Download ebook Web And Big Data 5Th International Joint Conference Apweb Waim 2021 Guangzhou China August 23 25 2021 Proceedings Part I Lecture Notes In Computer Science Leong Hou U Editor online pdf all chapter docx epub

Web and Big Data: 5th International
Joint Conference, APWeb-WAIM 2021,

Guangzhou, China, August 23–25, 2021,
Proceedings, Part I (Lecture Notes in
Computer Science) Leong Hou U
(Editor)
Visit to download the full and correct content document:
https://ebookmeta.com/product/web-and-big-data-5th-international-joint-conference-a
pweb-waim-2021-guangzhou-china-august-23-25-2021-proceedings-part-i-lecture-not
es-in-computer-science-leong-hou-u-editor/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
Cambridge IGCSE and O Level History Workbook 2C - Depth

Study: the United States, 1919-41 2nd Edition Benjamin
Harrison
https://ebookmeta.com/product/cambridge-igcse-and-o-level-
history-workbook-2c-depth-study-the-united-states-1919-41-2nd-
edition-benjamin-harrison/
Pattern Recognition and Computer Vision 4th Chinese

Conference PRCV 2021 Beijing China October 29 November
1 2021 Proceedings Part IV Lecture Notes in Computer
Science Huimin Ma (Editor)
https://ebookmeta.com/product/pattern-recognition-and-computer-
vision-4th-chinese-conference-prcv-2021-beijing-china-
october-29-november-1-2021-proceedings-part-iv-lecture-notes-in-
computer-science-huimin-ma-editor/
Image and Graphics 11th International Conference ICIG

2021 Haikou China August 6 8 2021 Proceedings Part III
Lecture Notes in Computer Science 12890 Yuxin Peng
(Editor)
https://ebookmeta.com/product/image-and-graphics-11th-
international-conference-icig-2021-haikou-china-
august-6-8-2021-proceedings-part-iii-lecture-notes-in-computer-
science-12890-yuxin-peng-editor/
Intelligent Computing Theories and Application 17th

International Conference ICIC 2021 Shenzhen China
August 12 15 2021 Proceedings Part I Lecture Notes in
Computer Science Book 12836 2nd Edition Katharina Rein
https://ebookmeta.com/product/intelligent-computing-theories-and-
application-17th-international-conference-icic-2021-shenzhen-
china-august-12-15-2021-proceedings-part-i-lecture-notes-in-
Artificial Intelligence and Security 7th International
Conference ICAIS 2021 Dublin Ireland July 19 23 2021
Proceedings Part I Lecture Notes in Computer Science
12736 Xingming Sun (Editor)
https://ebookmeta.com/product/artificial-intelligence-and-
security-7th-international-conference-icais-2021-dublin-ireland-
july-19-23-2021-proceedings-part-i-lecture-notes-in-computer-
science-12736-xingming-sun-editor/
Web Information Systems Engineering – WISE 2021: 22nd

International Conference on Web Information Systems
Engineering, WISE 2021, Melbourne, VIC, Australia,
October 26-29, 2021 Proceedings, Part I Wenjie Zhang
https://ebookmeta.com/product/web-information-systems-
engineering-wise-2021-22nd-international-conference-on-web-
information-systems-engineering-wise-2021-melbourne-vic-
australia-october-26-29-2021-proceedings-pa/
Quantitative Evaluation of Systems 18th International

Conference QEST 2021 Paris France August 23 27 2021
Proceedings Lecture Notes in Computer Science
Alessandro Abate (Editor)
https://ebookmeta.com/product/quantitative-evaluation-of-
systems-18th-international-conference-qest-2021-paris-france-
august-23-27-2021-proceedings-lecture-notes-in-computer-science-
alessandro-abate-editor/
Intelligent Computing Theories and Application 17th

International Conference ICIC 2021 Shenzhen China
August 12 15 2021 Proceedings Part III Lecture Notes in
Computer Science De-Shuang Huang (Editor)
https://ebookmeta.com/product/intelligent-computing-theories-and-
application-17th-international-conference-icic-2021-shenzhen-
china-august-12-15-2021-proceedings-part-iii-lecture-notes-in-
computer-science-de-shuang-huang-editor/
Advances in Swarm Intelligence 12th International

Conference ICSI 2021 Qingdao China July 17 21 2021
Proceedings Part I Lecture Notes in Computer Science
12689 Ying Tan (Editor)
https://ebookmeta.com/product/advances-in-swarm-
intelligence-12th-international-conference-icsi-2021-qingdao-
china-july-17-21-2021-proceedings-part-i-lecture-notes-in-
Leong Hou U
Marc Spaniol
Yasushi Sakurai
Junying Chen (Eds.)
LNCS 12858
Web and Big Data

5th International Joint Conference, APWeb-WAIM 2021
Guangzhou, China, August 23–25, 2021
Proceedings, Part I
Lecture Notes in Computer Science 12858
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
Gerhard Woeginger
RWTH Aachen, Aachen, Germany
Moti Yung
Columbia University, New York, NY, USA
More information about this subseries at http://www.springer.com/series/7409
Leong Hou U Marc Spaniol
• •
Yasushi Sakurai Junying Chen (Eds.)

•
Web and Big Data

5th International Joint Conference, APWeb-WAIM 2021
Guangzhou, China, August 23–25, 2021
Proceedings, Part I
123
Editors
Leong Hou U Marc Spaniol
University of Macau University of Caen Normandie
Macau, China Caen, France
Yasushi Sakurai Junying Chen
Osaka University South China University of Technology
Osaka, Japan Guangzhou, China
ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Computer Science
ISBN 978-3-030-85895-7 ISBN 978-3-030-85896-4 (eBook)
https://doi.org/10.1007/978-3-030-85896-4
LNCS Sublibrary: SL3 – Information Systems and Applications, incl. Internet/Web, and HCI
© Springer Nature Switzerland AG 2021

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
This volume (LNCS 12858) and its companion volume (LNCS 12859) contain the
proceedings of the fifth Asia-Pacific Web (APWeb) and Web-Age Information Man-
agement (WAIM) Joint Conference on Web and Big Data, called APWeb-WAIM.
With the increased focus on big data, the new joint conference is expected to attract
more professionals from different industrial and academic communities, not only from
the Asia-Pacific countries but also from other continents. The objective is to enable the
sharing and exchange of ideas, experiences, and results in the areas of the World Wide
Web and big data, thus covering web technologies, database systems, information
management, software engineering, and big data.
The fifth APWeb-WAIM conference was held in Guangzhou during August 23–25,
2021. As an Asia-Pacific flagship conference focusing on research, development, and
applications in relation to Web information management, APWeb-WAIM builds on the
successes of APWeb and WAIM: APWeb was previously held in Beijing (1998), Hong
Kong (1999), Xi’an (2000), Changsha (2001), Xi’an (2003), Hangzhou (2004),
Shanghai (2005), Harbin (2006), Huangshan (2007), Shenyang (2008), Suzhou (2009),
Busan (2010), Beijing (2011), Kunming (2012), Sydney (2013), Changsha (2014),
Guangzhou (2015), and Suzhou (2016); and WAIM was held in Shanghai (2000),
Xi’an (2001), Beijing (2002), Chengdu (2003), Dalian (2004), Hangzhou (2005), Hong
Kong (2006), Huangshan (2007), Zhangjiajie (2008), Suzhou (2009), Jiuzhaigou
(2010), Wuhan (2011), Harbin (2012), Beidaihe (2013), Macau (2014), Qingdao
(2015), and Nanchang (2016). The APWeb-WAIM conferences were held in Beijing
(2017), Macau (2018), Chengdu (2019), and Tianjin (2020). With the fast development
of web-related technologies, we expect that APWeb-WAIM will become an increas-
ingly popular forum that brings together outstanding researchers and developers in the
fields of the Web and big data from around the world.
The high-quality program documented in these proceedings would not have been
possible without the authors who chose APWeb-WAIM for disseminating their find-
ings. A total of 184 submissions were received and, after the double-blind review
process (each paper received at least three review reports), the conference accepted 44
regular papers (23.91%), 24 short research papers, and 6 demonstrations. The con-
tributed papers address a wide range of topics, such as graph mining, data mining, data
management, topic model and language model learning, text analysis, text classifica-
tion, machine learning, knowledge graphs, emerging data processing techniques,
information extraction and retrieval, recommender systems, and spatial and
spatio-temporal databases. The technical program also included keynotes by M. Tamer
Özsu (University of Waterloo, USA), Huan Liu (Arizona State University, Tempe,
USA), X. Sean Wang (Fudan University, China), and Xiaokui Xiao (National
University of Singapore, Singapore). We are grateful to these distinguished scientists
for their invaluable contributions to the conference program. As a joint conference,
teamwork is particularly important for the success of APWeb-WAIM. We are deeply
vi Preface
thankful to the Program Committee members and the external reviewers for lending
their time and expertise to the conference. Special thanks go to the local Organizing
Committee led by Yi Cai. Thanks also go to the workshop chairs (Yunjun Gao, An Liu,
and Xiaohui Tao), demo chair (Yanghui Rao), industry chair (Jianming Lv), tutorial
chair (Raymond Chi-Wing Wong), publication chair (Junying Chen), local arrange-
ment chairs (Guohua Wang and Junying Chen), and publicity chairs (Xin Wang and
Jianxin Li). Their efforts were essential to the success of the conference. Last but not
least, we wish to express our gratitude to the Webmaster (Jianwei Lu), for all the hard
work, and to our sponsors who generously supported the smooth running of the
conference.
We hope you enjoy the exciting program of APWeb-WAIM 2021 as documented in
these proceedings.
July 2021 Yi Cai

Tom Gedeon
Qing Li
Baltasar Fernández Manjón
Leong Hou U
Marc Spaniol
Yasushi Sakurai
Organization
Organizing Committee
General Chairs
Yi Cai South China University of Technology, China
Tom Gedeon Australia National University, Australia
Qing Li Hong Kong Polytechnic University, China
Baltasar Fernández Manjón UCM, Spain
Program Committee Chairs

Leong Hou U University of Macau, China
Marc Spaniol Université de Caen Normandie, France
Yasushi Sakurai Osaka University, Japan
Workshop Chairs
Yunjun Gao Zhejiang University, China
An Liu Soochow University, China
Xiaohui Tao University of Southern Queensland, Australia
Demo Chair
Yanghui Rao Sun Yat-sen University, China
Tutorial Chair
Raymond Chi-Wing Wong Hong Kong University of Science and Technology,
China
Industry Chair
Jianming Lv South China University of Technology, China
Publication Chair
Junying Chen South China University of Technology, China
viii Organization
Publicity Chairs
Xin Wang Tianjin University, China
Jianxin Li Deakin University, Australia
Local Arrangement Chairs

Guohua Wang South China University of Technology, China
Webmaster
Jianwei Lu South China University of Technology, China
APWeb-WAIM Steering Committee Representative

Yanchun Zhang Victoria University, Australia
Senior Program Committee Members

Feida Zhu Singapore Management University, Singapore
Lei Chen Hong Kong University of Science and Technology,
China
Mizuho Iwaihara Waseda University, Japan
Peer Kroger Christian-Albrechst-University Kiel, Germany
Reynold Cheng The University of Hong Kong, China
Wolf-Tilo Balke TU Braunschweig, Germany
Xiang Zhao National University of Defence Technology, China
Yunjun Gao Zhejiang University, China
Zhiguo Gong University of Macau, China
Program Committee Members

Alex Delis University of Athens, Greece
An Liu Soochow University, China
Aviv Segev KAIST, Korea
Baoning Niu Taiyuan University of Technology, China
Bin Cui Peking University, China
Bo Tang Southern University of Science and Technology, China
Bohan Li Nanjing University of Aeronautics and Astronautics,
China
Bolong Zheng Huazhong University of Science and Technology,
China
Carson K. Leung University of Manitoba, Canada
Cheqing Jin East China Normal University, China
Chih-Hua Tai National Taipei University, China
Organization ix
Defu Lian University of Electronic Science and Technology

of China, China
Dhaval Patel IBM TJ Watson Research Center, USA
Dimitris Sacharidis ULB, Belgium
Giovanna Guerrini University of Genoa, Italy
Guoqiong Liao Jiangxi University of Finance and Economics, China
Haibo Hu Hong Kong Polytechnic University, China
Hailong Sun Beihang University, China
Haiwei Zhang Nankai University, China
Han Su University of Southern California, USA
Hao Wang Nanjing University of Information Science
and Technology, China
Hiroaki Ohshima University of Hyogo, Japan
Hongzhi Wang Harbin Institute of Technology, China
Hua Wang Victoria University, Australia
Hui Li Xiamen University, China
Hui Luo RMIT University, Australia
Ilaria Bartolini University of Bologna, Italy
Jian Yin Sun Yat-sen University, China
Jianbin Huang Xidian University, China
Jianbin Huang XDU, China
Jianming Lv South China University of Technology, China
Jianxin Li Beihang University, China
Jianzhong Qi The University of Melbourne, Australia
Jieming Shi The Hong Kong Polytechnic University, China
Ju Fan Renmin University of China, China
Jun Gao Peking University, China
Junhu Wang Griffith University, Australia
Junjie Yao East China Normal University, China
Kai Yang City University of Hong Kong, China
Kai Zeng Microsoft, USA
Kai Zheng University of Electronic Science and Technology
of China, China
Krishna Reddy P. International Institute of Information Technology, India
Kyuseok Shim Seoul National University, Korea
Lei Duan Sichuan University, China
Leong Hou U University of Macau, China
Liang Hong Wuhan University, China
Lianghuai Yang Zhejiang University of Technology, China
Lu Chen Zhejiang University, China
Man Lung Yiu Hong Kong Polytechnic University, China
Maria Luisa Damiani University of Milan, Italy
Markus Endres University of Augsburg, Germany
Meng Wang Southeast University, China
Mirco Nanni ISTI-CNR, Italy
x Organization
Panagiotis Karras Aarhus University, Denmark

