Terminal Scheduling Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356 Ben Niu, Yuda Wang, Jia Liu, and Qianying Liu High Voltage Power Communication Network Security Early Warning and Monitoring System Based on HMAC Algorithm . . . . . . . . . . . . . . . . . . . . . . . 366 Zhengjian Duan Large Scale Network Intrusion Detection Model Based on FS Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380 Mei Hong, Yingyong Zou, and Chun Ai Research on Intelligent Detection Method of Automotive Network Data Security Based on FlexRay/CAN Gateway . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394 Jiatong Wei, Kai Ma, and Chunhua Kong Adversarial Attack and Defense on Natural Language Processing in Deep Learning: A Survey and Perspective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409 Huoyuan Dong, Jialiang Dong, Shuai Yuan, and Zhitao Guan A Novel Security Scheme for Mobile Healthcare in Digital Twin . . . . . . . . . . . . . 425 Nansen Wang, Wenbao Han, and Wei Ou Construction of Security Risk Prediction Model for Wireless Transmission of Multi Axis NC Machining Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442 Guoqiang Zhao, Meitao Zhang, and Yingying Wu Spiking Neural Networks Subject to Adversarial Attacks in Spiking Domain . . . 457 Xuanwei Lin, Chen Dong, Ximeng Liu, and Dong Cheng Diverse Web APIs Recommendation with Privacy-preservation for Mashup Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472 Shengqi Wu, Lianyong Qi, Yuwen Liu, Yihong Yang, Ying Miao, and Fei Dai

xvi

Contents – Part I

Network Security Evaluation Method of College Freshmen Career Counseling Service Based on Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 486 Shuilan Song and Xinjiu Liang FedTD: Efficiently Share Telemedicine Data with Federated Distillation Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501 Ning Li, Nansen Wang, Wei Ou, and Wenbao Han Increase Channel Attention Based on Unet++ Architecture for Medical Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 516 Fei Wu, Sikai Liu, Bo Li, and Jinghong Tang Distributed Power Load Missing Value Forecasting with Privacy Protection . . . . 521 Ying Miao, Lianyong Qi, Haoyang Wu, Yuxin Tian, Shengqi Wu, Yuqing Wang, Fei Dai, and Shaoqi Ding Differentially Private Generative Model with Ratio-Based Gradient Clipping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 535 Jianchen Lin and Yanqing Yao Differential Privacy Protection Algorithm for Data Clustering Center . . . . . . . . . 550 Mingyang Ma, Hongyong Yang, and Fei Liu Improved Kmeans Algorithm Based on Privacy Protection . . . . . . . . . . . . . . . . . . 560 Caixin Wang, Lili Wang, and Hongyong Yang Symmetry Structured Analysis Sparse Coding for Key Frame Extraction . . . . . . 568 Yujie Li, Benying Tan, Shuxue Ding, Christian Desrosiers, and Ahmad Chaddad Data Reconstruction from Gradient Updates in Federated Learning . . . . . . . . . . . 586 Xiaoxue Zhang, Junhao Li, Jianjie Zhang, Jijie Yan, Enmin Zhu, and Kongyang Chen Natural Backdoor Attacks on Speech Recognition Models . . . . . . . . . . . . . . . . . . . 597 Jinwen Xin, Xixiang Lyu, and Jing Ma Boarding Pass Positioning with Jointly Multi-channel Segmentation and Perspective Transformation Correction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 611 Jiahui Wu, Zuoyong Li, Pantea Keikhosrokiani, and Yuanzheng Cai AP-GCL: Adversarial Perturbation on Graph Contrastive Learning . . . . . . . . . . . 624 ZiYu Zheng, HaoRan Chen, and Ke Peng

Contents – Part I

xvii

An Overview of Opponent Modeling for Multi-agent Competition . . . . . . . . . . . . 634 Lu Liu, Jie Yang, Yaoyuan Zhang, Jingci Zhang, and Yuxi Ma Research on Potential Threat Identification Algorithm for Electric UAV Network Communication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649 Gebiao Hu, Zhichi Lin, Zheng Guo, Ruiqing Xu, and Xiao Zhang Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 665

Contents – Part II

AMAD: Improving Adversarial Robustness Without Reducing Accuracy . . . . . . Yujie Lin, Ximeng Liu, and Nan Jiang

1

Multi-party Secure Comparison of Strings Based on Outsourced Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xin Zhang, Chao Shan, and Yunfeng Zou

15

Highway: A Super Pipelined Parallel BFT Consensus Algorithm for Permissioned Blockchain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zui Luo, Chang Chen, and Wangjie Qiu

31

Overview of DDoS Attack Research Under SDN . . . . . . . . . . . . . . . . . . . . . . . . . . . Lei Guo, Shan Jing, and Chuan Zhao

39

A Medical Image Segmentation Method Based on Residual Network and Channel Attention Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sikai Liu, Fei Wu, Jinghong Tang, and Bo Li

54

Performance Improvement of Classification Model Based on Adversarial Sample Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Qian Jiang, Jie Kang, and Zhendong Wu

61

Research on Detection Method of Large-Scale Network Internal Attack Based on Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chang Liu, Chaozhong Long, Yuchuan Yu, and Ziqi Lin

75

Federated Community Detection in Social Networks . . . . . . . . . . . . . . . . . . . . . . . Zhiwei Zheng, Zekai Chen, Ximeng Liu, and Nan Jiang

89

A Textual Adversarial Attack Scheme for Domain-Specific Models . . . . . . . . . . . 104 Jialiang Dong, Shen Wang, Longfei Wu, Huoyuan Dong, and Zhitao Guan An Improved Conv-LSTM Method for Gear Fault Detection . . . . . . . . . . . . . . . . . 118 Yang Zhang, Jianwu Zhang, Guanhong Zhang, and Hong Li Extracting Random Secret Key Scheme for One-Time Pad Under Intelligent Connected Vehicle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 Junjie Chu, Mu Han, and Shidian Ma

xx

Contents – Part II

Semi-supervised Learning with Nearest-Neighbor Label and Consistency Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 Guolin Zheng, Zuoyong Li, Wenkai Hu, Haoyi Fan, Fum Yew Ching, Zhaochai Yu, and Kaizhi Chen Bipolar Picture Fuzzy Graph Based Multiple Attribute Decision Making Approach–Part I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 Shu Gong and Gang Hua Priv-IDS: A Privacy Protection and Intrusion Detection Framework for In-Vehicle Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 Simin Li and Mu Han Dynamic Momentum for Deep Learning with Differential Privacy . . . . . . . . . . . . 180 Guanbiao Lin, Hu Li, Yingying Zhang, Shiyu Peng, Yufeng Wang, Zhenxin Zhang, and Jin Li An Unsupervised Surface Anomaly Detection Method Based on Attention and ASPP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 Yuhui Huang, Xin Xie, Weiye Ning, Dengquan Wu, Zixi Li, and Hao Yang PCB Defect Detection Method Based on Improved RetinaNet . . . . . . . . . . . . . . . . 202 Yusheng Xu, Xinrong Cao, Rong Hu, Pantea Keikhosrokiani, and Zuoyong Li A Method of Protecting Sensitive Information in Intangible Cultural Heritage Communication Network Based on Machine Learning . . . . . . . . . . . . . . 214 Xiaoyu Zhang and Ye Jin Decision Making Analysis of Traffic Accidents on Mountain Roads in Yunnan Province . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228 Shu Gong and Gang Hua Deep Adaptively Feature Extracting Network for Cervical Squamous Lesion Cell Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238 Zhonghua Peng, Rong Hu, Fuen Wang, Haoyi Fan, Yee Wei Eng, Zuoyong Li, and Liwei Zhou DSGRAE: Deep Sparse Graph Regularized Autoencoder for Anomaly Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254 Shicheng Li, Xiaoguo Yang, Haoming Zhang, Chaoyu Zheng, and Yugen Yi A Lattice-Based Aggregate Signature Based on Revocable Identity . . . . . . . . . . . 266 Yang Cui, Huayu Cheng, Fengyin Li, and Domenico Santaniello