Peiquan Jin University of Science and Technology of China, China
Peng Wang Fudan University, China
Qingbao Huang Guangxi University, China
Raymond Chi-Wing Wong Hong Kong University of Science and Technology,
China
Rong-Hua Li Beijing Institute of Technology, China
Sanghyun Park Yonsei University, Korea
Sangkeun Lee Oak Ridge National Laboratory, USA
Sanjay Kumar Madria Missouri University of Science and Technology, USA
Senzhang Wang Central South University, China
Shaoxu Song Tsinghua University, China
Sheng Wang New York University, USA
Shengli Wu Jiangsu University, China
Shuyue Hu NUS, Singapore
Taketoshi Ushiama Kyushu University, Japan
Tao Wang King’s College London, UK
Tieyun Qian Wuhan University, China
Ting Deng Beihang University, China
Tingjian Ge University of Massachusetts, Lowell, USA
Vincent Oria NJIT, USA
Wee Siong Ng Institute for Infocomm Research, Singapore
Wei Lu Renmin University of China, China
Wei Song Wuhan University, China
Wei Wang University of New South Wales, Australia
Wen Zhang Wuhan University, China
Xiang Lian Kent State University, USA
Xiangmin Zhou RMIT University, Australia
Xiaochun Yang Northeastern University, USA
Xiaohui Tao The University of Southern Queensland, Australia
Xiaokui Xiao National University of Singapore, Singapore
Xiaowei Wu University of Macau, China
Xike Xie University of Science and Technology of China, China
Xin Cao University of New South Wales, Australia
Xin Huang Hong Kong Baptist University, China
Xin Wang Tianjin University, China
Xingquan Zhu Florida Atlantic University, USA
Xudong Mao Xiamen University, China
Yafei Li Zhengzhou University, China
Yajun Yang Tianjin University, China
Yanghua Xiao Fudan University, China
Yanghui Rao Sun Yat-sen University, China
Yang-Sae Moon Kangwon National University, South Korea
Yaokai Feng Kyushu University, Japan
Yijie Wang National University of Defense Technology, China
Yingxia Shao BUPT, China
Organization xi
Yongpan Sheng Chongqing University, China

Yongxin Tong Beihang University, China
Yu Gu Northeastern University, USA
Zakaria Maamar Zayed University, United Arab of Emirates
Zhaonian Zou Harbin Institute of Technology, China
Zhixu Li Soochow University, China
Zouhaier Brahmia University of Sfax, Tunisia
Keynotes
Approaches to Distributed RDF Data
Management and SPARQL Processing
M. Tamer Özsu
David R. Cheriton School of Computer Science, University of Waterloo, Canada

tamer.ozsu@uwaterloo.ca
Abstract. Resource Description Framework (RDF) has been proposed for

modelling Web objects as part of developing the “semantic web”, but its usage
has extended beyond this original objective. As the volume of RDF data has
increased, the usual scalability issues have arisen and solutions have been
developed for distributed/parallel processing of SPARQL queries over large
RDF datasets. RDF has also gained attention as a way to accomplish data
integration, leading to federated approaches. In this talk I will provide an
overview of work in these two areas.
Striving for Socially Responsible AI
in Data Science
Huan Liu
School of Computing and Augmented Intelligence,

Arizona State University, USA
huanliu@asu.edu
Abstract. AI has never been this pervasive and effective. AI algorithms are used
in news feeds, friend/purchase recommendation, making hiring and firing
decisions, and political campaigns. Data empowers AI algorithms and is then
collected again for further training AI algorithms. We come to realize that AI
algorithms have biases, and some biases might result in deleterious effects.
Facing existential challenges, we explore how socially responsible AI can help
in data science: what it is, why it is important, and how it can protect and inform
us, and help prevent or mitigate the misuse of AI. We show how socially
responsible AI works via use cases of privacy preservation, cyberbullying
identification, and disinformation detection. Knowing the problems with AI and
our own conflicting goals, we further discuss some quandaries and challenges in
our pursuit of socially responsible AI.
Democratizing the Full Data Analytics
Software Stack
X. Sean Wang
School of Computer Science, Fudan University, China

xywangcs@fudan.edu.cn
Abstract. Data analysis and machine learning is a complex task, involving a full
stack of hardware and software systems, from the usual compute systems, cloud
computing and supercomputing systems, to data collection systems, data storage
and database systems, data mining and machine learning systems, and data
visualization and interaction systems. A realistic and highly efficient data ana-
lytics and AI application often requires a smooth collaboration among the dif-
ferent systems, which becomes a big technical hurdle, especially to the
non-computing professionals. The history of computing may be viewed as a
technical democratizing processing, which in turn brings huge benefit to the
society and its economy. The democratizing process for data analysis and
machine learning has started to appear in various aspects, but it still needs
research and development in multiple directions, including human-machine
natural interaction, automated system selection and deployment, and automated
workflow execution and optimization. It can be expected that this democratizing
process will continue, and the research and development efforts by the computer
scientists are much needed.
Efficient Network Embeddings
for Large Graphs
Xiaokui Xiao
School of Computing, National University of Singapore, Singapore

xkxiao@nus.edu.sg
Abstract. Given a graph G, network embedding maps each node in G into a

compact, fixed-dimensional feature vector, which can be used in downstream
machine learning tasks. Most of the existing methods for network embedding
fail to scale to large graphs with millions of nodes, as they either incur signif-
icant computation cost or generate low-quality embeddings on such graphs. In
this talk, we will present two efficient network embedding algorithms for large
graphs with and without node attributes, respectively. The basic idea is to first
model the affinity between nodes (or between nodes and attributes) based on
random walks, and then factorize the affinity matrix to derive the embeddings.
The main challenges that we address include (i) the choice of the affinity
measure and (ii) the reduction of space and time overheads entailed by the
construction and factorization of the affinity matrix. Extensive experiments on
large graphs demonstrate that our algorithms outperform the existing methods in
terms of both embedding quality and efficiency.
Contents – Part I
Graph Mining
Co-authorship Prediction Based on Temporal Graph Attention . . . . . . . . . . . 3

Dongdong Jin, Peng Cheng, Xuemin Lin, and Lei Chen
Degree-Specific Topology Learning for Graph Convolutional Network . . . . . 20

Jiahou Cheng, Mengqing Luo, Xin Li, and Hui Yan
Simplifying Graph Convolutional Networks as Matrix Factorization . . . . . . . 35

Qiang Liu, Haoli Zhang, and Zhaocheng Liu
GRASP: Graph Alignment Through Spectral Signatures. . . . . . . . . . . . . . . . 44

Judith Hermanns, Anton Tsitsulin, Marina Munkhoeva, Alex Bronstein,
Davide Mottin, and Panagiotis Karras
FANE: A Fusion-Based Attributed Network Embedding Framework . . . . . . . 53

Guanghua Li, Qiyan Li, Jingqiao Liu, Yuanyuan Zhu, and Ming Zhong
Data Mining
What Have We Learned from OpenReview?. . . . . . . . . . . . . . . . . . . . . . . . 63

Gang Wang, Qi Peng, Yanfeng Zhang, and Mingyang Zhang
Unsafe Driving Behavior Prediction for Electric Vehicles. . . . . . . . . . . . . . . 80

Jiaxiang Huang, Hao Lin, and Junjie Yao
Resource Trading with Hierarchical Game for Computing-Power

Network Market . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Qingzhong Bao, Xiaoxu Ren, Chunfeng Liu, Xin Wang, Xiaofei Wang,
and Chao Qiu
Analyze and Evaluate Database-Backed Web Applications with WTool . . . . . 110

Zhou Zhou and XuJia Yao
Semi-supervised Variational Multi-view Anomaly Detection . . . . . . . . . . . . . 125

Shaoshen Wang, Ling Chen, Farookh Hussain, and Chengqi Zhang
A Graph Attention Network Model for GMV Forecast on Online

Shopping Festival . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Qianyu Yu, Shuo Yang, Zhiqiang Zhang, Ya-Lin Zhang, Binbin Hu,
Ziqi Liu, Kai Huang, Xingyu Zhong, Jun Zhou, and Yanming Fang
xx Contents – Part I
Suicide Ideation Detection on Social Media During COVID-19 via

Adversarial and Multi-task Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Jun Li, Zhihan Yan, Zehang Lin, Xingyun Liu, Hong Va Leong,
Nancy Xiaonan Yu, and Qing Li
Data Management
An Efficient Bucket Logging for Persistent Memory . . . . . . . . . . . . . . . . . . 149

Xiyan Xu and Jiwu Shu
Data Poisoning Attacks on Crowdsourcing Learning . . . . . . . . . . . . . . . . . . 164

Pengpeng Chen, Hailong Sun, and Zhijun Chen
Dynamic Environment Simulation for Database Performance Evaluation . . . . 180

Chunxi Zhang, Rong Zhang, Qian Su, and Aoying Zhou
LinKV: An RDMA-Enabled KVS for High Performance and Strict

Consistency Under Skew . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
Xing Wei, Huiqi Hu, Xuan Zhou, and Aoying Zhou
Cheetah: An Adaptive User-Space Cache for Non-volatile Main Memory

File Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
Tian Yan, Linpeng Huang, and Shengan Zheng
Topic Model and Language Model Learning
Chinese Word Embedding Learning with Limited Data . . . . . . . . . . . . . . . . 211

Shurui Chen, Yufu Chen, Yuyin Lu, Yanghui Rao, Haoran Xie,
and Qing Li
Sparse Biterm Topic Model for Short Texts . . . . . . . . . . . . . . . . . . . . . . . . 227

Bingshan Zhu, Yi Cai, and Huakui Zhang
EMBERT: A Pre-trained Language Model for Chinese Medical

Text Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
Zerui Cai, Taolin Zhang, Chengyu Wang, and Xiaofeng He
Self-supervised Learning for Semantic Sentence Matching with Dense

Transformer Inference Network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 258
Fengying Yu, Jianzong Wang, Dewei Tao, Ning Cheng, and Jing Xiao
An Explainable Evaluation of Unsupervised Transfer Learning for Parallel

Sentences Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Shaolin Zhu, Chenggang Mi, and Xiayang Shi
Contents – Part I xxi
Text Analysis
Leveraging Syntactic Dependency and Lexical Similarity for Neural

Relation Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Yashen Wang
A Novel Capsule Aggregation Framework for Natural Language Inference. . . 300

Chao Sun, Jianzong Wang, Fengying Yu, Ning Cheng, and Jing Xiao
Learning Modality-Invariant Features by Cross-Modality Adversarial

Network for Visual Question Answering . . . . . . . . . . . . . . . . . . . . . . . . . . 316
Ze Fu, Changmeng Zheng, Yi Cai, Qing Li, and Tao Wang
Difficulty-Controllable Visual Question Generation . . . . . . . . . . . . . . . . . . . 332

Feng Chen, Jiayuan Xie, Yi Cai, Tao Wang, and Qing Li
Incorporating Typological Features into Language Selection

for Multilingual Neural Machine Translation . . . . . . . . . . . . . . . . . . . . . . . 348
Chenggang Mi, Shaolin Zhu, Yi Fan, and Lei Xie
Removing Input Confounder for Translation Quality Estimation

via a Causal Motivated Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
Xuewen Shi, Heyan Huang, Ping Jian, and Yi-Kun Tang
Text Classification
Learning Refined Features for Open-World Text Classification . . . . . . . . . . . 367

Zeting Li, Yi Cai, Xingwei Tan, Guoqiang Han, Haopeng Ren, Xin Wu,
and Wen Li
Emotion Classification of Text Based on BERT and Broad

Learning System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
Sancheng Peng, Rong Zeng, Hongzhan Liu, Guanghao Chen,
Ruihuan Wu, Aimin Yang, and Shui Yu
Improving Document-Level Sentiment Classification with User-Product

Gated Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
Bing Tian, Yong Zhang, and Chunxiao Xing
Integrating RoBERTa Fine-Tuning and User Writing Styles for Authorship

Attribution of Short Texts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
Xiangyu Wang and Mizuho Iwaihara
Dependency Graph Convolution and POS Tagging Transferring

for Aspect-Based Sentiment Classification . . . . . . . . . . . . . . . . . . . . . . . . . 422
Zexin Li, Linjun Chen, Tiancheng Huang, and Jiagang Song
xxii Contents – Part I
Machine Learning 1
DTWSSE: Data Augmentation with a Siamese Encoder for Time Series . . . . 435
Xinyu Yang, Xinlan Zhang, Zhenguo Zhang, Yahui Zhao,
and Rongyi Cui
PT-LSTM: Extending LSTM for Efficient Processing Time Attributes

in Time Series Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
Yongqiang Yu, Xinyi Xia, Bo Lang, and Hongyu Liu
Loss Attenuation for Time Series Prediction Respecting Categories

of Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465
Jialing Zhang, Zheng Liu, Yanwen Qu, and Yun Li
PFL-MoE: Personalized Federated Learning Based on Mixture of Experts . . . 480

Binbin Guo, Yuan Mei, Danyang Xiao, and Weigang Wu
A New Density Clustering Method Using Mutual Nearest Neighbor . . . . . . . 487

Yufang Zhang, Yongfang Zha, Lintao Li, and Zhongyang Xiong
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495

Contents – Part II
Machine Learning 2
Unsupervised Deep Hashing via Adaptive Clustering. . . . . . . . . . . . . . . . . . 3

Shuying Yu, Xian-Ling Mao, Wei Wei, and Heyan Huang
FedMDR: Federated Model Distillation with Robust Aggregation . . . . . . . . . 18

Yuxi Mi, Yutong Mu, Shuigeng Zhou, and Jihong Guan
Data Augmentation for Graph Convolutional Network

on Semi-supervised Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Zhengzheng Tang, Ziyue Qiao, Xuehai Hong, Yang Wang, Fayaz Ali
Dharejo, Yuanchun Zhou, and Yi Du
Generating Long and Coherent Text with Multi-Level Generative

Adversarial Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Tianyi Tang, Junyi Li, Wayne Xin Zhao, and Ji-Rong Wen
A Reasonable Data Pricing Mechanism for Personal Data Transactions

with Privacy Concern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Zheng Zhang, Wei Song, and Yuan Shen
Knowledge Graph
A Probabilistic Inference Based Approach for Querying Associative