Contents – Part II

xxi

Research and Design of an Emergency Supply Assurance Monitoring System in the Post-epidemic Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275 Yongbo Li Face Presentation Attack Detection Based on Texture Gradient Enhancement and Multi-scale Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 Fei Peng, Shao-hua Meng, and Min Long Optimal Revenue Analysis of the Stubborn Mining Based on Markov Decision Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299 Yiting Zhang, Ming Liu, Jianan Guo, Zhaojie Wang, Yilei Wang, Tiancai Liang, and Sunil Kumar Singh Bipolar Picture Fuzzy Graph Based Multiple Attribute Decision Making Approach-Part II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 Shu Gong, Gang Hua, and Xiaomei Zhang Machine Learning Based Method for Quantifying the Security Situation of Wireless Data Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321 Jie Xu Overlapping Community Discovery Algorithm Based on Three-Level Neighbor Node Influence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335 Shishuang Chen, Guanru Huang, Sui Lin, Wenchao Jiang, and Zhiming Zhao Content-Aware Deep Feature Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 Weice Wang, Zuoyong Li, Xiangpan Zheng, Taotao Lai, and Pantea Keikhosrokiani F2DLNet: A Face Forgery Detection and Localization Network Based on SSIM Error Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355 Fei Peng, Xin-lin Zhang, and Min Long An Eye-Gaze Tracking Method Based on a 3D Ocular Surface Fitting Model . . 370 Zhang Ling, Ma Zongxin, Yan MingYu, Jiang Wenchao, and Muhammad A Certificateless-Based Blind Signature Scheme with Message Recovery . . . . . . 382 Xiao Li, Mengwen Wang, and Fengyin Li Fault Detection of Rolling Bearings by Using a Combination Network Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390 Tingting Chen, Guanhong Zhang, and Tong Wu

xxii

Contents – Part II

zkChain: An Efficient Blockchain Privacy Protection Scheme Based on zk-SNARKs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400 Jiahui Huang, Teng Huang, and Jiehua Zhang Research on Influential Factors of Online Learning Behavior Based on Big Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411 Fangqin Ma and JunHan Qiu Short Speech Key Generation Technology Based on Deep Learning . . . . . . . . . . . 422 Zhengyin Lv, Zhendong Wu, and Juan Chen Domain Adversarial Interaction Network for Cross-Domain Fault Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 436 Weikai Lu, Jian Chen, Hao Zheng, Haoyi Fan, Eng Yee Wei, Xinrong Cao, and Deyang Zhang A Vehicle Data Publishing System with Privacy-Awares in VANETs Based on Blockchain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447 Xueqing Sun, Xiao Li, and Fengyin Li Highsimb: A Concrete Blockchain High Simulation with Contract Vulnerability Detection for Ethereum and Hyperledger Fabric . . . . . . . . . . . . . . . . 455 Pengfei Huang, Wanqing Jie, Arthur Sandor Voundi Koe, Ruitao Hou, Hongyang Yan, Mourad Nouioua, Phan Duc Thien, Jacques Mbous Ikong, and Camara Lancine Research on Key Technologies for the Trusted Perception of Network Information for Big Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469 Yuxiang Li and Fakariah Hani Mohd Ali Micro-expression Recognition Method Combining Dual-Stream Convolution and Capsule Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 479 Lanwei Zeng, Yudong Wang, Chang Zhu, Wenchao Jiang, and Jiaxing Li Security Scheduling Method of Cloud Network Big Data Cluster Based on Association Rule Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495 Teng Peng and Xiaohong Wang Towards Differentially Private Contrastive Learning . . . . . . . . . . . . . . . . . . . . . . . . 510 Wenjun Li, Anli Yan, Taoyu Zhu, Teng Huang, Xuandi Luo, and Shaowei Wang Two-Stage High Precision Membership Inference Attack . . . . . . . . . . . . . . . . . . . . 521 Shi Chen and Yubin Zhong

Contents – Part II

xxiii

Secure Storage Method for Network Resources of Professional Works Based on Decision Tree Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536 Bomei Tan and Rong Yu Vehicle CAN Network Intrusion Detection Model Based on Extreme Learning Machine and Feature Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550 Jiaoxing Xu A Broad Learning System Based on the Idea of Vertical Federated Learning . . . 565 Junrong Ge, Xiaojiao Wang, Fengyin Li, and Akshat Gaurav PAMP: A New Atomic Multi-Path Payments Method with Higher Routing Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575 Jianan Guo, Lei Shang, Yilei Wang, Tiancai Liang, Zhaojie Wang, and Hui An Privacy-Preserving Searchable Encryption Scheme Based on Deep Structured Semantic Model over Cloud Application . . . . . . . . . . . . . . . . . . . . . . . . 584 Na Wang, Jian Jiao, Shangcheng Zhang, Jianwei Liu, Kaifa Zheng, Junsong Fu, and Jiawen Qiao A Event Extraction Method of Document-Level Based on the Self-attention Mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 609 Xueming Qiao, Yao Tang, Yanhong Liu, Maomao Su, Chao Wang, Yansheng Fu, Xiaofang Li, Mingrui Wu, Qiang Fu, and Dongjie Zhu Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621

Contents – Part III

Design of Active Defense System for Railway Communication Network Based on Deep Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zhenguo Wu

1

Human Resource Network Information Recommendation Method Based on Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Xiao Wang

15

Machine Learning Based Abnormal Flow Analysis of University Course Teaching Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shaobao Xu and Yongqi Jia

30

A Learned Multi-objective Bacterial Foraging Optimization Algorithm with Continuous Deep Q-Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tianwei Zhou, Wenwen Zhang, Pengcheng He, and Guanghui Yue

44

Webpage Text Detection Based on Improved Faster-RCNN Model . . . . . . . . . . . . Junling Gao, Moran Zhao, Jianchao Wang, Yijin Zhao, Lifang Ma, and Pingping Yu

54

Turbo: A High-Performance and Secure Off-Chain Payment Hub . . . . . . . . . . . . . Jinchun He, Wangjie Qiu, Rixin He, Shengda Zhuo, and Wanqing Jie

67

A Complete Information Detection Method for Vehicle CAN Network Gateway Based on Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Shen Jiang and Hailan Zhang

76

Optimization of Data Transmission Efficiency of Wireless Communication Network for Chemical Energy Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Chengmi Xiang

91

Deep Spatio-Temporal Decision Fusion Network for Facial Expression Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 Xuanchi Chen, Heng Yang, Xia Zhang, Xiangwei Zheng, and Wei Li A Tabu-Based Multi-objective Particle Swarm Optimization for Irregular Flight Recovery Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Tianwei Zhou, Yichen Lai, Xiaojie Huang, Xumin Chen, and Huifen Zhong

xxvi

Contents – Part III

A Robot Foreign Object Inspection Algorithm for Transmission Line Based on Improved YOLOv5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 Zhenzhou Wang, Xiaoyue Xie, Xiang Wang, Yijin Zhao, Lifang Ma, and Pingping Yu Path Planning Algorithm Based on A_star Algorithm and Q-Learning Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 Xiaodong Zhao, Mengying Cao, Jingfang Su, Yijin Zhao, Shuying Liu, and Pingping Yu Belief χ 2 Divergence-Based Dynamical Complexity Analysis for Biological Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158 Lang Zhang and Fuyuan Xiao Local Feature Acquisition Method of Multi-layer Vision Network Image Based on Virtual Reality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 Jinzhu Liu and Shuai Zheng Channel Selection for EEG Emotion Recognition via an Enhanced Firefly Algorithm with Brightness-Distance Attraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 Ben Niu, Gemin Liang, Bang Tao, Chao Fu, Shuang Geng, Yang Wang, and Bowen Xue Security Risk Assessment Method of High Voltage Power Communication Network Based on Fuzzy Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 Zhengjian Duan and Xingguo Li Trimodal Fusion Network Combined Global-Local Feature Extraction Strategy and Spatial-Frequency Fusion Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 Danyang Yao, Jinyu Wen, Amei Chen, Meie Fang, Xinhua Wei, and Zhigeng Pan Data Security Risk Prediction of Labor Relationship Rights Protection Network Platform Based on Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227 Min Yu Research on LSTM Based Traffic Flow Prediction Adaptive Beacon Transmission Period and Power Joint Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242 Botao Tu, Guanxiang Yin, Guoqing Zhong, Nan Jiang, and Yuejin Zhang Construction of Color Network Model of Folk Painting Based on Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 Rong Yu and Bomei Tan