Entities in Knowledge Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
JianYu Li, Kun Yue, Jie Li, and Liang Duan
BOUNCE: An Efficient Selective Enumeration Approach for Nested

Named Entity Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Liujun Wang and Yanyan Shen
PAIRPQ: An Efficient Path Index for Regular Path Queries

on Knowledge Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Baozhu Liu, Xin Wang, Pengkai Liu, Sizhuo Li, and Xiaofei Wang
A Hybrid Semantic Matching Model for Neural Collective Entity Linking . . . 121
Baoxin Lei, Wen Li, Leung-Pun Wong, Lap-Kei Lee, Fu Lee Wang,
and Tianyong Hao
Multi-space Knowledge Enhanced Question Answering over

Knowledge Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Ye Ji, Bohan Li, Yi Liu, Yuxin Zhang, and Ken Cai
xxiv Contents – Part II
Emerging Data Processing Techniques
A Distribution-Aware Training Scheme for Learned Indexes. . . . . . . . . . . . . 143

Youyun Wang, Chuzhe Tang, and Xujia Yao
AIR Cache: A Variable-Size Block Cache Based on Fine-Grained

Management Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Yuxiong Li, Yujuan Tan, Congcong Xu, Duo Liu, Xianzhang Chen,
Chengliang Wang, Mingliang Zhou, and Leong Hou U
Learning an Index Advisor with Deep Reinforcement Learning. . . . . . . . . . . 178

Sichao Lai, Xiaoying Wu, Senyang Wang, Yuwei Peng,
and Zhiyong Peng
SardineDB: A Distributed Database on the Edge of the Network. . . . . . . . . . 186

Min Dong, Haozhao Zhong, Boyu Sun, Sheng Bi, and Yi Cai
DLSM: Distance Label Based Subgraph Matching on GPU . . . . . . . . . . . . . 194

Shijie Jiang, Yang Wang, Guang Lu, and Chuanwen Li
Information Extraction and Retrieval
Distributed Top-k Pattern Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

Xin Wang, Mingyue Xiang, Huayi Zhan, Zhuo Lan, Yuang He,
Yanxiao He, and Yuji Sha
SQKT: A Student Attention-Based and Question-Aware Model

for Knowledge Tracing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Qize Xie, Liping Wang, Peidong Song, and Xuemin Lin
Comparison Question Generation Based on Potential Compared

Attributes Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Jiayuan Xie, Wenhao Fang, Yi Cai, and Zehang Lin
Multimodal Encoders for Food-Oriented Cross-Modal Retrieval . . . . . . . . . . 253

Ying Chen, Dong Zhou, Lin Li, and Jun-mei Han
Data Cleaning for Indoor Crowdsourced RSSI Sequences . . . . . . . . . . . . . . 267

Jing Sun, Bin Wang, Xiaoxu Song, and Xiaochun Yang
Recommender System
A Behavior-Aware Graph Convolution Network Model

for Video Recommendation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
Wei Zhuo, Kunchi Liu, Taofeng Xue, Beihong Jin, Beibei Li,
Xinzhou Dong, He Chen, Wenhai Pan, Xuejian Zhang, and Shuo Zhou
Contents – Part II xxv
GRHAM: Towards Group Recommendation Using Hierarchical

Attention Mechanism. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Nanzhou Lin, Juntao Zhang, Xiandi Yang, Wei Song, and Zhiyong Peng
Multi-interest Network Based on Double Attention for Click-Through

Rate Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
Xiaoling Xia, Wenjian Fang, and Xiujin Shi
Self-residual Embedding for Click-Through Rate Prediction . . . . . . . . . . . . . 323

Jingqin Sun, Yunfei Yin, Faliang Huang, Mingliang Zhou,
and Leong Hou U
GCNNIRec: Graph Convolutional Networks with Neighbor Complex

Interactions for Recommendation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
Teng Mei, Tianhao Sun, Renqin Chen, Mingliang Zhou,
and Leong Hou U
Spatial and Spatio-Temporal Databases
Velocity-Dependent Nearest Neighbor Query . . . . . . . . . . . . . . . . . . . . . . . 351

Xue Miao, Xi Guo, Xiaochun Yang, Lijia Yang, Zhaoshun Wang,
and Aziguli Wulamu
Finding Geo-Social Cohorts in Location-Based Social Networks . . . . . . . . . . 368

Muhammad Aamir Saleem, Toon Calders, Torben Bach Pedersen,
and Panagiotis Karras
Modeling Dynamic Spatial Influence for Air Quality Prediction

with Atmospheric Prior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
Dan Lu, Le Wu, Rui Chen, Qilong Han, Yichen Wang, and Yong Ge
Learning Cooperative Max-Pressure Control by Leveraging Downstream

Intersections Information for Traffic Signal Control . . . . . . . . . . . . . . . . . . . 399
Yuquan Peng, Lin Li, Qing Xie, and Xiaohui Tao
Privacy-Preserving Healthcare Analytics of Trajectory Data . . . . . . . . . . . . . 414

Carson K. Leung, Anifat M. Olawoyin, and Qi Wen
Demo
PARROT: An Adaptive Online Shopping Guidance System . . . . . . . . . . . . . 423

Da Ren, Yi Cai, Zhicheng Zhong, Zhiwei Wu, Zeting Li, Weizhao Li,
and Qing Li
gStore-C: A Transactional RDF Store with Light-Weight Optimistic Lock . . . 429

Zhe Zhang and Lei Zou
xxvi Contents – Part II
Deep-gAnswer: A Knowledge Based Question Answering System . . . . . . . . 434

Yinnian Lin, Minhao Zhang, Ruoyu Zhang, and Lei Zou
ALMSS: Automatic Learned Index Model Selection System. . . . . . . . . . . . . 440

Rui Zhu, Hongzhi Wang, Yafeng Tang, and Bo Xu
GPKRS: A GPU-Enhanced Product Knowledge Retrieval System . . . . . . . . . 446

Yuming Lin, Hao Song, Chuangxin Fang, and You Li
Standard-Oriented Standard Knowledge Graph Construction

and Applications System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 452
Haopeng Ren, Yi Cai, Mingying Zhang, Wenjian Hao, and Xin Wu
Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459

Graph Mining
Co-authorship Prediction Based
on Temporal Graph Attention
Dongdong Jin1 , Peng Cheng1(B) , Xuemin Lin2 , and Lei Chen3

1
East China Normal University, Shanghai, China
ddjin@stu.ecnu.edu.cn, pcheng@sei.ecnu.edu.cn
2
University of New South Wales, Sydney, Australia
lxue@cse.unsw.edu.au
3
Hong Kong University of Science and Technology, Hong Kong, China
leichen@cse.ust.hk
Abstract. The social network analysis has received significant inter-

ests and concerns of researchers recently, and co-authorship prediction
is an important link prediction problem. Traditional models inefficiently
use multi-relational information to enhance topological features. In this
paper, we focus on the co-authorship prediction in the co-authorship
knowledge graph (KGs) to show that multi-relation graphs can enhance
feature expression ability and improve prediction performance. Cur-
rently, the main models for link prediction in KGs are based on KG
embedding learning, such as several models using convolutional neural
networks and graph neural networks. These models capture rich and
expressive embeddings of entities and relations, and obtain good results.
However, the co-authorship KGs have much temporal information in real-
ity, which cannot be integrated by these models since they are aimed at
static KGs. Therefore, we propose a temporal graph attention network to
model the temporal interactions between the neighbors and encapsulate
the spatiotemporal context information of the entities. In addition, we
also capture the semantic information and multi-hop neighborhood infor-
mation of the entities to enrich the expression ability of the embeddings.
Finally, our experimental evaluations on all dataset verify the effective-
ness of our approach based on temporal graph attention mechanism,
which outperforms the state-of-the-art models.
Keywords: Co-authorship prediction · Temporal KGs · Graph

attention
1 Introduction
Recently, social network analysis has attracted great interest from researchers,
such as social recommendation [31], influence analysis [9] and sentiment analysis
[25]. Co-authorship prediction [26] is an important problem of social network
analysis, i.e., predicting the possibility of collaboration between two authors in
the future, which can be regarded as academic user recommendation. Traditional
c Springer Nature Switzerland AG 2021
L. H. U et al. (Eds.): APWeb-WAIM 2021, LNCS 12858, pp. 3–19, 2021.
https://doi.org/10.1007/978-3-030-85896-4_1
4 D. Jin et al.
co-authorship prediction models [4,23] often ignore or underutilize the informa-

tion of various relations, such as the conference of papers, the topic of papers,
and the affiliation of authors. These information are very important for the co-
authorship prediction. For example, the authors in the same affiliation are more
likely to collaborate in future.
To utilize the relational information, we apply knowledge graphs on co-
authorship data. As a multi-relational graph widely used in many fields, knowl-
edge graphs (KGs) can integrate multiple relations to significantly improve per-
formance such as semantic analysis [16], question answering [15] and recommen-
dation systems [30]. A KG consists of triples in the form of (head entity, relation,
tail entity), where entities are nodes, a relation is a directed edge from the entity
ei to the entity ej and ei is called head entity as the starting point of the edge,
ej is called tail entity as the end of the edge. It is promising for co-authorship
prediction to use KGs considering various relations in a holistic way. Currently,
most existing works apply KG embedding learning for link prediction in KGs
[2,5,7], where entities and relations are mapped into the low-dimensional con-
tinuous embedding spaces through embedding models. Unfortunately, most of
them are time independent [5,22], while many real-world KGs (e.g., Wikidata
[10] and YAGO3 [21]) carry rich temporal information (e.g., the valid timestamp
of triples). With the temporal information, the performance of KG embedding
models in link prediction tasks can be effectively improved [6,11].
In this paper, we integrate the temporal information into KGs to predict the
potential co-author relations (collaborations) among authors, which can provide
valuable suggestions for the collaborations of papers or projects and facilitate
the exchange of ideas. To illustrate the effect of the temporal information on
the entities, we provide an example in Fig. 1. Each triple has its establishment
time, such as the time when two authors collaborated. With time information,
for author entity A1 , the difference in its relationships with two 1-hop neighbors
(A4 , A5 ) is easier to distinguish. We can find that author A1 was closer to author
A5 in 2013, and was closer to author A4 during 2015 to 2016. The future co-
authors of an author depend on his current state, such as the research fields,
affiliation and existing co-authors. For instance, author A1 and A5 published
papers at a same conference K1 in 2012 and 2010, respectively. Then, A1 and
A5 had collaboration in 2013 since they have similar research fields. Another
example is that in 2014 A3 had collaborations with A1 and A4 , which promoted
the subsequent collaborations between A1 and A4 in 2015 and 2016. In summary,
there are many interactions with temporal attributes among the entities.
Therefore, encapsulating the temporal information of triples of entities as
their embedding can generate better signals in model learning [11]. In order
to reasonably obtain the expressive embeddings, especially the embeddings of
author entities, we have the following challenges: 1) capturing the temporal inter-
actions between neighbors; 2) effectively aggregating representation of entities
at different distances (time and space) in KG.
Inspired by the success of the recurrent neural networks (RNNs) [14] in the
graph attention mechanism [34], we propose a novel temporal graph attention
Co-authorship Prediction Based on Temporal Graph Attention 5
Fig. 1. A knowledge graph of temporal co-authorship, where the double arrow indicates
the bidirectional relation between entities, the solid lines represent the actual relations.
network based on Bi-LSTM [12] to solve these challenges and obtain the embed-
dings of entities and relations. We first serialize the triples related to an entity,
then send them to the attention layer to compute the corresponding attention
coefficient and aggregate triples representation. When calculate the attention
coefficient of a neighbor j to the entity i, we encapsulate the influence of sub-
sequent neighbors (the interaction time with i is later) and previous neighbors
(the interaction time with i is earlier) on them. By stacking the temporal graph
attention layers, we can implicitly aggregate representation of entities at dif-
ferent distances in the KGs. In addition, we set up a constrained neighborhood
strategy to obtain the information of multi-hop entities more directly. Our model
is consisted of three parts: (a) considering the importance of the natural seman-
tics of the entities such as paper topic, we use a Semantic Encoder to initialize
the embeddings of entities; (b) the temporal graph attention network based
Embedding Encoder is used to learn the embeddings of entities and relations;
(c) multi-scale convolution kernels are added to ConvKB [5] as a Decoder to
predict the author’s future co-authors.
Our main contributions are as follows:
• We utilize the time and semantics information to enhance the expressive

ability of embeddings in KGs to predict future co-authors.
• We propose a temporal graph attention network to model the temporal inter-
actions between neighbors. Combining the temporal graph attention network
and the neighborhood strategy, we effectively aggregate the neighborhood
information of different distances into a low-dimensional space in Sect. 4.
6 D. Jin et al.
• We conduct extensive experiments on three real-world co-authorship datasets