Contents – Part III

xxvii

Research on Intrusion Prevention Optimization Algorithm of Power UAV Network Communication Based on Artificial Intelligence . . . . . . . . . . . . . . . . . . . 265 Gebiao Hu, Zhichi Lin, Zheng Guo, Ruiqing Xu, and Xiao Zhang Design of Network Big Data Anti Attack System for Carbon Emission Measurement Based on Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279 Sida Zheng, Shuang Ren, Jun Wang, Chang Wang, and Yaoyu Wang An Efficient Particle YOLO Detector for Urine Sediment Detection . . . . . . . . . . . 294 Zejian Chen, Rong Hu, Fukun Chen, Haoyi Fan, Fum Yew Ching, Zuoyong Li, and Shimei Su Evolutionary Factor-Driven Concise Bacterial Foraging Optimization Algorithm for Solving Customer Clustering Problems . . . . . . . . . . . . . . . . . . . . . . 309 Lijing Tan, Kuangxuan Qing, Chen Guo, and Ben Niu Brain Storm Optimization Algorithm with Multiple Generation Strategies for Patient Data Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321 Chen Guo, Xikun Liu, and Keqin Yao TGPFM: An Optimized Framework for Ordering and Transporting Raw Materials for Production . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334 Dongni Hu, Wenjun Li, Yada Yu, Junhao Li, and Hongyang Yan Visual Analysis of Facial Expression Recognition Research Based on Knowledge Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350 Yuan Bo, Fan Jiajia, and Jin Zhuang Brainstorming-Based Large Scale Neighborhood Search for Vehicle Routing with Real Travel Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358 Jia Liu, Nanqing Guo, and Bowen Xue A SAR Image Preprocessing Algorithm Based on Improved Homomorphic Wavelet Transform and Retinex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371 Jianchao Wang, Hainan Cheng, Xiang Wang, Yijin Zhao, Shuying Liu, and Pingping Yu Medical Data Clustering Based on Multi-objective Clustering Algorithm . . . . . . 385 Shilian Chen, Yingsi Tan, Junkai Guo, Yuqin He, and Shuang Geng Self-supervised Visual-Semantic Embedding Network Based on Local Label Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400 Zhukai Jiang and Zhichao Lian

xxviii

Contents – Part III

A New Deep Network Model for Stock Price Prediction . . . . . . . . . . . . . . . . . . . . 413 Min Liu, Hui Sheng, Ningyi Zhang, Yu Chen, and Longjun Huang A Method for Residual Network Image Classification with Multi-scale Feature Fusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427 Guo Ru, Peng Sheng, Anyang Tong, and Zhenyuan Li Face Morphing Detection Based on a Two-Stream Network with Channel Attention and Residual of Multiple Color Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 439 Min Long, Cheng-kun Jia, and Fei Peng DU-Net: A Novel Architecture for Retinal Vessels Segmentation . . . . . . . . . . . . . 455 Yan Jiang, Ziji Zeng, Lingxia Chen, Jiyong Hu, and Ping Li Sub-pixel Level Edge Extraction Technology for Industrial Parts for Smart Manufacturing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469 Bowen Zhang and Yingjie Liu USDSE: A Novel Method to Improve Service Reputation Based on Double-Side Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 484 Jianmao Xiao, Jia Zeng, Xu Miao, Yuanlong Cao, Jing Zhao, and Zhiyong Feng Morphology-Based Soft Label Smoothing Strategy for Fine-Grained Domain Adaptationming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499 Kangshun Li, Yi Wang, Tian Feng, Hassan Jalil, and Huabei Nie MOOC Performance Prediction and Online Design Instructional Suggestions Based on LightGBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507 Yimin Ren, Jun Wang, Jia Hao, Jianhou Gan, and Ken Chen MOOC Dropout Prediction Based on Bayesian Network . . . . . . . . . . . . . . . . . . . . 520 Shuang Shi, Shu Zhang, Jia Hao, Ken Chen, and Jun Wang Knowledge Enhanced BERT Based on Corpus Associate Generation . . . . . . . . . . 533 Lu Jiarong, Xiao Hong, Jiang Wenchao, Yang Jianren, and Wang Tao Multi-objective Particle Swarm Optimization Based on Archive Control Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548 Meilan Yang, Fei Chen, Qian Zhang, Jie Yang, Xiaoli Shu, and Yanmin Liu A Hybrid Multi-objective Genetic-Particle Swarm Optimization Algorithm for Airline Crew Rostering Problem with Fairness and Satisfaction . . . . . . . . . . . 563 Tianwei Zhou, Xuanru Chen, Xusheng Wu, and Chen Yang

Contents – Part III

xxix

Plant Leaf Area Measurement Using 3D Imaging: A Comparative Study Between Dynamic Structured Light Stereo and Time-of-Flight . . . . . . . . . . . . . . . 576 Yi Wang and KangShun Li Subgraph Matching Based on Path Adaptation for Large-Scale Graph . . . . . . . . . 585 Xinmiao Hu, Sui Lin, Guangsi Xiong, and Wenchao Jiang Photovoltaic Panel Intelligent Management and Identification Detection System Based on YOLOv5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595 Xueming Qiao, Dan Guo, Yuwen Li, Qi Xu, Baoning Gong, Yansheng Fu, Rongning Qu, Jingyuan Tan, Hongwei Zhao, and Dongjie Zhu Image Encryption Algorithm Based on a New Five-Dimensional Lorenz System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607 Xiujun Zhang, Xiaohong Zha, and Ke Luo Research on the Construction of an Accurate Procurement System for Library e-Resources in Foreign Language Under Dig Data Analysis . . . . . . . 619 Ke Luo and Xiujun Zhang Recognition Method of Wa Language Isolated Words Based on Convolutional Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 631 Jinsheng Liu, Jianhou Gan, Ken Chen, Di Wu, and Wenlin Pan Binding Number and Fractional (k, m)-Covered Graph . . . . . . . . . . . . . . . . . . . . . . 641 Linli Zhu, Yu Pan, and Wei Gao Domination Based Federated Learning Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 655 Yikuan Chen, Liang Li, and Wei Gao Structured Representation of Fuzzy Data by Bipolar Fuzzy Hypergraphs . . . . . . 663 Juanjuan Lu, Linli Zhu, and Wei Gao Deep Knowledge Tracing with GRU and Learning State Enhancement . . . . . . . . 677 Xiaoyu Han, Shu Zhang, Juxiang Zhou, Zijie Li, and Jun Wang Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687

Traditional Chinese Medicine Health Status Identification with Graph Attention Network Amin Fu1 , Jishun Ma1 , Chuansheng Wang2 , Changen Zhou3 , Zuoyong Li4(B) , and Shenghua Teng1(B) 1

2

College of Electronic and Information Engineering, Shandong University of Science and Technology, Qingdao 266590, China [email protected] Department of Automatic Control Technical, Polytechnic University of Catalonia, Barcelona 08034, Spain 3 Research Base of Traditional Chinese Medicine Syndrome, Fujian University of Traditional Chinese Medicine, Fuzhou 350122, China 4 Fujian Provincial Key Laboratory of Information Processing and Intelligent Control, College of Computer and Control Engineering, Minjiang University, Fuzhou 350121, China [email protected]