in Sect. 5. The results show the clear superiority of our model compared with
several state-of-the-art models.
2 Related Work
Firstly, we review some of the existing work on co-authorship prediction. Path-
Predict [26] is the first work for heterogeneous co-authorship networks. It con-
structs several meta-paths between authors to capture topological features and
makes prediction through supervised learning. Existing work studies the impact
of different topological features between authors in medical co-authorship net-
works on the results, and uses supervised models to learn the best weights asso-
ciated with different topological features [23]. LDAcosin [4] incorporates the
content of the paper into the similarity measure between authors, and proposes
a new metric for evaluating similarity. However, the features extracted by these
models are relatively weak in expression, which is insufficient for high-quality
learning and prediction.
And the widely used knowledge graphs can obtain the powerful expression
for the author nodes through embedding learning. The existing KG embed-
ding learning models in the KGs are mainly aimed at static and temporal KGs.
We first introduce several excellent embedding learning models for static KGs.
TransE [2], DistMult [33], and ComplEx [27] are three translation-based models.
TransE is the first work to take the translation of entities and relations in the
embedding space as an optimization goal; DistMult is a simple expression of
bilinear model and ComplEx introduces the complex space into the KG embed-
ding. ConvE [7] and ConvKB [5] are two convolution-based models. They both
use two-dimensional convolution to mine the global relationships between enti-
ties and relations. R-GCN [24] and KBGAT [22] are two models based on graph
neural networks (GNNs) [1,18,29]. They use the neighborhood of each entity to
learn the embeddings. The difference between the two lies in the way of treat-
ing node neighbors. R-GCN assigns the same weight matrix to neighbors that
have the same relation type with an entity and KBGAT treats every neighbor
differently. Although the above embedding learning models have achieved good
results, none of them take into account the temporal information in the KGs.
Embedding learning models for temporal KGs are also important. TAE [17]
takes the time order as a constraint to improve the quality of embedding rep-
resentation. However, it does not use the exact time information of triples, but
mines the time order of relations from the occurrence time of triples. TTransE
[19] models the establishment and expiration time of the relations between the
head and tail entities, and predicts the valid time of unlabeled triples. TA-
TransE [11] strengthens the expressive ability of embeddings by learning the
time-aware representation of relation types. It can directly act on the current
common scoring functions for link prediction, and enable them to obtain better
performance and robustness on the temporal KGs. HyTE [6] explicitly merges
the time information into the space of entities and relations by associating each
time stamp with the corresponding hyperplane. These temporal KG embedding

models have excellent performance in link prediction, but the drawback is that
they are unable to model the temporal interaction between entities.
3 Problem Description
Given a temporal co-authorship KG G = [E, R], where E is the entity (node)
set, R is the relation (edge) set. G is composed of many temporal triples of
the form (h, r, t, time), which means that an interaction of type r has occurred
between entity h and entity t at time time. For example, in Fig. 1, the triple
(A1, W rite, A2, 2011) means that the authors A1 and A2 have formed a co-
author relation at time 2011. For the purpose of observing the time and fre-
quency of interactions between entities, unlike the general case where the same
relation between an entity pair is only considered once, we record all interactions
with time information. Link prediction in the temporal co-authorship KGs is to
predict whether a co-author relation will be established between two existing
authors in the future. We consider that author i and author j form a co-author
relation if and only if they jointly write and publish a paper.
In particular, we are only interested in possible co-author relation between
two authors who have never collaborated before, and repeated co-authors are
not within our forecast.
4 Our Approach
4.1 Semantic Encoder
For KG embedding learning, most methods of initializing entities, such as ran-

dom and TransE, ignore the semantic information of entities. We apply a pre-
trained language model BERT [8] to initialize the embeddings, which can gen-
erate word embeddings with semantics. Specifically, an entity e is defined by
a sequence of words, i.e., e = s = [w1 , . . . , wL ], where wl is the l-th word
of sequence s and L is the number of the words. We feed the word sequence
s into BERT to obtain the word embeddings, denoted as [w1 , . . . , wL ] =
BERT([w1 , . . . , wL ]), where wl ∈ RdBERT represents the word embedding of
wl and dBERT represents the dimension of the word embeddings. We apply an
averaging strategy and a fully connected layer FCs to get the low-dimensional
representation of s and use it as the initial embedding of e:
1
L
einit = FCs (Average(BERT(s))) = FCs ( wl ) (1)
L
l=1
8 D. Jin et al.
4.2 Embedding Encoder

Graph attention networks (GATs) [29] introduced the attention mechanism into
graph-structure data, and KBGAT better integrated the graph attention mech-
anism into KGs since GATs ignore the information of the relations. However, as
we mentioned before, their models are time-unaware and unable to model the
temporal information of triples.
Fig. 2. Semantic Encoder and Embedding Encoder with two temporal graph attention
layers. The form of the triples is reversed for convenience. The colored squares indicate
the embeddings of entities and relation at different layers in the two Encoders. The
arrows represent the input and output of each module or the operations on embeddings.
(Color figure online)
Therefore, we propose a temporal graph attention network to encode the

time-aware triples. We introduce a single temporal graph attention layer, and
the attention network can be implemented by simply stacking multiple attention
layers, as shown in Fig. 2. We show the process of updating the embedding of
entity t (indicated by the blue square) from the perspective of entity t as the
tail entity. The input of each layer is the embeddings of entities and relations,
and the output is the new embeddings learned by the layer.
To obtain the new embedding of an entity t, the attention coefficient will
be calculated for each triple associated with t (the triples with t as the tail
entity). We serialize the triples and denote them as V = [v1 , . . . , vY ], where
vy = [hi , rj , t, timea ], Y represents the number of triples in V and these triples
are sorted in ascending order of time. The process of calculating the normalized
attention coefficient of each triple in V is shown in Fig. 3. We first apply a linear
transformation to each triple vy with a weight matrix W1 :
vy = W1 [ehi erj et ] (2)
where represents concatenation operation, ehi , erj and et represent the embed-
dings of head entity hi , relation rj and tail entity t, respectively. The obtained
vector set is denoted as V = [v1 , . . . , vY ].
Fig. 3. The process of attention calculation. The input is a sequence of triples related
to the entity t, and the output is the normalized attention coefficient of each triple vy .
We choose Bi-LSTM [12] as an important part of the attention layer to

capture the information of the past and future triples and model the temporal
interactions between neighbors. A Bi-LSTM layer consists of two LSTM layers
[14], which calculate the forward sequence and the reverse sequence respectively,
called forward-layer and backward-layer. Specifically, − my−1 and ←
→ m−
y−1 are the
hidden states of forward-layer and backward-layer at the previous time-step y−1,
then the hidden states − my and ←
→ m− of two layers and the hidden state m of
y y
Bi-LSTM layer at the current time-step y are calculated as:
my = forward layer(vy , −
→
− →
my−1 )
m = backward layer(v , ←
←− −
m ) (3)
y y y−1
my = [−
m→←
y m−y ]
Then we obtain the attention value of the triple vy by a fully connected layer:
αvy = FCa (my ) = σ(W2 my + b) (4)
where the weight matrix W2 and the bias b are trainable parameters, σ is the
activation function such as LeakyReLU. Such a structure can provide the atten-
tion value with complete past and future contextual information. We apply a
softmax function to compute the normalized attention coefficient of triple vy :
exp(αvy )
βvy = softmax(αvy ) = Y (5)
n=1 exp(αvn )
The entity t obtains a new embedding representation by weighted aggregation

of the associated triples in V :
Y
et = σ( βvy vy ) (6)
y=1
To ensure the stability of the attention mechanism and focus on different aspects
of the neighborhood, similar to GATs, we employ the multi-head attention [28],
which can be described as:
Y
K

et = σ k k
βvy vy (7)
k=1 y=1
10 D. Jin et al.
where βvky , vky are calculated by the k-th independent attention head, K is the
number of attention heads, and represents the concatenation of the output of
each attention head. If the current temporal graph attention layer is the final
layer, then the entity embeddings output by the multi-head attention will not
be concatenated but averaged to achieve better results as done in GATs:

1 k k
K Y

et = σ βvy vy (8)
K y=1
k=1
To prevent the problem of gradient dispersion and degradation as well as

the disappearance of semantic information due to the deepening of the network
depth, we adopt the skip connection [13] in each attention layer, represented as:
et = W3 etinit + et (9)
where etinit is the embedding of the entity t output by the Semantic Encoder.
To facilitate the subsequent training process and model design, we use a weight
matrix W4 to transform the embedding space of the relations, where W4 is a
shared parameter, independent of a specific relation rj .
One of the advantages of graph attention networks is that they can naturally
get the information about multi-hop neighbors by stacking attention layers. For
example, as shown in Fig. 1, in the first attention layer, entities A1 and A4
aggregate the information of their 1-hop neighbors. And when A1 aggregates
the information of neighbor A4 again in the second layer, it gets the information
of all 1-hop neighbors of entity A4 , such as entity A6 . Therefore, a network with n
layers can get n-hop neighborhood information through iterative accumulation.
To get the multi-hop neighborhood information more directly, we artificially
construct the additional relations between an entity and it’s multi-hop neigh-
bors [20,22], for example, the dashed line between entity A1 and A6 in Fig. 1.
We set the time information of the additional relation to be the same as the
latest time on the multi-hop path. In addition, the most important entities and
relations in the temporal co-authorship KGs are the author entities and the co-
author relations. Therefore, it is unnecessary to construct additional relations
for all multi-hop neighbors. We develop a neighborhood strategy, the first is to
construct additional relations between each entity and all their 2-hop neighbors;
the second is that for an 3-hop neighbor, an additional relation is constructed
only if they are both author entities.
4.3 Decoder
Similar to KBGAT, we build the Decoder based on ConvKB [5]. However, unlike
KBGAT which uses ConvKB directly, we use the multi-scale convolution kernel
strategy instead of the same-scale strategy to capture more information of triples,
as shown in Fig. 4. Given a triple (h, r, t), eh , er and et represent the embedding
of h, r and t output by the Embedding Encoder, respectively. The input of the
Fig. 4. The process involved in our Decoder based on ConvKB (with the embedding
size d = 7, the number of multi-scale filter set M = 1)
Decoder is a matrix A = [eh , er , et ] ∈ Rd×3 , where d represents the embedding

dimension of entities and relations.
In the convolutional layer, we use three convolution kernels of different sizes,
denoted as Ω = [ω 1 ∈ R1×3 , ω 2 ∈ R3×3 , ω 3 ∈ R5×3 ]. The convolution ker-
nels convolve on the triple matrix independently to generate the feature maps
[τ 1 , τ 2 , τ 3 ], where τ i ∈ Rci ×1 represents the feature map output by ω i and ci
represents the dimension of τ i . After feature extraction by the convolutional
layer, the output feature maps are concatenated, i.e., p = [τ 1 τ 2 τ 3 ] ∈ Rc×1 ,
c = c1 + c1 + c3 .
We apply M convolution kernel sets to stabilize the process of feature extrac-
tion, denoted as Ω = [Ω1 , . . . , ΩM ] and the outputs of M filter sets are concate-
nated into a single vector, i.e., P = [p1 . . . pM ] ∈ RM c×1 . Finally, we define
the scoring function f of Decoder based on ConvKB as:
M
f (h, r, t) = σ([A ∗ Ωm + b]) · wd = P · wd (10)
m=1
where ∗ represents the convolution operator, b is the bias, wd is the weight

vector and · represents the dot product.
4.4 Optimization
Encoder. The basic idea of the translation-based model [2] is that given a
golden triple (h, r, t), the sum of the embedding eh of h and the embedding er
of r is as close as possible to the embedding et of t, i.e., eh + er ≈ et , where the
embeddings are learned by the model. Here the “close” can be measured with
d(h,r,t) =eh + er − et l1/l2 , where l1 represents L1-norm and l2 represents L2-
norm. We borrow this idea and minimize the loss LEncoder to train the Semantic
Encoder and Embedding Encoder jointly:

LEncoder = [d(h,r,t) − d(h ,r,t ) + γ]+ (11)
(h,r,t)∈Δ (h ,r,t )∈Δ
12 D. Jin et al.
where [x]+ represents Max[x, 0], γ is the safety margin distance, Δ and Δ indi-
cate positive and negative training sets, respectively. Additionally, a constraint of
the entity embeddings is added that the L2-norm is 1 (no such constraint on the
relations) to prevent the training process from trivially minimizing LEncoder [3].
Decoder. The Decoder optimizes model parameters by minimizing the soft-

margin loss:
λ
LDecoder = log 1 + exp l(h,r,t) · f (h, r, t) + wd 22 (12)
2
(h,r,t)∈{Δ∪Δ }
1 for(h, r, t) ∈ Δ
where l(h,r,t) = and λ
wd 22 is the L2 regularization on
−1 for(h, r, t) ∈ Δ 2
the weight vector.
5 Experiments
5.1 Datasets
Table 1. Datasets Statistics where |E| and |R| represent the number of entities and
relation types. And we give the number of authors and co-author relations in parenthe-
ses after the size of the entity set |E| and the size of the triple training set, respectively.
Triples
Datasets |E| |R|
Training Valid Test-full Test-2hop
Management 16327 (1241) 39952 (12314) 1450 640 464
Computer 18454 (2233) 9 54276 (17092) 2998 1104 888
Physics 59100 (13863) 365046 (126290) 5900 10786 7452
We verify the superiority of our model on three temporal co-authorship datasets.