Abstract. The Traditional Chinese Medicine Health Status Identification plays an important role in TCM diagnosis and prescription recommendation. In this paper, we propose a method of Status Identification via Graph Attention Network, named SIGAT, which captures the complex medical correlation in the symptom-syndrome graph. More specifically, we construct a symptom-syndrome graph in that symptoms are taken as nodes and the edges are connected by syndromes. And we realize automatic induction of symptom to state element classification by using the attention mechanism and perceptron classifier. Finally, we conduct experiments by using hamming loss, coverage, 0/1 error, ranking loss, average precision, macro-F1 score, and micro-F1 score as evaluation metrics. The results demonstrate that the SIGAT model outperforms comparison algorithms on Traditional Chinese Medicine Prescription Dictionary dataset. The case study results suggest that the proposed method is a valuable way to identify the state element. The application of the graph attention network classification algorithm in TCM health status identification is of high precision and methodological feasibility. Keywords: Graph attention network TCM health status identification

1

· Multi-label classification ·

Introduction

With the development of emerging medical technology, people’s demand for medical treatment has translated into health demand. Since China has gradually c The Author(s), under exclusive license to Springer Nature Switzerland AG 2023  Y. Xu et al. (Eds.): ML4CS 2022, LNCS 13655, pp. 1–14, 2023. https://doi.org/10.1007/978-3-031-20096-0_1

2

A. Fu et al.

entered the center of the world stage, TCM has received national attention and benefits people all over the world [18]. The rapid development of big data and artificial intelligence in recent years has made traditional Chinese medicine more intelligent [7]. The theory of TCM state identification effectively promotes clinical treatment, and the induction of state elements is the key to TCM diagnosis and treatment. The state element is a generalization of the local or overall functions and conditions of the human body, and is the key to grasping the state of human health [1]. The theory of health state identification is based on the theory of Traditional Chinese Medicine. It is a process that identifies the location, nature, degree, and other state elements of the disease according to the fourdiagnosis information such as observation, auscultation and olfaction, inquiry, and pulse feeling and palpation [6,20]. State elements can be divided into two types: disease location (such as stomach, spleen, etc.) and disease (such as heat, cold, etc.) [5]. The essence of the state element is to summarize the characteristics of the human body’s Yin and Yang conflict state under specific conditions, including the physiological and pathological characteristics of the disease, the type of cold and heat deficiency and excess of the disease, and the environment’s cold, heat, summer and dampness characteristics [1]. Traditional status identification methods aim to classify state elements based on symptoms. There are two types of state elements induction methods. One is based on traditional machine learning methods, such as multi-label learning with label-specific features (LIFT) based on generic attributes, multi-label learning K nearest neighbor (ML-KNN), rank support vector machines (Rank SVM), and other methods [15,17]. The second is the deep learning method based on the graph neural network. Since the dialectical relationship among TCM symptoms, state elements, and syndromes can be concretized through graph representation learning, this paper uses the graph attention network to realize the task of state element induction. In this paper, we propose a graph attention network-based method for TCM health status identification, named SIGAT. More specifically, SIGAT captures complex medical correlations between nodes and edges for graph representation learning. We construct a symptom-syndrome graph in this model by using a graph attention network and a perceptron classifier to inductive state elements. Then we conduct comparative experiments based on several different machine learning methods and demonstrate that SIGAT achieves higher state element multi-label classification precision, laying a foundation for further research. The contributions of this paper are as follows: – We develop a graph representation based on Traditional Chinese Medicine health status identification, where symptoms are regarded as nodes and syndromes are regarded as edges in a symptom-syndrome graph. – We propose a graph attention network framework for Traditional Chinese Medicine health status identification, named the SIGAT model. Different from previous methods, we build a graph attention mechanism architecture along with a perceptron classifier for multi-label classification.

TCM Health Status Identification with Graph Attention Network

3

– We conduct extensive experiments on real-world datasets and demonstrate the feasibility and effectiveness of the proposed SIGAT compared with other machine learning methods.

2

Related Work

The state element induction work based on status identification theory can be defined as single-label classification and multi-label classification in the field of artificial intelligence [16]. The accuracy of single-label classification is low due to the rough abstraction of the features of TCM symptoms, in this paper, the status factor induction task is defined as a multi-label classification problem. The features of the symptom’s characterization parameters are used as input, and the state element is used as the output label of the model. In addition, every prescription record has one or more status elements, so it is more reasonable to define the task as a multi-label classification problem. In recent years, the task of multi-label text classification originated from deep learning techniques of text classification. And large numbers of graph neural network embedding techniques have been proposed for multi-label text classification tasks [3]. The Graph Attention Network proposed by Veliˇckovi´c et al. [14] introduces the attention mechanism into the spatial domain-based graph neural network [8]. The network is trained by aggregating the features of neighbor nodes, and the feature representation of the central node is updated, which improves the generalization ability of the model [14]. Through the attention mechanism, the graph attention network enables each neighbor node to learn different attention weight parameters and achieves a more efficient feature extraction ability by achieving effective aggregation of neighbor nodes [11]. Attention mechanisms have been shown to be effective in many domains [9]. In the task of state element classification, not all symptoms and syndromes contribute to the classification results, and the attention mechanism enables different symptoms to play a moderating role in the final node representation, thereby improving the classification accuracy.

3

Methods

In this section, we introduce the graph attention network for TCM health status identification called SIGAT, as shown in Fig. 1. The SIGAT model is mainly divided into three modules: TCM graph construction module, attention layer construction module, and state element prediction module. 3.1

TCM Graph Constrcuction Module

Firstly, we introduce the graph construction based on TCM state identification knowledge. We define the number of TCM prescriptions as M , and define the symptom set as V , where the number of symptoms is N . Let E stand for syndrome. Y is the set of state elements, and the number of state elements is L. For

4

A. Fu et al.

G

…

…… Leaky ReLU

∑ …

…

…

TCM Graph Construction

…

Output Sigmoid

Cough, Chilly Fever, Retch

Z

Leaky ReLU

MLP Classifier

Input

Lung Stomach Surface Cold Hot

Graph Attention Layer Attention Layer Construction

State Element Prediction

Fig. 1. The framework of the proposed SIGAT model.

every TCM prescription, symptom, syndrome, and state element are multi-hot coded into a vector. As shown in Fig. 2, the graph data we constructed is a series of interconnected nodes representing symptoms, and the edge relationship of the TCM syndrome type structure contains the TCM state identification theory. In the symptom-syndrome graph, the symptom node correlation is represented by the interconnected edge relation 1, that is, the two symptoms exist in the same syndrome. If the unconnected edge relationship between two nodes is 0, it indicates that these are two unrelated symptom nodes. The solid line in the graph is the symptom with solar typhoid syndrome as the edge, and the connection of the dotted line is due to other syndromes. We define a graph structure G = (V, E), where the set of symptoms nodes is V , and the edge set of syndromes is E. Every symptom node contains unique features, F is the initial feature of the node. X = {x1 , x2 , . . . , xN } is the multi-hot vector of prescription symptoms.

Sweating Headache Deaf

Bitter

Cough Chill

Pulse floating

Fever Solar Typhoid Syndrome

Retch

Ventosity

Fig. 2. Symptom-syndrome graph for TCM state identification.

TCM Health Status Identification with Graph Attention Network

3.2

5

Attention Layer Construction Module

The main idea of the graph attention network is that the generation of any node feature in the graph needs to pay attention to the contribution degree of its different nodes. After learning the correlation between this central node and other adjacent nodes, we assign different weight coefficients to all nodes [14]. This section first introduces a single graph attention layer and then stacks multiple graph attention layers to construct the SIGAT model. Calculation of Attention Coefficient. Let the input of the graph attention layer be the feature H of all nodes, where N is the total number of nodes (symptoms). The goal of this layer is to aggregate neighbour information of nodes to obtain new node feature representation [14]. Firstly, the attention coefficient of a single node i is calculated. Assuming that the central node i has an adjacent point j ∈ Ni , which are W Hi and W Hj respectively after linear transformation, the attention coefficient of the adjacent node j to the node i is: eij = a ([W Hi W Hj ]) , j ∈ Ni .

(1)

W is a shared parameter of linear mapping, which enhances the features of nodes, [··] concatenates the transformed features of nodes i and j. Here, a(·) stacks the concatenated higher-dimensional features to a real number. Obviously, through the learnable parameter W and mapping a(·), eij represents the importance of adjacent node j to the new features of central node i, and the correlation between nodes i and j is learned. When node i has multiple adjacent points, we normalize the correlation coefficient to avoid the problem that the attention coefficient is too large and is not conducive to training [10]. Meanwhile, in order to generalize the fitting ability of the model, the nonlinear activation function can be added to the linearly changed value as LeakyReLU (·), and the final attention coefficient calculation formula is: exp(eij ) . (2) αij = LeakyReLU (eij )  k∈Ni exp(eik ) Then, the weighted summation of neighbour node features is performed to obtain new node features:  Hi = σ( αij W Hj ). (3) j∈Ni

Hi is the new feature output by the graph attention layer that incorporates neighbourhood information for every node i, and σ(·) is the activation function. As described above, the features of a single layer of graph attention layer can be obtained. Multiple Attention Mechanism. Using k independent attention mechanisms, feature information of different subspaces can be obtained. In every graph attention layer, K independent transformation matrices W k are calculated. We

6

A. Fu et al.

aggregate the attention information of multiple subspaces, which is calculated in parallel according to Eq. (3), and then we concatenate the obtained results again: k 1   (k) (k) αij W Hj ), (4) Hi = σ( k i=1 j∈Ni

(k) αij

is the weight coefficient calculated by the K transformation matrix where  W k , which cause Hi to have higher dimensions (1, kd ). So it can only be used as a middle layer and not an output layer. For the output layer, one aggregation  approach is the weighted average of each attention mechanism H . Attention Layer Polymerization. After obtaining the attention coefficients of the single-layer symptom node and its adjacent nodes above, the model aggregates multiple graph attention layers to obtain the final node embedding representation:  (l) (l) (l−1) αij W (l) Hj )(1 < l < P ), (5) Hi = σ( j∈Ni

where l represents the number of network layers and the feature H of the first layer is the initial feature H (0) = F . The graph attention layer aggregates the features of symptom neighbour nodes to the central symptom nodes, and it learns new node feature expressions by using the nodes connection strategy based on the graph. Finally, the embedding of a symptom group is defined as Z. The fusion of all symptoms in a symptom group is expressed as the following formula: (P )

Z = σ(XHi 3.3

).

(6)

State Element Prediction Module

After using the graph attention mechanism to obtain the graph embedding representation Z of symptoms and syndromes, the perceptron classifier and sigmoid function are set in the output layer, and every neuron corresponds to a label: y = Sigmoid(W Z),

(7)

where, W is the trainable weight parameter, and Sigmoid(·) is the activation function that converts the output value into probability. The cross-entropy loss function [2,4] is used in network training as follows: L

1 (yi log2 (yi ) + (1 − yi ) log2 (1 − yi ). Loss = − L i=1

(8)

There, yi ∈ {0, 1} represents the true value of prescription sample, yi ∈ [0, 1] is the predicted probability value. L is the number of state elements.

TCM Health Status Identification with Graph Attention Network

4

7

Experiments

In this section, we introduce the experimental setups and analyze the experimental results. 4.1

Experimental Setup

Dataset. The data set used in this paper is collated from the Large Dictionary of Prescriptions. The dataset includes 2685 diagnosis and treatment prescriptions, 298 symptoms, 69 status elements, and 747 syndromes. The information statistics and examples are shown in Table 1. Table 1. Dataste information statistics. Traditional Chinese Medicine Prescription Dictionary

Counts

Prescription

2685

Symptoms

Syndrome

Dajianzhong Decoction

Yiqi sanfeng Decoction

Shortness of breath, Heaviness in waist, Hot flash, Aversion to wind, Asthma, Adiapneustia, Cough Headache Spleen and stomach Superficies tightened Yang deficiency By wind cold

Yang deficiency, State Elements Stomach, Spleen, Cold

Qi deficiency, Kidney, Wet, Hot

298

747

69

Implementation. The implementation of our method is based on the deep learning framework of PyTorch. All experiments are run on NVIDIA 1650Ti GPU with 4 GB memory and a Windows 10 operating system. Adaptive moment estimation (Adam) is used to minimize the final objective function with a learning rate of 1.5 × 10−4 , batch size of 8, and weight decay of 1 × 10−8 . Comparison Algorithm. The following machine learning methods are selected to compare with the proposed SIGAT: (1) Support Vector Machines (SVM): The kernel function of the support vector machine is set to a gaussian kernel, the kernel coefficient is 0.3, and the penalty coefficient of the relaxation coefficient is set to 3. (2) Bayesian Network (NB): The naive bayes classifier of the multivariate bernoulli model binarizes (maps to boolean values) the threshold of sample features, with the smoothing parameter set to 1.

8

A. Fu et al.

(3) K Nearest Neighbors (KNN): The number of neighbours used in the query is 10, and the nearest 10 points are selected with the size of the ball tree and KD tree constructed to 3. (4) Random Forests (RF): We set the number of decision trees in the forest is 10, and the minimum number of samples needed to divide internal nodes is 2. The minimum number of samples needed to be on leaf nodes is set to 1. Evaluation Metrics. Performance evaluation for multi-label classification is complicated, and the prediction result of each instance is a set of labels. We define the total number of prescription samples in the dataset of the Traditional Chinese Medicine Prescription Dictionary as M . The total state element label set is defined as Y , and the total number of categories is defined as L. For any sample xi , Yi represents the predicted label set, and Yi represents the true label set corresponding to the sample. And ri (λ) stands for the ranking position of label λ among all predicted labels. In this paper, we employ 7 different metrics, which are widely used to evaluate the performance of classifiers. Hamming loss (HLoss) is used to measure the number of misclassified labels. Δ calculates the commission errors and omission errors between the true labels and predicted results. Hamming loss represents the proportion of wrong samples in all labels [12], which is defined as follows: M 1 1 |Yi ΔYi |. HLoss = M i=1 L

(9)

Coverage indicates that the prediction results of the state element label are sorted from probability and the average value of the probability values requires to cover the true labels [19]. And maxλ∈Yi ri (λ) is the ranking of true labels among the predicted labels. Similarly, a small value indicates high network prediction performance. The calculation formula is as follows: Coverage =

M 1  max ri (λ) − 1. M i=1 λ∈Yi

(10)

0/1 error indicates the number of irrelevant labels with the highest ranking in the ranking labels of the prediction results [13]. In prescription sample xi , argmax(ri (λ)) is the top-ranked label. When the prediction label λ is not in Yi , the prediction result is wrong, then δ(λ) = 1. In the same way, when the prediction is correct, δ(λ) = 0. A small 0/1 error indicates better network prediction performance as follows: M 1  δ(argmax(ri (λ)), λ ∈ / Yi ). 0/1 Error = M i=1

(11)

Ranking loss (RLoss) is the number that the probability ranking ri (λb ) of irrelevant label λb is higher than the probability ri (λa ) of the related label λa .