We extract three academic relationships of different disciplines with a time span
of 2013–20161 , and process them into three temporal co-authorship KGs. We
delete the authors who publish only one paper in each dataset since they con-
tribute less to our task and are not within our prediction range. We use the
data of the first three years as the training sets, the new co-author relations of
the last year as the test sets, and filter the most recent part of the co-author
relations from the training set as the validation sets. We set up two test sets for
each dataset according to different purposes. One is a complete test set used to
predict all new collaborations, denoted as Test-full. And in the other, to avoid
1
http://cdblp.ruc.edu.cn/.
the excessive calculations between unrelated authors, we only test the new col-
laborations between two existing authors who are 2hop neighbors in the KGs,
denoted as Test-2hop. The number of positive and negative samples in all test
sets is balanced. For Test-full, we randomly construct non-existent co-author
relations as negative triples; for Test-2hop, we construct negative triples with
authors who are 2-hop neighbors to each other. The specific statistical informa-
tion of the datasets is given in Table 1.
5.2 Training Protocol
We use the “bern” method [32] to generate negative triples for training. Given
a golden triple, it sets different replacement probabilities for the head and the
tail to generate negative triples. This method implies the mapping property of
the relation r, i.e., 1-N, N-1, or N-N. The training process consists of two steps.
The first is jointly training the Semantic Encoder and the Embedding Encoder
to obtain the embeddings of entities and relations, and the second is training the
Decoder to predict future co-authors. The input of the Embedding Encoder is
the semantic embeddings of the entities and relations outputting by the Semantic
Encoder (the dimension size is 50), and its output is the new embeddings (the
dimension size is 200). The Embedding Encoder contains two temporal graph
attention layers, where each attention layer has four attention heads, and the
dropout probability set for each layer is 0.2. We choose ELU as the activation
function when aggregating neighborhood information. The margin distance in
the loss function and the training epoch are set to 5 and 4500 in the Physics,
1 and 3500 in other datasets. The input of the Decoder is the embeddings of
entities and relations learned by the Embedding Encoder, and the output is the
score of the triples. The dropout is set to 0.3, and the number of multi-scale
convolution kernel sets M is 50. As for training epoch, all datasets are set to
180 except for the Physics which is 300. We choose RELU as the activation
function, and the L2-regularizer λ in the loss function is 0.001. Finally, we use
Adam optimizer to update all model parameters, where the initial learning rate
is 1e−3 and the decay rate is 1e−5 .
5.3 Evaluation Protocol
The goal of future co-authorship prediction task is to predict whether the author
entities h and t will collaborate in the future, i.e., whether a triple (h, r, t) is
correct (r is fixed as the co-author relation). Specifically, if the score of a triple
output by the Decoder is higher than the threshold, then we consider it to be a
golden triple that will occur in the future, otherwise a wrong triple that will not
occur in the future. The threshold here is obtained by testing on the validation
set. We do training and testing separately on each dataset. Two main metrics
are compared for them, i.e., the accuracy of binary classification (ACC), and the
area under the ROC curve (AUC).
14 D. Jin et al.
5.4 Results and Analysis
We choose six models to compare with our model, where PathPredict [26] is
a model specifically aimed at co-authorship prediction, TransE [2], DistMult
[33], ComplEx [27] and KBGAT [22] are static KG embedding learning models
(without considering the temporal information) for link prediction, and TA-
TransE [11] is a temporal KG embedding learning model for link prediction.
Results on Test-Full. The results of all models on the Test-full of three

datasets are given in Table 2. The results show that our model has achieved
significant improvements on each dataset. For example, on the Management,
our model has an improvement of 2.65% on ACC compared to the sub-optimal
model (TA-TransE), and 2.23% on AUC compared to the sub-optimal model
(KBGAT). In addition, we find that the models based on GNNs (our model and
KBGAT) are obviously better than other models. For example, our model and
KBGAT have achieved the optimal and sub-optimal performance on the Com-
puter and Physics, which reflects the superiority of GNNs in graph structured
data and the role of graph attention mechanism in aggregating neighborhood
information. Table 2 shows that our model outperforms KBGAT in every com-
puted measure, which illustrates the importance of considering time informa-
tion and modeling the temporal interactions between neighborhood entities. As
a temporal KG embedding learning model, TA-TransE is not competitive com-
pared with other models such as TransE. Since most of the triples in the temporal
co-authorship KGs are co-author relations with time information and they are
always effective. Meanwhile the unique temporal interactions also makes the per-
formance of TA-TransE more mediocre. PathPredict achieves acceptable results
on Management, but performs poorly on the other two data sets. Although it
considers multiple types of nodes, and designs several meta-paths to capture the
topological features between author node pairs, the features are only between
two author nodes, without considering the influence of other authors on them.
We also conducted experiments on the unbalanced test set of Physics, with a
ratio of positive and negative samples of 1:10. The results in Table 2 show that
the performance of all models has improved, and our model maintains optimality.
Results on Test-2hop. If two author entities that have never collaborated are
2-hop neighbors (we call it related authors) such as they have all collaborated
with the same author, then they will be more likely to form a new co-author
relation. We summarize the comparison results among these models on the Test-
2hop of three datasets in Table 3. From the results, our model still achieves better
results in all metrics. In addition, the results on most datasets have a certain
degree of improvement compared with the Test-full, especially on the Physics.
Therefore, predicting relevant authors can achieve better results and is more
suitable in application.
Table 2. Experimental results on the Test-full of all datasets. All results are displayed
in percentage. The best score is in bold and second best score is underlined. Physics
represents the unbalanced test-full of Physics, where the ratio of positive and negative
samples is 1:10.
Management Computer Physics Physics

Methods
ACC AUC ACC AUC ACC AUC ACC AUC
PathPredict 90.16 93.83 90.31 92.52 81.32 85.42 89.65 87.02
TransE 88.91 92.33 92.03 94.38 83.06 86.74 92.97 86.85
DistMult 89.38 90.59 92.03 92.93 83.47 86.20 93.89 86.12
ComplEx 90.63 92.85 91.30 94.56 83.23 86.65 93.73 86.70
KBGAT 90.31 94.00 93.39 95.37 85.63 90.39 94.06 92.07
TA-TransE 90.94 93.03 92.39 94.40 83.91 88.48 92.56 88.91
Ours 93.59 96.23 95.11 97.22 86.98 92.10 94.76 94.30
Table 3. Experimental results on the Test-2hop of all datasets. All results are displayed
in percentage. The best score is in bold and second best score is underlined.
Management Computer Physics

Methods
ACC AUC ACC AUC ACC AUC
PathPredict 90.30 94.86 88.18 93.06 88.59 94.11
TransE 88.58 94.06 91.33 96.36 91.43 95.55
DistMult 88.36 94.23 90.77 93.61 91.68 95.89
ComplEx 90.30 94.18 88.63 91.10 91.73 95.84
KBGAT 91.59 95.91 93.47 96.50 92.37 96.78
TA-TransE 89.44 94.68 89.86 93.26 90.98 94.40
Ours 94.61 96.59 94.59 97.00 93.58 97.51
Fig. 5. The influence of the embedding dimension on the accuracy and training time
of the model. The experimental results come from Test-full of Physics.
16 D. Jin et al.
Embedding Dimension Test and Author’s Frequency Test. We study

the influence of the embedding dimension of entities and relations outputting by
the Embedding Encoder on the accuracy and the training time of our model.
We test the embedding dimensions of 50, 100, 200 and 400 on Physics, and the
results are shown in Fig. 5. From the figure, we can conclude that the larger the
embedding dimension is, the more expressive it is, which effectively improves the
capability of our model. However, when the embedding dimension reaches 200,
continuing to increase the embedding dimension no longer improves the accuracy
significantly, and the training time will increase dramatically.
We analyze the effect of the frequency of author collaboration on prediction
accuracy. We divide the authors in the test set into several groups according to
their current collaboration frequency, and give the test set size for each group.
The results are shown in Fig. 6. It should be noted that a test sample may be
divided into two groups since the two authors belong to different groups. There-
fore, the number of test samples in the figure will be larger than that in Table 1.
As a whole, as the number of current collaborations of an author increases, the
performance of predicting becomes worse. The results on other datasets are con-
sistent with Physics. This is reasonable as the number of potential co-authors of
the author also increases.
5.5 Ablation Study

Our model is a joint learning model composed of several modules. To explore
the contribution of each module, we conduct the following ablation studies from
three aspects. Table 4 shows the results of two evaluation metrics (ACC and
AUC) on the Computer and Physics.
• (Ours-SEMANTIC) We study the effectiveness of integrating the semantic

information of the entities into the embedding learning. We use the TransE
instead of the Semantic Encoder to generate the initial embeddings of the
entities. As shown in Table 4, our model achieves a more accurate prediction
Fig. 6. The influence of the author’s collaboration frequency on the accuracy. The
experimental results come from Test-full of Physics.
Table 4. Ablation learning results on computer and physics. All results are displayed
in percentage and the best score is in bold.
Management Computer Physics

Methods
ACC AUC ACC AUC ACC AUC
Ours-SEMANTIC 92.97 95.38 94.66 96.41 85.93 91.56
Ours-TGA 91.88 93.38 93.39 96.07 85.71 91.15
Ours-MSCK 92.66 94.91 94.29 95.67 86.01 91.69
Ours 93.59 96.23 95.11 97.22 86.98 92.10
than the variant, indicating that it is effective to consider the entity semantic
information.
• (Ours-TGA) We analyze the impact of the temporal graph attention mecha-
nism in the Embedding Encoder. We replace it with the attention mechanism
used in KBGAT. In Computer, the ACC value of our model is 95.11, almost
2 points higher than the variant, which reflects the importance of capturing
the temporal interactions between neighbors. In addition, the variant still
has better performance than KBGAT, which also directly demonstrates the
superiority of other modules in our model.
• (Ours-MSCK) Finally, we analyze the impact of the multi-scale convolution
kernel strategy in the Decoder. We fix the number of convolution kernels of
size ω ∈ R1×3 instead of multi-scale convolution kernels. From the results,
our model is significantly better than this variant, which demonstrates the
effectiveness of the multi-scale convolution kernel strategy in capturing multi-
dimensional features.
6 Conclusion and Future Work
In this paper, we study the problem of predicting possible future co-authors

between existing authors in the temporal co-authorship KGs. Aiming at the
temporal interactions among the neighbors of each entity, we propose a tem-
poral graph attention network to encapsulate the current state of an entity.
Moreover, we apply the extended temporal graph attention network and neigh-
borhood strategy to capture the multi-hop neighborhood information of entities.
Experimental results show that our model with multiple modules outperforms
other state-of-the-art models. In the future, we will make slight modifications to
this model and extend it to other types of temporal KGs.
Acknowledgements. Peng Cheng’s work is partially sponsored by Shanghai Pujiang

Program 19PJ1403300. Lei Chen’s work is partially supported by National Key
Research and Development Program of China Grant No. 2018AAA-0101100, the Hong
Kong RGC GRF Project 16207617, CRF Project C6030-18G, C1031-18G, C5026-
18G, AOE Project AoE/E-603/18, Theme-based project TRS T41-603/20R, China
NSFC No. 61729201, Guangdong Basic and Applied Basic Research Foundation
18 D. Jin et al.
2019B151530001, Hong Kong ITC ITF grants ITS/044/18FX and ITS/470/18FX,

Microsoft Research Asia Collaborative Research Grant, HKUST-NAVER/LINE AI
Lab, Didi-HKUST joint research lab, HKUST-Webank joint research lab grants.
References
1. Abu-El-Haija, S., et al.: MixHop: higher-order graph convolutional architectures
via sparsified neighborhood mixing. In: ICML, pp. 21–29 (2019)
2. Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., Yakhnenko, O.: Translating
embeddings for modeling multi-relational data. In: NeurIPS, pp. 1–9 (2013)
3. Bordes, A., Weston, J., Collobert, R., Bengio, Y.: Learning structured embeddings
of knowledge bases. In: AAAI, pp. 301–306 (2011)
4. Chuan, P.M., Ali, M., Khang, T.D., Dey, N., et al.: Link prediction in co-authorship
networks based on hybrid content similarity metric. Appl. Intell. 48(8), 2470–2486
(2018)
5. Dai Quoc Nguyen, T.D.N., Nguyen, D.Q., Phung, D.: A novel embedding model for
knowledge base completion based on convolutional neural network. In: NAACL-
HLT, pp. 327–333 (2018)
6. Dasgupta, S.S., Ray, S.N., Talukdar, P.: HyTE: hyperplane-based temporally aware
knowledge graph embedding. In: EMNLP, pp. 2001–2011 (2018)
7. Dettmers, T., Minervini, P., Stenetorp, P., Riedel, S.: Convolutional 2d knowledge
graph embeddings. In: AAAI, pp. 1811–1818 (2018)
8. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: pre-training of deep
bidirectional transformers for language understanding. In: NAACL, pp. 4171–4186
(2019)
9. Embar, V.R., Bhattacharya, I., Pandit, V., Vaculin, R.: Online topic-based social
influence analysis for the Wimbledon championships. In: SIGKDD (2015)
10. Erxleben, F., Günther, M., Krötzsch, M., Mendez, J., Vrandečić, D.: Introducing
Wikidata to the linked data web. In: ISWC, pp. 50–65 (2014)
11. Garcia-Duran, A., Dumančić, S., Niepert, M.: Learning sequence encoders for tem-
poral knowledge graph completion. In: EMNLP, pp. 4816–4821 (2018)
12. Graves, A., Schmidhuber, J.: Framewise phoneme classification with bidirectional
LSTM and other neural network architectures. Neural Netw. 18, 602–610 (2005)
13. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition.
In: CVPR, pp. 770–778 (2016)
14. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Comput. 9(8),
1735–1780 (1997)
15. Huang, X., Zhang, J., Li, D., Li, P.: Knowledge graph embedding based question
answering. In: WSDM, pp. 105–113 (2019)
16. Jain, N.: Domain-specific knowledge graph construction for semantic analysis. In:
Harth, A., et al. (eds.) ESWC 2020. LNCS, vol. 12124, pp. 250–260. Springer,
Cham (2020). https://doi.org/10.1007/978-3-030-62327-2 40
17. Jiang, T., Liu, T., Ge, T., Sha, L., Li, S., Chang, B., Sui, Z.: Encoding temporal
information for time-aware link prediction. In: EMNLP, pp. 2350–2354 (2016)
18. Kipf, T.N., Welling, M.: Semi-supervised classification with graph convolutional
networks. arXiv preprint arXiv:1609.02907 (2016)
19. Leblay, J., Chekol, M.W.: Deriving validity time in knowledge graph. In: WWW,
pp. 1771–1776 (2018)
20. Lin, Y., Liu, Z., Luan, H., Sun, M., Rao, S., Liu, S.: Modeling relation paths for
representation learning of knowledge bases. In: EMNLP, pp. 705–714 (2015)
21. Mahdisoltani, F., Biega, J., Suchanek, F.: YAGO3: a knowledge base from multi-
lingual Wikipedias. In: CIDR (2014)
22. Nathani, D., Chauhan, J., Sharma, C., Kaul, M.: Learning attention-based embed-
dings for relation prediction in knowledge graphs. In: ACL, pp. 4710–4723 (2019)
23. Qi, Y., Chao, L., Yanhua, L., Hongfang, S., Peifeng, H.: Predicting co-author rela-
tionship in medical co-authorship networks. PLOS ONE 9(7), e101214 (2014)
24. Schlichtkrull, M., Kipf, T.N., Bloem, P., van den Berg, R., Titov, I., Welling,
M.: Modeling relational data with graph convolutional networks. In: Gangemi, A.,
et al. (eds.) ESWC 2018. LNCS, vol. 10843, pp. 593–607. Springer, Cham (2018).
https://doi.org/10.1007/978-3-319-93417-4 38
25. Severyn, A., Moschitti, A.: Twitter sentiment analysis with deep convolutional
neural networks. In: SIGIR, pp. 959–962 (2015)
26. Sun, Y., Barber, R., Gupta, M., Aggarwal, C.C., Han, J.: Co-author relationship
prediction in heterogeneous bibliographic networks. In: ASONAM (2011)
27. Trouillon, T., Welbl, J., Riedel, S., Gaussier, É., Bouchard, G.: Complex embed-
dings for simple link prediction. In: ICML, pp. 2071–2080 (2016)
28. Vaswani, A., et al.: Attention is all you need. In: NeurIPS, pp. 5998–6008 (2017)
29. Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., Bengio, Y.: Graph
attention networks. In: ICLR (2018)
30. Wang, X., He, X., Cao, Y., Liu, M., Chua, T.S.: KGAT: knowledge graph attention
network for recommendation. In: SIGKDD, pp. 950–958 (2019)
31. Wang, X., Lu, W., Ester, M., Wang, C., Chen, C.: Social recommendation with
strong and weak ties. In: CIKM, pp. 5–14 (2016)
32. Wang, Z., Zhang, J., Feng, J., Chen, Z.: Knowledge graph embedding by translating
on hyperplanes. In: AAAI, pp. 1112–1119 (2014)
33. Yang, B., Yih, W., He, X., Gao, J., Deng, L.: Embedding entities and relations for
learning and inference in knowledge bases. In: ICLR (2015)
34. Zhang, C., Yao, H., Huang, C., Jiang, M., Li, Z., Chawla, N.V.: Few-shot knowledge
graph completion. In: AAAI, vol. 34, pp. 3041–3048 (2020)
Degree-Specific Topology Learning
for Graph Convolutional Network
Jiahou Cheng1 , Mengqing Luo1 , Xin Li3 , and Hui Yan1,2(B)