TCM Health Status Identification with Graph Attention Network

9

Yi is the relevant label set in the true prescription, but Y¯i is the label that is not relevant to the prescription [13]. The low the ranking loss, the few unrelated labels are predicted to the front, which means the higher the prediction accuracy of the true relevant labels as follows: RLoss =

M 1 1  |(λa , λb ) : ri (λa ) > ri (λb ), (λa , λb ) ∈ Yi × Y¯i |. M i=1 |Yi ||Y¯i |

(12)

Average precision (AP) refers to the average probability that in the ranking labels of the predicted prescription, the label of the predicted sample in front is also the truly related label Yi . For the predicted label λ , ri (λ ) is the probability ranking of the predicted samples, and ri (λ) is the ranking of the true label of the samples. The calculation formula is as follows: AP =

M 1  1  |λ ∈ Yi |ri (λ ) ≤ ri (λ) | . M i=1 |Yi | ri (λ)

(13)

λ∈Yi

Macro-F1 focuses on the difference between state element labels. we first calculate Macro-P and Macro-R for all classes, then calculate Macro-F1 as follows: Macro-P =

Macro-R =

TP 1 1 , Pi = L L TP + FP

(14)

1 TP 1 , Ri = L L TP + FN

(15)

L

L

i=1

i=1

L

L

i=1

i=1

2Macro-P · Macro-R . (16) Macro-P + Macro-R Micro-F1 ignores the difference between labels and establishes the global confusion matrix without classifying every instance in the data set. We calculate Micro-P and Micro-R for all labels, resulting in Micro-F1: L i=1 TPi Micro-P = L , (17) L i=1 TPi + i=1 FPi L i=1 TPi , (18) Micro-R = L L i=1 TPi + i=1 FNi Macro-F1 =

Micro-F1 = 4.2

2Micro-P · Micro-R . Micro-P + Micro-R

(19)

Experimental Results and Analysis

In this section, we demonstrate the effectiveness of the proposed SIGAT method by presenting quantitative results and contrasting them with comparison algorithms for TCM status identification. In addition, we analyze ablation studies and case studies of state element classification.

10

A. Fu et al.

TCM Status Identifiaction. Table 2 shows the different training ratios of the dataset to evaluate our method, including 50%, 60%, 70%, 80% and 90%, and the rest are used as test data, respectively. The maximum macro-F1 is 37.79% when the proportion of the training set is 90%. When the training set is 80%, the hamming loss of SIGAT network training is least to 0.1060 and the 0/1 error is least to 0.2486. When the proportion of the training set is 70%, the coverage is least to 21.9862 and the ranking loss is least to 0.0858. The maximum average precision is 65.92%, and the maximum micro-F1 score is 58.29%. Table 2. Results of SIGAT model carried out on different training ratios set. Training ratio

HLoss

Coverage RLoss

0/1 Error AP

Macro-F1 Micro-F1

90%

0.1108

22.2100

0.0946

0.2500

0.6499

0.3779

0.5811

80%

0.1060 23.3618

0.0933

0.2486

0.6569

0.3094

0.5759

70%

0.1101

21.9862 0.0858 0.2594

0.6592 0.3460

0.5829

60%

0.1089

22.0839

0.0884

0.2799

0.6439

0.3127

0.5601

50%

0.1120

22.0636

0.0870

0.2897

0.6436

0.3332

0.5658

The experimental results in Table 3 show that the proposed SIGAT significantly outperforms the compared machine learning methods. For example, given 80% training data, SIGAT has the smallest coverage, ranking loss, and 0/1 error compared to SVM, NB, KNN, and RF, with an increase of 2%–10%, 1%–13%, and 6%–10% in average precision, macro-F1, and micro-F1, respectively. Table 3. Experiment results of comparison algorithms. Algorithm HLoss

Coverage RLoss

0/1 Error AP

Macro-F1 Micro-F1

SVM

0.0985 25.0317

0.1018

0.2998

0.6348

0.3000

0.5118

NB

0.1003

25.4380

0.1018

0.2810

0.6392

0.2176

0.4605

KNN

0.1026

39.1174

0.1830

0.2991

0.5735

0.1712

0.3761

RF

0.1043

45.2436

0.2153

0.3370

0.5592

0.2840

0.4718

SIGAT

0.1060

23.3618 0.0933 0.2486

0.6569 0.3094

0.5759

Ablation Study. We evaluate the contributions of the graph attention layer construction module with the ablation studies: 1. MLP as the baseline model for 80% training dataset; 2. The number of attentional heads; 3. The number of graph attention layers. According to Fig. 3, SIGAT outperforms MLP on various metrics. Experimental results show that graph attention performs better in terms of average precision, macro-F1 and micro-F1, and lower in ranking loss and 0/1 error. This

TCM Health Status Identification with Graph Attention Network

11

shows that the embedded graph attention layer module plays an important role in SIGAT, and this component is effective. Therefore, it indicates that SIGAT integrates the features of symptoms and syndromes through graph representation learning, which further verifies the importance of embedding syndromes in the model for the classification of state elements.

0.7

0.6387

0.6569

0.35

0.65

0.5759

0.6 0.55

0.4927

0.20

0.45

0.15

0.35

0.2974

0.0997

0.1060 0.0965 0.0933

0.10

0.3094

0.05

0.3 0.25 AP

0.2486

0.25

0.5 0.4

0.3033

0.30

Macro-F1

Micro-F1

0.00 MLP SIGAT ↑

HLoss

Rloss

0/1 Error MLP SIGAT ↓

Fig. 3. The effectiveness of SIGAT model compared with MLP.

To learn the graph feature of TCM data and reduce the calculation time, Fig. 4 shows the influence of different attentional heads, that is, K takes different values on the performance of the SIGAT model. The attentional heads K determine how many times the model calculates the data attention weights. The more heads, the more computationally expensive the model is, and the longer it takes. It can be seen from the experimental results that when the number of heads of attentional heads is 4, the performance of the model reaches the optimal value. It indicates that too many heads lead to over-fitting of the network. Not only fails to achieve better results but also increases the amount of calculation and training time.

0.7 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0.2

0.6504 0.6569 0.6481 0.6569 0.6550 0.6535

0.5762 0.5759 0.5621 0.5759 0.5718 0.5715

0.3515

0.3 0.2690 0.2486 0.26160.2486 0.2542 0.2579 0.25 0.2 0.15

0.3094

0.3094 0.3052 0.3055

0.2937

0.1

0.1148 0.1123 0.1060 0.1060 0.1051 0.1087 0.0948 0.0933 0.0947 0.0933 0.09330.0955

0.05 1 AP

2

3

Micro-F1

4 5 6 Number of Heads Macro-F1

1

2 0/1 Error

3

4 5 6 Number of Heads HLoss RLoss

Fig. 4. The influence of the number of attentional heads.

12

A. Fu et al.

Nodes of the graph attention network collect information from neighbouring nodes. The first layer collects the information of the first-order neighbours, and the second layer collects the information of the second-order neighbours. The experimental results in Fig. 5 show that setting the graph attention layer to 3 is the optimal parameter for SIGAT. The hamming loss, ranking loss, and 0/1 error rates reached the lowest value of 0.1060, 0.0933, and 0.2486 respectively. And the average precision and micro-F1 are the highest at 65.69% and 59.59%. 0.7 0.65 0.6 0.55 0.5 0.45 0.4 0.35 0.3

0.6569

0.6504

0.6481

0.3

0.2690

0.2486

0.25 0.5762

0.5759

0.2616

0.2 0.5621

0.15

0.1123

0.1060

0.1148

0.1 0.3515 2 AP

0.3094 3

0.3055

0.05

0.0948

Macro-F1

0.0947

0 2

4 Number of Layers

Micro-F1

0.0933 3

4 Number of Layers

0/1 Error

HLoss

RLoss

Fig. 5. The influence of the number of graph attention layer.