1
School of Computer Science and Engineering, Nanjing University of Science
and Technology, Nanjing 210094, China
yanhui@njust.edu.cn
2
The Key Laboratory of Intelligent Perception and Systems for High Dimensional
Information of Ministry of Education, Nanjing 210094, China
3
School of Internet of Things, Nanjing University of Posts and Telecommunications,
Nanjing 210003, China
Abstract. Graph Convolutional Networks (GCNs) have gained great

popularity in various graph data learning tasks. Nevertheless, GCNs
inevitably encounter over-smoothing with increasing the network depth
and over-fitting specifically on small datasets. In this paper, we present
an experimental investigation which clearly shows that the misclassifica-
tion rates of nodes with high and low degrees are obviously distinctive.
Thus, previous methods such as Dropedge that randomly removes some
edges between arbitrary pairwise nodes are distant from optimal or even
unsatisfactory. To bridge the gap, we propose the degree-specific topology
learning for GCNs to find a hidden graph structure that augments the
initial graph structure. Particularly, instead of training with uncertain
connections, we remove edges between the high-degree nodes and their
first-order neighbors with low similarity in the attribute space; besides,
we add edges between the low-degree nodes and their connectionless
neighbors with high similarity in the attribute space. Experiments con-
ducted on several popular datasets demonstrate the effectiveness of our
topology learning method.
Keywords: Degree-specific · Over-smoothing · Over-fitting ·

Dropedge
1 Introduction
Graphs or networks which consist of an edge set and a node set are widely used
to describe complex relationships among objects, such as citation relationship
among papers in citation networks, the relationship among users in social net-
work, etc. GCN [15] and their numerous variants, have shown to be successful
in graph representation learning tasks [1–4], such as node classification, graph
classification, and link prediction.
Despite their strong task performance, most of existing methods have two
major obstacles [5,6,9–11]: (1) Over-smoothing. GCNs transmit information
c Springer Nature Switzerland AG 2021
L. H. U et al. (Eds.): APWeb-WAIM 2021, LNCS 12858, pp. 20–34, 2021.
https://doi.org/10.1007/978-3-030-85896-4_2
Degree-Specific Topology Learning for Graph Convolutional Network 21
(a) (b)
Fig. 1. Receptive field of nodes with different degrees. Target node is colored in red,
and the 1-st, 2-nd, 3-rd, 4-rd order neighboring nodes are colored in orange, blue, green,
yellow and purple, respectively. (Color figure online)
through edges, and thus they can be approximately understood that the rep-
resentations of the central nodes are obtained by averaging their first-order
neighbor nodes. Since the information of low-order neighbor is not enough in
the shallow model, deep GCNs have merged, where more and more layers are
stacked. Unfortunately, by repeatedly applying Laplacian smoothing too often,
the final embeddings of all nodes will converge to the same values extremely. (2)
Over-fitting. GCNs, like other deep models, may suffer from over-fitting problem
which is generally caused by a model fitting a distribution with insufficient train-
ing data. It leads to fitting the training data very well but generalizing poorly
to the test data with such over-parameterized model learned.
There exist a variety of methods to address the aforementioned problems.
For example, [13] uses linear combination of the multi-order adjacency matrix
instead of the original adjacency matrix to expand the receptive field. IDGL [31]
adjusts the adjacency matrix iteratively and increases the range of the receptive
field by adding edges to learn the graph structure. [27] proposes that high-order
graph convolution can help model capture the global structure of graph. And
inspired by convolutional neural networks (CNNs), [26] builds deep GCN model
by adding residual or dilation convolution which construct graph in the attribute
space. Very recently, Dropedge [8] reduces part of the information transmission
by randomly selecting some edges to remove in each training procedure. As a
widely used method, Dropedge mitigates the over-fitting and over-smoothing
problems.
Since randomly removing some edges improves the performance, we take
the degree of the node into consideration and formulate a set of rules to add
and delete some edges, the performance of GCNs may be further improved. On
the one hand, low-degree nodes may possess high sparsity. Once their corre-
sponding edges are removed, the sharp reduction of their receptive field may
weaken the performance of the model. On the other hand, high-degree nodes are
more likely to be the authoritative or active individuals that have an important
effect in networks and communities, and are often bridges or hubs in connecting
different communities. Compared with low-degree nodes, their receptive fields
are larger, leading into more redundant and noisy neighbor nodes. Therefore,
Another random document with
no related content on Scribd:
“Did you see that, chief?” whispered Chick.
“Yes. Keep quiet. We want the papers. But we want him, too.”
“That’s what,” put in Patsy. “And that Keshub and the other coffee-
colored guy, too. There may be others in the house as well as them. There
are some maids, we know.”
“They are probably in another part of the house,” answered Nick. “We
need not trouble about the maids. What we want is this fellow, papers and
all. Keep ready!”
Ched Ramar stepped over the red lamp and looked carefully at the papers
he had got from the lap of the image. His sinister smile again spread over his
dark countenance, and he muttered to himself in his own tongue.
“This is all!” he suddenly exclaimed in English. “I will take these records
to Sang Tu in the morning. Meanwhile, they shall not leave me. I do not trust
any one. I will not go to bed. Such sleep as I need I can get here, in this
chair.”
He walked over to the chair in which the girl had sat. It was very large,
and when she had been in it had seemed actually to swallow her up. Even
Ched Ramar, tall as he was, had plenty of room to curl up in it.
He tried it in this way. Then he arose and strode over to the big idol, as if
to look behind it. Nick Carter, Chick, and Patsy were standing ready to fling
themselves upon him.
But he changed his mind, when nearly up to them, and contented himself
with calling sternly:
“Swagara!”
For a moment Patsy Garvan had forgotten his assumed name. He made
no move to go out. Instead, he held his automatic pistol ready to be used
either as a club or a firearm. Nick Carter brought him to himself with a sharp
tug at his elbow.
“Go out, confound you!” he whispered. “You are Swagara!”
“Gee! So I am!”
“Swagara!” called Ched Ramar, again, in a fiercer tone. “Come here!”
Patsy slipped out from behind the statue and made his Swagara bow with
due humility.
Ched Ramar raised his fist, as if he would bring it down on Patsy’s
shoulder. It was as well that he did not carry out his intention, for Patsy
surely would have forgotten his assumed character and retaliated with
another and harder blow.
“You deserve to be kicked, you dog!” snarled Ched Ramar. “You are to
come quickly when I call. But let that pass. You will keep awake in this
room till I tell you that you may sleep. Understand?”
Patsy bowed. He never had spoken more than a word or two to the
Indian. He had a presentiment that if ever he did so, he would be known as a
bogus Swagara at once.
“Very well,” went on Ched Ramar. “I would sleep for an hour—in this
chair. Keshub and Meirum are asleep in the hall without. They will not come
in unless I summon them. But you! You are not to sleep at all. Now, walk
over there to the large Buddha and let me see that you are quite awake now.
Go over and march back. Do as I bid you.”
Somehow, Patsy Garvan did not exactly understand what was meant by
this command, and he hesitated when he got to the idol. Turning toward
Ched Ramar, he was about to give him a pleading look, which would mean
that he wanted clearer instructions.
This angered Ched Ramar, and he bounded from the chair, drawing a
large jeweled scimitar that he generally wore, concealed by the folds of his
robe.
Flourishing this weapon, he flew at Patsy, as if he would strike him down
with it. The belligerent action was a great deal like his former one, only that
this time he held a deadly weapon, instead of merely menacing with his fist.
“Gee!” shouted Patsy, forgetting entirely the part he was playing. “If you
don’t drop that cheese knife, I’ll plug you as if you were a rat!”
He drew his pistol as he spoke and leveled it at the head of the surprised
Indian.
Instantly it occurred to the cunning mind of Ched Ramar that there was
treachery somewhere, and he leaped forward to seize Patsy. Rascal as he
might be, there was no cowardice about Ched Ramar.
He did not catch Patsy, however. Instead, the supposed Japanese suddenly
stooped, and just as the Indian got to him, he arose and sent his fist into the
brown neck.
Ched Ramar uttered a choking gasp, and dashed behind the Buddha. As
he got there, he found himself facing not only Patsy Garvan, but Nick Carter
and Chick, as well. All three were in hostile attitudes that could not be
mistaken for a moment.
CHAPTER XI.
THE CRASH OF THE IDOL.
The utter astonishment in the face of Ched Ramar when he saw these
three men where he had expected to find one only—and he a submissive
servant—made Patsy Garvan emit a shrill chuckle. Patsy never would hold
back his emotions when they got a good grip on him.
“Gee! Look at the map of him!” he shouted.
“Who are you?” roared Ched Ramar. “You’re not Swagara!”
“Not by a jugful!” returned Patsy Garvan. “There isn’t anything like that
in me. Say, chief! We want to work quick! There’s two more right outside
the door.”
Nick Carter stepped in front of the East Indian and held up his hand for a
chance to speak.
“Ched Ramar,” he said in his usual cool tones, “the game is up. You have
some papers in your pocket that you stole from Professor Matthew Bentham.
You got them with the help of the man you call Swagara, who is already my
prisoner.”
“Prisoner?” broke from Ched Ramar’s lips before he knew that he was
speaking. “Prisoner? Who are you?”
“My name is Nicholas Carter,” answered Nick.
“Nicholas Carter? Ah! Yes! I never saw you before. But your picture is in
our archives. We all know what you look like. If it had been lighter here, I
should have recognized you at once. Well, Mr. Nicholas Carter, all I have to
say to you is—this!”
The curved scimitar, with its richly jeweled hilt and its heavy, Damascus-
steel blade, swept through the air like a great half moon of fire, as it caught
and reflected the red glow of the lamp. The next moment, it circled Nick
Carter’s neck, and seemed as if it must actually sever his head from his body.
But the detective had been in critical situations of this kind before, and he
knew how to meet even an attack by such an unusual weapon as this cruel,
curved saber.
He stooped just in time. He had very little to spare, for the keen blade
caught the top of his soft hat and actually shaved away a thin sliver as clean
as if done by a razor. In fact, the convex edge of the scimitar was ground
almost to a razor edge.
The force of the blow made Ched Ramar swing around, so that he could
not recover himself immediately. Nick took advantage of this momentary
confusion to close with the tall Indian and grasp the handle of the saber.
There was a short and desperate struggle. The muscles of Ched Ramar
were as tough and flexible as Nick Carter’s, and the detective knew he had a
foe worthy of his best endeavors.
Up and down in the narrow space behind the big idol they fought, each
trying to gain possession of the scimitar.
Nick did not want to make noise enough to attract outside attention. But
he soon realized that this was something he could not prevent—the more so
as Ched Ramar seemed desirous of causing as much disturbance as possible.
A banging at the door explained why Ched Ramar had made as much
noise as he could.
“Now, Mr. Nicholas Carter,” hissed the tall Indian, “I think you will find
you have stepped into a trap. I have two men outside that door who will do
anything they are commanded, and never speak of it afterward. You have
been in countries where men are slaves to other men, I know. You shall see
what my men will do for me.”
During this speech, which was delivered jerkily, as the two struggled for
possession of the scimitar, the banging at the door increased in violence.
Chick and Patsy were against it on the inside, trying to prevent its being
battered down.
“Chick!” called Nick. “Come here!”
Chick looked over his shoulder.
“If I leave this door, Patsy can’t hold it by himself. It takes all we can
both do to hold those fellows back.”
“Never mind!” returned Nick. “Come here!”
As Chick came toward the two powerful fighters, Ched Ramar laughed
derisively.
“The door will fall,” he shouted. “When it does, you will wish you were
out of this place. I’m glad you are here. It is fortunate.”
He wrenched with tremendous energy to get the scimitar away from Nick
Carter. But the detective’s grip was not to be shaken. He held the handle of
the weapon at top and bottom, with the Indian’s two hands doubled around it
between. Neither could gain any advantage over the other.
“What am I to do?” queried Chick, looking at his chief, and making a
grab at the handle of the scimitar.
“Don’t bother with this,” directed Nick sharply. “Feel in the front of this
man’s robe and get the papers he has hidden there.”
“What?” bellowed Ched Ramar. “You’ll try such a thing as that? Ha, ha,
ha!” he laughed, as the door broke down, throwing Patsy Garvan to the floor.
“Get these men, Keshub! And you, Meirum! You did well to come! You
heard the noise? Yes? Now to your duty!”
Instantly there was a fray in which all six were engaged. The two guards
were nearly as strong as their employer, and all three of the Indians were
vindictive, and determined to be victorious.
“Get the one who is trying to rob me!” shouted Ched Ramar.
The two big guards rushed on Chick together, and with such sudden
violence that they hurled him away before he could set himself for
resistance.
“Look out, Patsy!” cried Chick. “Get those papers! The chief wants them!
Didn’t you hear him?”
“Did I hear him?” roared Patsy Garvan. “Well, I guess I did! Let me in
there!”
As Chick was hurled aside, Patsy rushed at Ched Ramar and sent his head
full into the Indian’s stomach. Patsy had had training in rough-and-tumble
warfare in the Bowery in his younger days, and he still remembered the
tricks that had availed him then.
The concussion was too much for Ched Ramar. It doubled him up, so that
Nick Carter got a better hold on the handle of the scimitar than he had been
able to obtain heretofore. At first he thought he had won the weapon
altogether. But Ched Ramar’s hold was too sure for that. He still retained his
grip, but not quite so good a one as he had had, because there was not so
much room for his fingers.
As Ched Ramar bent forward, still intent on not letting the scimitar out of
his grasp, Patsy reached in among the flowing robes that were flying in all
directions in the turbulence of the fight, and, after a little fumbling, felt the
end of the packet of papers sticking from an inner pocket.
“Got them!” he shouted, as he dragged out the papers and passed them to
Chick. “Gee! This is where we make the riffle!” cried Patsy delightedly.
“Hand them to the chief!”
Nick Carter shook his head quickly. He was holding Ched Ramar with
both hands.
“No! Keep them yourself, Chick, until I’ve got this man where I want
him. They’ll be safe enough now. Patsy, lay out that big fellow behind you
with your gun, before it is too late.”
Patsy employed a little ruse, and grinned as he saw how successful it was.
Turning swiftly, he presented his automatic pistol at the head of Meirum, and
there was a glint in the eye looking along the barrel which convinced the
man Patsy meant business.
As a result of his terror, Meirum backed away quickly, and let go of
Patsy’s arm, which he had seized as Patsy handed the papers to Chick.
On the instant, Patsy changed ends with his pistol, and brought the heavy
butt down on Meirum’s turbaned head with a crash that made nothing of the
white linen swathed about it. A turban is not much protection against a hard
blow with a steel-bound pistol butt.
As Meirum went down, there were only the two left—Keshub and Ched
Ramar.
“Take those papers, Keshub!” cried Ched Ramar. “Quick! Before he goes
away.”
“I’m not going away!” interposed Chick. “I’ve something else to do
before I go.”
He threw his arms suddenly around the big Keshub as he spoke, and
forced him backward.
“Pull that turban off the other fellow’s head!” he shouted to Patsy. “It will
make a good rope.”
This was a happy thought. Patsy unceremoniously stripped the white
turban from the head of the unconscious Meirum, and found himself with a
long strip of strong, white linen, which would, indeed, make a serviceable
rope.
But Keshub had not been overcome yet. He was almost as powerful as
Ched Ramar, and quite as full of fight. He tore himself out of Chick’s grasp
and rushed to the aid of his employer. The two of them set to work to get the
papers from Chick.
Nick Carter was equally resolved that Ched Ramar should not interfere
with Chick. He argued that Patsy Garvan and Chick were quite able to deal
with Keshub together—even if Chick could not do it alone.
“But Chick could do it himself,” he muttered. “Only that it might require
a little more time.”
It seemed as if Ched Ramar might have guessed what was passing in the
mind of Nick Carter, for he redoubled his efforts to get away, scimitar and
all, to go to the aid of his man.
“You may as well give up, Ched Ramar,” panted Nick Carter—for the
long fight was beginning to tell on his wind, just as it did on his foe’s.
“We’ve got you. We have the papers, and one of your men is done right here.
Another is a prisoner in my house. What is more, I know who you are.”
“I am Ched Ramar!” cried the Indian proudly.
“Perhaps. I don’t know what your name may be. The main thing is that
you are a member of the Yellow Tong, and that you are trying to steal these
papers for your chief, the infamous Sang Tu.”
“He is not infamous!” shouted Ched Ramar indignantly. “He is the
greatest man in the world to-day, and it will not be long before he will
control every nation on earth.”
“Beginning with the United States, I suppose?” exclaimed Nick Carter
ironically.
“Yes. We have this country of yours mapped out and given to different
sections of our great organization already,” snarled Ched Ramar. “As for
giving up, why—see here!”
He bent almost double, as he exerted every ounce of his immense
strength to tear the scimitar away from the detective. The latter felt the
handle slipping through his fingers. But he had strength, too, and in another
instant he had gained a firmer hold than ever, as he pushed with all his might
against the powerful bulk of his towering antagonist.
For a moment neither side gave way. It was like two mountains pressing
against each other. No one could say what the end might be. They might
stand thus for an indefinite period.
But they didn’t. Nick Carter felt his foe yield ever so little—not more
than a fraction of an inch. But the fact remained that he had given way
slightly, and Nick was quick to take advantage of anything that would help
him in such a desperate fight as this.
He pushed harder, and back went Ched Ramar two or three inches this
time.
“Keshub!” shouted Ched Ramar.
But Keshub had his own troubles just now. Chick had applied a backheel
to him, and was slowly pushing him backward, until he must fall flat on his
back, while Patsy hovered above them and grumbled because he couldn’t get
into the fight.
“Keep off, Patsy!” cried Chick. “Don’t come into this, or you’ll spoil it.
Don’t you see that?”
“Gee! I can see it, all right. But it’s mighty tough on me. I’ve been shut
out of this whole circus. When this is over, I’m a goat if I don’t go out and
hit a policeman. I’ve got to get action somehow.”
Nick Carter saw that he had Ched Ramar giving way now, and he
determined to make an end of the struggle without further waste of time. The
fight had been conducted very quietly. It had not even disturbed the two
maids, asleep upstairs, and there was no reason to suppose the fracas had
been heard on the street.
“You think you have me, I suppose?” hissed Ched Ramar, as he fought
with all the energy he had left.
Nick Carter did not answer. He knew that the cunning Indian was trying
to make him talk, and thus divert his attention. Instead, he gave his enemy a
sudden and harder twist that took him an inch farther back.
There was an inarticulate ejaculation of rage from the Indian, and his
black eyes glowed fiercely through his glasses. He stopped for a second the
onward rush of his assailant. Then, as he was obliged to give way, he jerked
up his arms and tried to bring the edge of the scimitar across Nick Carter’s
face.
The attempt failed, but it brought the battle to an abrupt end.
As Nick Carter leaped aside to avoid the scimitar, he kicked the feet of
Ched Ramar from under him. Back went the Indian, crashing against the
gigantic image of Buddha behind him.
For a moment the enormous idol rocked on its pedestal. Then, as it lost its
balance, down it came, pedestal and all, toward the two fighters!
One corner of the pedestal struck Nick Carter on the shoulder and laid
him out flat on his back.
He was not hurt, and he jumped to his feet on the instant. As he did so, he
shook his head—partly in satisfaction, but still more in horror.
The body of Ched Ramar lay under the great idol, and the brazen knees
were pressed into its victim’s head, crushing it out of all semblance to what
it had been!
Ched Ramar had paid the penalty of his rascality through the very agent
he had employed to make an innocent girl a participant in his crime.
“Look out, Chick!” shouted Nick Carter, as he saw Keshub breaking
away from his assistant’s hold. “He’s going to get out, if you don’t hurry.”
But Patsy Garvan was on the alert. He was only too glad to get into the
fight in any way, and he tripped Keshub, just as he leaped through the
doorway, in a very skillful and workmanlike manner.
“Oh, I guess not!” observed Patsy. “I saw you getting up after Chick had
laid you out, and I was looking for you to make a break like this. Come back
here!”
The cloth from Meirum’s turban was bound about Keshub, and he was
laid on the floor by the side of the knocked-out Meirum. Then, with
considerable exertion, the image of Buddha was rolled completely away
from the body of Ched Ramar, so that Nick could look it over with his flash
light.
“He died on the instant,” decided Nick. “Cover it with one of those
curtains, while I go downstairs and telephone the police station.”
In due course, the remains of Ched Ramar were viewed by the coroner,
and a verdict of “accidental death” was rendered.
Very little got into the papers about it. This was arranged by Nick Carter.
He did not want too much publicity while any of the Yellow Tong were still
likely to be active. It might interfere with work he had yet to do.
Keshub and Meirum, as well as Swagara, were not prosecuted. Nick
made up his mind that he could better afford to let them escape than to draw
general attention to the rascality they had been carrying on.
So he put them aboard a tramp steamer bound for Japan, and India, and
which would not touch anywhere until it got to Yokohama. Swagara was to
be put off there.
The next port would be Bombay. Both Keshub and Meirum said they
would never leave Indian soil again if once they could get back to it, and
there is no reason to suppose they were telling anything but the truth.
Matthew Bentham never knew the part his daughter had played in taking
and returning the precious papers. Nick Carter decided that no good end
would be served by letting him find it out.
Even Clarice herself was quite unaware of what she had done. The subtle
influence of hypnotism had permeated her whole being at the time, and
when she came to herself, it was entirely without recollection of what she
had passed through when in the power of another and stronger will.
Hypnotism is a wonderful science.
“Is this all of the Yellow Tong, chief?” asked Chick, smiling.
“There will be no end to this investigation until I have my hands on Sang
Tu,” replied Nick Carter sternly.
“I thought so,” was Chick’s reply.
THE END.
“The Doom of Sang Tu; or, Nick Carter’s Golden Foe,” will be the title of
the long, complete story which you will find in the next issue, No. 153, of
the Nick Carter Stories, out August 14th. In this story you will read of the
great detective’s ultimate triumph over the shrewd leader of the Yellow
Tong. Then, too, you will also find an installment of a new serial, together
with several other articles of interest.
Sheridan of the U. S. Mail.
By RALPH BOSTON.
(This interesting story was commenced in No. 148 of Nick Carter Stories. Back numbers can
always be obtained from your news dealer or the publishers.)
CHAPTER XXII.
A QUESTION OF COLOR.
After Owen had seen Jake Hines safely locked up in a local police
station, he went back to Dallas to fulfill this mission which had brought him
to Chicago. “I want you to explain to me about that letter you got from the
mail box,” he said. “You got the wrong letter by a mistake, of course?
Instead of the one which you had mailed to your brother, you got the pink
envelope which the Reverend Doctor Moore dropped into the box?”
“Yes,” answered Dallas, “when the letter carrier opened the box and took
out the mail, and I caught sight of that square, pink envelope lying on top of
the heap, I jumped to the conclusion that it was mine, and I grabbed it and
hurried away, fearing that he might change his mind about giving it to me.
You see, Owen, I was very much excited. The letter which I had received
from my scapegrace brother that day was very startling. It informed me that
he was in great trouble, and was about to do something desperate—the
letter didn’t state what—and that the only thing which could prevent him
from taking this step was my coming to Chicago immediately. It warned
me, too, that I mustn’t let a soul in New York know where I was going.”
“That was Hines’ work, of course,” said Owen. “He couldn’t come to
you in New York, so he contrived that scheme to bring you out to him.”
“Yes; but I didn’t suspect anything like that. I was very much worried.
From the tone of Chester’s letter I feared that he contemplated suicide, and I
was awfully scared. But I didn’t very well see how I could get out to him,
because”—she hesitated, and blushed painfully—“because I—I didn’t have
the fare, Owen. I had been sending more than I could spare to Chester
recently, to help him to get out of a scrape, and I was very hard up. So I had
to write him that I was very sorry, but I really couldn’t come to Chicago.”
“And then?” said Owen eagerly.
“Then, after I had mailed that letter, I suddenly thought of the
engagement ring which you had given to me, dear. I hated to pawn it, of
course, but I was so scared about Chester, and I—I thought you wouldn’t
mind, under the circumstances.”
“So that’s how you raised the fare to Chicago!” said Owen, with a smile
of great relief.
“Yes; and when I found that I could go, naturally I wanted to get back
that letter; for I feared the effect it might have upon my brother.”
“So you waited at the box until Pop Andrews came to collect the mail,
and you prevailed upon him to violate the rules and let you have it, and he
handed you the wrong letter,” said Owen. “So far, so good. And now,
Dallas, when you found that you had the Reverend Doctor Moore’s pink
envelope, with the hundred-dollar bill inside, what did you do with it?”
“When I got to my room at the boarding house, I started to tear the letter
up without opening it, still thinking, of course, that it was the one which I
had sent Chester. When I caught sight of the money inside, and realized the
mistake I had made, I was in a quandary. The hundred-dollar bill and the
letter which the envelope contained were each in four pieces. I was afraid to
go to the post office and explain how it had happened, because I knew that
if I did so it would get Carrier Andrews into trouble for violating the rules.
So I decided to cut some sticking plaster into small strips, and paste the
pieces together. I made quite a neat job of it; then I addressed a fresh
envelope, inclosed the patched-up letter and hundred-dollar bill, and
dropped it into a mail box.”
Owen drew a deep breath of relief. “And I suppose the envelope which
you addressed was a white one?”
“Yes. I didn’t have any pink ones at the boarding house.”
“And that explains, of course, why they thought at Branch X Y that the
letter was missing from the mail. Naturally they didn’t think to go through
the white envelopes. No doubt by this time the Reverend Doctor Moore’s
friend in Pennsylvania is in receipt of his hundred-dollar bill. Your
explanation, Dallas, clears the mystery! What a gink I am not to have
thought of that solution before!”
But suddenly a puzzled look came to Inspector Sheridan’s face. “There’s
one point that isn’t cleared up yet: If you got the wrong pink envelope,
Dallas, what became of the right one? The letter which you sent to your
brother ought to have been in the mail still.”
“And so it was,” answered Dallas, with a smile. “When I reached here I
found that Chester was already in receipt of it.”
“But how could that be? They searched all through the mail at Branch X
Y, and failed to find any square, pink envelope.”
“The letter which Chester received was in a square, white envelope,”
said Dallas. “I noticed that as soon as he showed it to me. And,” she went
on, with a puzzled frown, “that’s something which I can’t understand at all.
I know that it sometimes happens that in a box of colored stationery a white
envelope will get mixed with the tinted ones, but I am ready to take oath
that the envelope in which I inclosed Chester’s letter was pink. If it wasn’t
so perfectly ridiculous I should be inclined to believe that it must have
changed color while in the mail.”
“Wait a minute!” exclaimed Owen, an inspiration coming to him. “I
think I’ve got the answer. This envelope was exactly the same shape and
design as the rest in the box, wasn’t it, Dallas?”
“Yes; exactly the same as the others, except that it was white instead of
pink.”
“And it appeared to you to be pink?”
“Yes; and I am not color blind—if that is what you are going to imply,”
replied Dallas, mildly indignant.
“I’m not quite so sure of that,” said Owen, with a smile. “I’ll grant that
you are not color blind under ordinary conditions, but these were not
ordinary conditions.”
“What do you mean by that?”
“It was a dark afternoon when you addressed that envelope, and the
electric light over your desk at the office was turned on, wasn’t it?”
The girl nodded. “Yes, that’s so; but still——”
“And the electric globe over your desk throws such a strong light that
you have a piece of paper around it to shade it, haven’t you, Dallas? A piece
of red paper; I noticed it the other day.”
A look of enlightenment came to the girl’s face. “Why, yes; I understand
how it happened now. That red shade around the electric light made that
white envelope look pink, just like the rest.”
“Exactly!” cried Owen happily; “and that solves the mystery of the
missing pink envelope. I’m mighty glad now that I followed you to
Chicago, Dallas.”
CHAPTER XXIII.
UNTO THE LAST.
When Samuel J. Coggswell learned that his disciple and confidential