Case Study. Table 4 shows two prediction cases of SIGAT, the model correctly predicts the bolded state elements. In the first case, three state elements in the label are correct. There is one misreported state element: Qi deficiency. In the second case, four state elements in the label are all predicted by SIGAT correctly. These results suggest that SIGAT has substantial practical value. Table 4. State elements predicted by SIGAT. Symptoms

Syndrome Ground-Truth

Prediction

Spontaneous perspiration, Thin sloppy stool, Heavy body, Aversion to cold, Edema

Deficiency Of spleen And Kidney Yang

Yang deficiency, Cold, Wet, Spleen, Surface

Cold Limbs, Dysphagia, Headche, Vomiting and diarrhea, Nausea

Phlegm Stomach, Retention Spleen, Syndrome phlegm, Cold

Qi deficiency, Yang deficiency, Wet, Spleen

Spleen, Stomach, Cold, Wet, Phlegm

TCM Health Status Identification with Graph Attention Network

5

13

Conclusion

In this paper, we propose a graph attention network for health state identification based on TCM knowledge, called SIGAT. We define symptoms as nodes and syndrome as edges in this model. In SIGAT, automatic induction of TCM symptoms to state elements is realized with relatively high average precision, macro-F1 score, and micro-F1 score. And we implement state element classification with low hamming loss, coverage, 0/1 error, and ranking loss. The application of graph attention network classification algorithm in TCM health status identification contributes to computer-assisted diagnosis and treatment. In future work, we will consider other better composition methods to achieve a more accurate prediction of state elements, and we will also consider other work such as syndrome classification and herbal medicine recommendations. Acknowledgments. This work is partially supported by National Natural Science Foundation of China (61972187), Natural Science Foundation of Fujian Province (2020J02024).

References 1. Chen, S., Wang, Y.: Analysis of TCM health management mode with status identification as core. Asia-Pacific Tradit. Med. 15(11), 165–166 (2019) 2. Fan, H., Zhang, F., Wang, R., Huang, X., Li, Z.: Semi-supervised time series classification by temporal relation prediction. In: Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 3545–3549 (2021) 3. Fan, H., et al.: Heterogeneous hypergraph variational autoencoder for link prediction. IEEE Trans. Pattern Anal. Mach. Intell. 44(8), 4125–4138 (2021). https:// doi.org/10.1109/TPAMI.2021.3059313 4. Fan, H., Zhang, F., Xi, L., Li, Z., Guanghai, L., Xu, Y.: LeukocyteMask: an automated localization and segmentation method for leukocyte in blood smear images using deep neural networks. J. Biophotonicse 12(7) (2019) 5. Feng, S.: Classical prescriptions syndrome differentiation based on symptomatic response. China J. Tradit. Chinese Med. Pharm. 36(1), 22–26 (2021) 6. Li, C., Yang, X., Gan, H., Lai, X., Changen, Z., Chen, M.: China journal of traditional Chinese medicine and pharmacy. Asia-Pacific Tradit. Med. 26(6), 1351–1355 (2011) 7. Liang, W., Lin, X., Yu, J., Min, L., Li, C.: Big database of real world promotes health management of traditional Chinese medicine into artificial intelligence era. China J. Tradit. Chinese Med. Pharm. 33(4), 1213–1215 (2018) 8. Liu, J., Xie, Y.: Collaborative filtering recommendation algorithm based on graph attention network representation learning. Comput. Syst. Appl. 31(4), 273–280 (2022) 9. Nie, T.: Research on short text classification based on graph attention networks. Northeast Normal University (2021) 10. Pan, M.: Research on social recommender system based on graph attention network. Northeast Normal University (2021) 11. Shi, Y., Zhou, K., Li, S., Zhou, M., Liu, W.: Heterogeneous graph attention network for food safety risk prediction. J. Food Eng. 323, 111005 (2022)

14

A. Fu et al.

12. Song, Z., Li, Y., Li, D., Li, S.: Multi-label classification of legal text with fusion of label relations. Pattern Recogn. Artif. Intell. 35(2), 185–192 (2022) 13. Tsoumakas, G., Katakis, I., Vlahavas, I.: Mining multi-label data. In: Data Mining and Knowledge Discovery Handbook, pp. 667–685 (2010) 14. Veliˇckovi´c, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. In: Proceedings of the International Conference on Learning Representations (2017) 15. Xin, J.: Research of health status identification algorithm based on TCM theory of state. Fujian University of Traditional Chinese Medicine (2021) 16. Xin, J., Li, S., Zhang, J., Lei, H., Candong, L.: Dicovery of identification method of traditional Chinese medicine health status. China J. Tradit. Chinese Med. Pharm. 34(7), 3151–3153 (2019) 17. Xin, J., Zhang, J., Li, S., Li, C.: Research on multi-label classification methods for the identification of health state in traditional Chinese medicine. China J. Tradit. Chinese Med. Pharm. 34(9), 3952–3955 (2019) 18. Xu, J., et al.: Rules of traditional Chinese medicine state identification based on artificial intelligence algorithm. J. Tradit. Chinese Med. 61(3), 204–208 (2020) 19. Zhang, M., Wu, L.: Lift: multi-label learning with label-specific features. IEEE Trans. Pattern Anal. Mach. Intell. 37(1), 107–120 (2015) 20. Zhao, W., Lu, W., Li, Z., Zhou, C., Fan, H., Yang, Z.: TCM herbal prescription recommendation model based on multi-graph convolutional network. J. Ethnopharmacol. 297, 115109 (2022)

Flexible Task Splitting Strategy in Aircraft Maintenance Technician Scheduling Based on Swarm Intelligence Bowen Xue1 , Huifen Zhong1 , Junrui Lu1,2 , Tianwei Zhou1 , and Ben Niu1(B) 1 College of Management, Shenzhen University, Shenzhen 518060, China

[email protected] 2 Greater Bay Area International Institute for Innovation, Shenzhen University,

Shenzhen 518060, China

Abstract. Working overtime is a problem that airlines and maintenance technicians pay great attention to and have not been effectively solved. To alleviate this phenomenon as well as save the maintenance cost and time, this paper introduces a flexible task splitting strategy (FTSS) into the original aircraft maintenance technician scheduling (AMTS) model and presents a novel model called AMTS-FTSS. In AMTS-FTSS, one maintenance task can be completed by multiple multi-skilled maintenance technicians. When the maintenance task needs to be completed over time, it can be flexibly split, and the splitting standard is controlled by the flexible time. Finally, ant colony optimization, particle swarm optimization, bacterial foraging optimization, and artificial bee colony algorithms and their variants are used to verify the effectiveness and universality of FTSS. The experimental results show that compared with the AMTS model, AMTS-FTSS can save the maintenance cost and time in most circumstances. Keywords: Flexible task splitting strategy · Aircraft maintenance technician scheduling · Swarm intelligence algorithm

1 Introduction Proper scheduling and control of fatigue levels of maintenance technicians are the requirements of the civil aviation administration for airlines, which provides an essential guarantee for the safety and high-quality development of civil aviation [1]. The early literature on aircraft maintenance technician scheduling (AMTS) mainly focused on scheduling maintenance teams, which spanned a long maintenance cycle and did not involve specific maintenance task assignment [2]. Then, Gang et al. [3] refined the aircraft maintenance technician scheduling to assign specific maintenance tasks and simplified the problem to a maintenance task that only needs one maintenance technician to complete. On this basis, Qin et al. [4] and Niu et al. [5] conducted further research. Qin et al. [4] combined maintenance staff scheduling with aircraft allocation to build a two-stage aircraft maintenance joint scheduling model. Niu et al. [5] improved the assignment mode of aircraft maintenance technicians and built a distributed aircraft © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 Y. Xu et al. (Eds.): ML4CS 2022, LNCS 13655, pp. 15–26, 2023. https://doi.org/10.1007/978-3-031-20096-0_2