man, Jake Hines, had been brought back to New York under arrest, he was
greatly perturbed.
“And what does he say?” he asked the reporter who brought him the
news. “What does the misguided young man say? I suppose he has been
making some sensational and, of course, absolutely false statements about
me, eh?” He looked at his visitor anxiously.
“On the contrary,” the newspaper man replied, “they can’t get a thing out
of Jake, Mr. Coggswell. He refuses to talk.”
An expression of great relief came to the district leader’s face. “Ah!” he
exclaimed, his ears wiggling rapidly as he spoke. “Poor Jake, poor Jake! So
they can’t get a word out of him, eh? Jake always was a stubborn young
man—a very stubborn young man.”
After the newspaper man had gone, Boss Coggswell sat in his private
office at the clubhouse, smiling confidently to himself.
“I might have known Jake wouldn’t squeal,” he mused. “He’s not that
kind. Even though they’ve got him, I guess I’m safe.”
Even in the worst of men there is usually some redeeming trait. Crook,
grafter, and scoundrel as Jake Hines was, there was one thing which,
perhaps, should be put down to his credit—his unswerving loyalty to his
master.
The prosecuting attorney, certain that Samuel J. Coggswell was behind
the conspiracy against Owen Sheridan, which had landed Jake in the toils,
and anxious to get the bigger fish in his net, if possible, offered to deal
leniently with Hines if he would make a confession involving the boss. But
Jake stubbornly refused.
“No,” he said, “I ain’t convicted yet, and while Boss Coggswell’s my
friend I won’t give up hope of beatin’ this case. But if the worst comes to
worst, and I have to go up—well, I’ll be the goat. You won’t get a squeal
out of me!”
Coggswell made every effort to keep his subordinate from going to jail;
that is to say, every effort which it was possible to make in secret. He got a
bondsman for Jake, even though the latter’s bail was set at a very high
figure, and arranged for the young man to skip his bail and escape beyond
the jurisdiction of the courts before the case came up for trial.
But this plan was defeated by the vigilance of the prosecuting attorney,
who, anticipating such a move, had Hines watched so closely by detectives
that it was impossible for him to get away.
Failing in this attempt, Coggswell retained the very best lawyers
obtainable to defend his faithful follower; and when this array of legal
talent met with defeat, and Hines was found guilty by a jury, the politician
exerted all his powerful influence to save the convicted man from a jail
sentence. But this attempt also failed, and Jake Hines had to go to prison.
CHAPTER XXIV.
A SAD FAREWELL.
The young politician took his medicine with a stoicism worthy of a

better cause. There was actually a broad grin on his beefy face as he heard
the judge utter the words which condemned him to several years behind
prison bars. But it was not wholly stoicism. His attitude was partly due to
the fact that even at that desperate stage of the game he had not quite lost
faith in the power of his master and mentor to aid him.
“I won’t be in the jug long,” he declared confidently to the deputy sheriff
who led him, shackled, out of the courtroom. “Boss Coggswell will get me
out. His pull will win me a pardon, all right. So long as he’s my friend I’m
not worryin’. And not only will he get me free,” he added, a glint coming
into his beady eyes, “but you can bet he’ll make it hot for everybody that’s
had a hand in sending me up. That judge’ll get his for handing me such a
stiff sentence; the district attorney will be made to regret that he wouldn’t
let up when the boss gave him the hint; and as for that big stiff of a
Sheridan—well, I’m willing to bet a thousand to a hundred that he won’t be
holding that inspector’s job very long. They’ll all be made to feel that it
ain’t healthy to defy a man like Samuel J. Coggswell.”
Just as the train which was to carry him off to prison was about to pull
out of the station, Jake received a visit from the man in whom he had such
faith. Coggswell rarely yielded to sentiment when it was against his interest
to do so, but in this instance, although he realized that he could ill afford to
be seen shaking hands with the convicted man, he decided that the latter’s
loyalty in refusing to “squeal” was deserving of this tribute; so he was there
to say farewell to his faithful henchman.
“I need scarcely say,” he explained unctuously to the group of newspaper
men who were on the platform to see Hines depart, “that there is no man
who condemns and deplores more than I the atrocious crime for which that
wretched young man is about to pay the penalty. Still, I cannot quite forget
the time when poor, misguided Jake Hines was an honest man, who enjoyed
my esteem and friendship. It is in memory of those days, gentlemen, that I
am here now to give him a parting handclasp. Who knows,” he added,
raising his eyes piously toward the ceiling of the train shed, “but what the
lingering recollection of that last touch of an old friend’s hand may soften
his heart and cause his feet to seek once more the straight and narrow path
after he emerges from his gloomy prison cell?”
Having delivered himself of this worthy sentiment, and noting, with
satisfaction, that several of the scribes were taking it down verbatim, Mr.
Coggswell stepped aboard the train and approached the seat which
contained Jake and the deputy sheriff.
“How do you feel, my boy?” he inquired, in a sympathetic whisper.
“First class, boss,” Hines assured him, with a cheerful grin. “Say, it’s
mighty white of you to come to see me off, but you shouldn’t have done it.
It might cause talk.”
“Let evil tongues wag if they will,” was the sententious response. “You
ought to know me better, Jake, than to think for a moment that I would
consider myself at all in a case like this. I hope, my boy, that you are
accepting this unfortunate situation with philosophy and—er—are still
determined not to talk.”
“Don’t worry, boss,” said Hines, with another grin. “They’re not going
to get a word out of me, even though I have to stay in the jug for the full
term of my sentence. I’m no squealer.”
Hearing which, Coggswell exhaled a sigh of relief, and, as the train was
about to get under way, took a hurried leave of his unfortunate lieutenant.
“Boss,” Hines said to him wistfully, as they once more clasped hands,
“I’m sorry I won’t be there to help you at the coming primary fight. I’m
afraid you’ll miss me.”
“I’m afraid I shall, Jake,” Coggswell answered, taking care not to speak
above a whisper. “I’m afraid I shall.”
And his ears were not wiggling as he said it.
CHAPTER XXV.
THE LAST STAND.
Deprived of the services of his able lieutenant, Boss Coggswell faced the
coming primary-election contest with some misgivings. He realized that he
was up against the biggest battle of his political career.
Several times in the past attempts had been made to wrest the district
leadership from him, but in all those cases his opponents had been so weak,
and their campaigns so poorly organized, that he had been able to defeat
them without much effort. The Honorable Sugden Lawrence, he had reason
to believe, would prove a much more formidable foeman. The ex-judge
possessed a personality which made him an opponent to be feared even by
so powerful a boss as Samuel J. Coggswell. Therefore the latter had spoken
with the utmost sincerity when he told Jake Hines that he would miss him.
He feared that in order to win, much dirty work would have to be done; and
Boss Coggswell disliked dirty work—when he had to do it himself. It
would have been so much pleasanter to have the indefatigable Jake on hand
to take care of the hiring of “guerrillas,” the “fixing” of election inspectors,
and various other details of a similarly sordid and disagreeable character
which Jake had always taken care of so faithfully.
Perhaps it is needless to say that the enforced absence of his trusty helper
did not increase the boss’ good will toward the man who was directly
responsible for that calamity. Coggswell promised himself grimly that if the
primary election went his way Mr. Owen Sheridan’s chances of holding
down his job as post-office inspector wouldn’t be worth a plugged nickel.
True, Sheridan was protected—to some extent, by the civil-service laws;
but that fact did not worry Coggswell. He had his own little ways for
overcoming such obstacles.
It was not only a desire for vengeance which actuated him; fear and self-
preservation were also his motives. He considered it positively dangerous to
have Sheridan remain in the detective branch of the postal service, for there
were certain transactions past, present, and contemplated, with which he
was closely identified, which would not bear the scrutiny of a post-office
inspector.

Download ebook Web And Big Data 5Th International Joint Conference Apweb Waim 2021 Guangzhou China August 23 25 2021 Proceedings Part I Lecture Notes In Computer Science Leong Hou U Editor online pdf all chapter docx epub

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Download ebook Web And Big Data 5Th International Joint Conference Apweb Waim 2021 Guangzhou China August 23 25 2021 Proceedings Part I Lecture Notes In Computer Science Leong Hou U Editor online pdf all chapter docx epub

Uploaded by

Copyright:

Available Formats

Web and Big Data: 5th International

Joint Conference, APWeb-WAIM 2021,

Cambridge IGCSE and O Level History Workbook 2C - Depth

Pattern Recognition and Computer Vision 4th Chinese

Image and Graphics 11th International Conference ICIG

Intelligent Computing Theories and Application 17th

Web Information Systems Engineering – WISE 2021: 22nd

Quantitative Evaluation of Systems 18th International

Intelligent Computing Theories and Application 17th

Advances in Swarm Intelligence 12th International

Web and Big Data

Editorial Board Members

Yasushi Sakurai Junying Chen (Eds.)

Web and Big Data

ISSN 0302-9743 ISSN 1611-3349 (electronic)

© Springer Nature Switzerland AG 2021

July 2021 Yi Cai

Program Committee Chairs

Local Arrangement Chairs

APWeb-WAIM Steering Committee Representative

Senior Program Committee Members

Program Committee Members

Defu Lian University of Electronic Science and Technology

Panagiotis Karras Aarhus University, Denmark

Yongpan Sheng Chongqing University, China

David R. Cheriton School of Computer Science, University of Waterloo, Canada

Abstract. Resource Description Framework (RDF) has been proposed for

School of Computing and Augmented Intelligence,

School of Computer Science, Fudan University, China

School of Computing, National University of Singapore, Singapore

Abstract. Given a graph G, network embedding maps each node in G into a

Co-authorship Prediction Based on Temporal Graph Attention . . . . . . . . . . . 3

Degree-Specific Topology Learning for Graph Convolutional Network . . . . . 20

Simplifying Graph Convolutional Networks as Matrix Factorization . . . . . . . 35

GRASP: Graph Alignment Through Spectral Signatures. . . . . . . . . . . . . . . . 44

FANE: A Fusion-Based Attributed Network Embedding Framework . . . . . . . 53

What Have We Learned from OpenReview?. . . . . . . . . . . . . . . . . . . . . . . . 63

Unsafe Driving Behavior Prediction for Electric Vehicles. . . . . . . . . . . . . . . 80

Resource Trading with Hierarchical Game for Computing-Power

Analyze and Evaluate Database-Backed Web Applications with WTool . . . . . 110

Semi-supervised Variational Multi-view Anomaly Detection . . . . . . . . . . . . . 125

A Graph Attention Network Model for GMV Forecast on Online

Suicide Ideation Detection on Social Media During COVID-19 via

An Efficient Bucket Logging for Persistent Memory . . . . . . . . . . . . . . . . . . 149

Data Poisoning Attacks on Crowdsourcing Learning . . . . . . . . . . . . . . . . . . 164

Dynamic Environment Simulation for Database Performance Evaluation . . . . 180

LinKV: An RDMA-Enabled KVS for High Performance and Strict

Cheetah: An Adaptive User-Space Cache for Non-volatile Main Memory

Topic Model and Language Model Learning

Chinese Word Embedding Learning with Limited Data . . . . . . . . . . . . . . . . 211

Sparse Biterm Topic Model for Short Texts . . . . . . . . . . . . . . . . . . . . . . . . 227

EMBERT: A Pre-trained Language Model for Chinese Medical

Self-supervised Learning for Semantic Sentence Matching with Dense

An Explainable Evaluation of Unsupervised Transfer Learning for Parallel

Leveraging Syntactic Dependency and Lexical Similarity for Neural

A Novel Capsule Aggregation Framework for Natural Language Inference. . . 300

Learning Modality-Invariant Features by Cross-Modality Adversarial

Difficulty-Controllable Visual Question Generation . . . . . . . . . . . . . . . . . . . 332

Incorporating Typological Features into Language Selection

Removing Input Confounder for Translation Quality Estimation

Learning Refined Features for Open-World Text Classification . . . . . . . . . . . 367

Emotion Classification of Text Based on BERT and Broad

Improving Document-Level Sentiment Classification with User-Product

Integrating RoBERTa Fine-Tuning and User Writing Styles for Authorship