16

B. Xue et al.

maintenance technician scheduling model. However, few studies can consider the urgent problems in aircraft maintenance technician scheduling, such as overtime work and high fatigue levels of maintenance technicians. Therefore, based on the research of Niu et al. [5], this paper designs an aircraft maintenance technician scheduling model more suitable for the actual maintenance scenario. In this model, the maintenance technicians have the attribute of multi-skills. One maintenance task is no longer limited by the number of maintenance technicians but can be assigned to multiple maintenance technicians to complete together. In addition, to alleviate the overtime work of maintenance technicians, this paper proposes a flexible task splitting strategy for the scheduling model, namely AMTS-FTSS. Finally, as the research models in this paper are multi-dimensional, complex, and multi-constraint nonlinear models, this paper uses the swarm intelligence optimization algorithms suitable for NP-hard problems as the solution methods. At the same time, in order to verify the effectiveness of AMTS-FTSS in controlling cost and saving time, this paper applies eight common swarm intelligence algorithms to solve the models, including artificial ant colony optimization (ACO) [6], particle swarm optimization (PSO) [7], comprehensive learning particle swarm optimizer (CLPSO) [8], hybrid firefly and particle swarm optimization (HFPSO) [9], phasor particle swarm optimization (PPSO) [10], bacterial foraging optimization (BFO) [11], bacterial colony optimization (BCO) [12], and artificial bee colony (ABC) [13] algorithms. The remainder of this paper is organized as follows. Section 2 explains the definitions of aircraft maintenance technician scheduling and proposes the flexible task splitting strategy. The formulations of our proposed model are given in Sect. 3. In Sect. 4, eight swarm intelligence algorithms are briefly introduced, and the encoding scheme is presented. Section 5 displays the experiments and results analysis, after which the whole paper is concluded in Sect. 6.

2 Problem Definition and Flexible Task Splitting Strategy Aircraft maintenance technician scheduling (AMTS) mainly focuses on the practical assignment of aircraft maintenance technicians for maintenance tasks to save maintenance costs and time. Besides, it is a dynamic process involving aircraft, shift, maintenance task, maintenance technician, etc. The whole scheduling process mainly depends on the completion progress of the maintenance task sequence. 2.1 Problem Description The execution of aircraft maintenance work is often guided by cost control and timesaving. When an aircraft arrives at the maintenance hangar, it begins to be repaired. Each aircraft has a fixed task sequence, and each maintenance task has requirements for license levels and numbers of maintenance technicians holding each license level. Each aircraft is equipped with an independent maintenance team, including multiple maintenance technicians with multiple license levels. During the scheduling process, technicians shall be assigned to each maintenance task in turn according to the arrival time of each aircraft. The implementation of technician assignment shall be carried out in

Flexible Task Splitting Strategy in AMTS

17

strict accordance with the requirements of maintenance tasks on the number and license level of technicians. Among them, technicians with higher license levels can replace technicians with lower license levels to complete maintenance tasks. 2.2 Flexible Task Splitting Strategy In the technician assignment process, different technicians may spend different times completing the same maintenance task. Therefore, the work progress of the task sequence depends on the technician scheduling scheme, and the exact maintenance shift of maintenance tasks cannot be predicted, leading to the technicians working overtime for some complex and time-consuming tasks. In order to alleviate the problem of overtime work, this section proposes a flexible task splitting mechanism. It introduces the technician exchange method on this basis in order to save maintenance time. Then the model can be called an aircraft maintenance technician scheduling model with a flexible task splitting strategy (AMTS-FTSS). At the end of maintenance shift t, if maintenance task s has not been completed, it is necessary to consider whether task s needs to be split and give the reasonable assignment of maintenance technicians. This kind of tasks can be defined in formula (1), where rt ism1 is the time required for technician m1 to complete task s of aircraft i, and st is is the start time of task s of aircraft i. rtism1 > H · t − stis , ∀i ∈ A, ∀m1 ∈ Mi , ∀s ∈ Si , ∀t ∈ T

(1)

In particular, if the start time of this task is slightly earlier or its estimated end time is slightly later than the end time of shift t, it can be considered not to split this task. In other words, the time threshold range can be set according to the needs of the actual scenario to split the maintenance tasks satisfying formula (1) flexibly. Then, the time threshold range is set to [H · t − f , H · t + f ], where H represents the length of a shift and f is called the flexible time to adjust this range. When a maintenance task spans two shifts and its start time or estimated end time is outside this range, it is necessary to split the maintenance task into subtasks s1 and s2 . Therefore, the conditions for a flexible task splitting strategy include formulas (2) and (3). stis < H · t − f , ∀i ∈ A, ∀m1 ∈ Mi , ∀s ∈ Si , ∀t ∈ T

(2)

stis + rtism1 > H · t + f , ∀i ∈ A, ∀m1 ∈ Mi , ∀s ∈ Si , ∀t ∈ T

(3)

2.3 Technician Assignment Methods As mentioned in Sect. 2.2, if maintenance task s is split, there will be two subtasks, i.e., s1 and s2 . The time left for subtask s1 is fixed, which is represented by lt ist , and the calculation of lt ist is shown in formula (4). ltist = H · t − stis

(4)

18

B. Xue et al.

Specifically, s1 and s2 have the same requirements for maintenance technicians as task s. Then, another group of technicians m2 should be chosen for task s. Generally, the time required for technicians m2 to complete task s is different from m1 , and it can also be understood that the work efficiencies of technicians m1 and m2 are distinctive. Therefore, there may be a case that the maintenance task s does not need to be split when replacing technicians m1 with m2 , i.e., rt ism2 + st is ≤ H · t + f . Except for the above circumstance, in order to save maintenance time, it is necessary to assign technicians m1 and m2 to subtasks s1 and s2 according to the work efficiency and the estimated maintenance time of task s. Obviously, there are two assignment methods, which are illustrated as follows. No Technician Exchange Method. With this method, technicians m1 are assigned to subtask s1 and technicians m2 are assigned to subtask s2 . At the end of shift t, when subtask s1 is finished, the work progress of task s can be represented by lt ist /rt ism1 , then the left work progress for technicians m2 is (1 − lt ist /rt ism1 ). After that, the maintenance time of task s with no technician exchange method is:   mtis0 = ltist + 1 − ltist /rtism1 · rtism2 (5) Technician Exchange Method. With this method, technicians m2 are assigned to subtask s1 and technicians m1 are assigned to subtask s2 . Compared with the no technician exchange method, this method exchanges the sequence of technicians m1 and m2 . Thus, at the end of shift t, the work progress of task s is lt ist /rt ism2 . Then the left work progress for technicians m1 is (1 − lt ist /rt ism2 ). Similarly, the maintenance time of task s with the technician exchange method is shown in formula (6).   mtis1 = ltist + 1 − ltist /rtism2 · rtism1 (6) Then, by comparing the maintenance time obtained by the above two methods, the method with the shortest maintenance time is selected as the final technician assignment method, and the following assignment principles are obtained. Finally, the flowchart of aircraft maintenance technician scheduling with the flexible task splitting strategy is demonstrated in Fig. 1. • If 0 < lt ist ≤

rt ism1 ·rt ism2 rt ism1 +rt ism2 ,

1

then mt 0is ≤ mt is , and no technician exchange method is

applied, the maintenance time mt is of task s is equal to mt 0is . • If

rt ism1 ·rt ism2 rt ism1 +rt ism2

1

< lt ist ≤ H , then mt 0is > mt is , and the technician exchange method is

applied, the maintenance time mt is of task s is equal to mt 1is .

3 Model Design and Formulations In this section, the objective and constraints of the AMTS-FTSS model are given. The variables and their meanings are shown in Table 1.

Flexible Task Splitting Strategy in AMTS

19

Start

N

N

a