Welcome to the DASFAA-2022 Virtual Conference, April 11-14, 2022, Hyderabad, India
The DASFAA-2022 conference is held in online mode. All timings are as per Indian Standard Time (IST) (UTC + 05:30).
We are hosting the conference on Airmeet virtual event platform.

Instructions to join the event:

Workshop Schedule
April 11, 2022 (Monday)
Timings	Hall 1	Hall 2	Hall 3	Hall 4	Hall 5
9.00-13.00	IWBT-2022 International Workshop on Blockchain Technologies (IWBT2022) Organizers: Sanjay Chaudhary (Ahmedabad University, India), Krishnasuri Narayanam (IBM Research, India)	MAQTDS-2022 Managing Air Quality Through Data Science (MAQTDS 2022) Organizers: Girish Agrawal (O.P. Jindal Global University, India), Jai Ganesh (Mphasis Ltd., India)	PMBD-2022 1^st Workshop on Pattern mining and Machine learning in Big complex Databases (PMBD 2022) Organizers: Philippe Fournier-Viger (Shenzhen University, China), Mourad Nouioua (Harbin Institute of Technology Shenzhen, China), Hamido Fujita (Iwate Prefectural University, Japan), Lin Zhang (Tencent.com), Vincent S. Tseng (National Chiao Tung University, Taiwan)	BDMS-2022 The 8^th International Workshop on Big Data Management and Service (BDMS 2022) Organizers: Xiaoling Wang (East China Normal University, China), Kai Zheng (University of Electronic Science and Technology of China), An Liu (Soochow University, China)	BDQM-2022 The 7^th International Workshop on Big Data Quality Management (BDQM 2022) Organizers: Xiaoou Ding (Harbin Institute of Technology, China), Xueli Liu (Tianjin University, China)
13.00-14.00	Lunch break
14.00-18.00	IWBT2022 (Afternoon session)	MAQTDS 2022 (Afternoon session)	GDMA-2022 The 6^th International Workshop on Graph Data Management and Analysis (GDMA 2022) Organizers: Lei Zou (Peking University, China)

Main Conference Schedule (April 12-14, 2022)
April 12, 2022 (Tuesday)
Timings	Hall 1	Hall 2	Hall 3	Hall 4	Hall 5
08.30-09.00	Inauguration (Hall 1) Lei Chen (Hong Kong University of Science and Technology, Hong Kong), P. J. Narayanan (IIIT Hyderabad, India), S. Sudarshan (IIT Bombay, India), Masaru Kitsuregawa (University of Tokyo, Japan), P. Krishna Reddy (IIIT Hyderabad, India), Mukesh Mohania (IIIT Delhi, India), Anirban Mondal (Ashoka University, India), Arnab Bhattacharya (IIT Kanpur, India), Lee Mong Li Janice (National University of Singapore, Singapore), Divyakant Agrawal (University of California, Santa Barbara, USA), Prasad M. Deshpande (Google), Rajeev Gupta (Microsoft, India)
09.00-10.00	Keynote by Gautham Das (University of Texas at Arlington) (Hall 1) Title: Fairness in Database Querying Chair: Mukesh Mohania (IIIT, Delhi)
10.00-10.15	Tea break
10.15-12.30	Research Session #1 (Queries) Chair: P Radha Krishna (NIT Warangal, India)	Research Session #2 (Text and Image-I) Chair: Anil Kumar Vuppala (IIIT Hyderabad, India)	Research Session #3 (Applications of ML-I) Chair: Rajeev Gupta (Microsoft)	Industry Session #1 (Advancing Recommendation Systems) Chair: Prasad Deshpande (Google)	Tutorial Session #2 (Title: Reachability on Large-scale Graphs: Models, Techniques, and Trends) Chair: Lini Thomas (IIIT Hyderabad)
12.30-13.30	Lunch break
13.30-14.45	Demo and PhD Consortium (Hall 1) Chair: Anirban Mondal (Ashoka University, India)
14.45-16.30	Research Session #4 (Graphs-I) Chair: Vasudha Bhatnagar (University of Delhi)	Research Session #5 (Text and Image-II) Chair: Arvind Agarwal (IBM)	Research Session #6 (Recommendation-I) Chair: Manish Singh (IIT Hyderabad)		Tutorial Session #1 (Title: Make Wise Decisions for Your DBMSs: Workload Forecasting and Performance Prediction Before Execution) Chair: Vikram Pudi (IIIT Hyderabad)
16.30-17.00	Tea break
17.00-18.30	Panel Session (Hall 1) Moderator: Kurt Stockinger (Zurich University of Applied Sciences, Switzerland) Panelists: Georgia Koutrika (Athena Research, Greece), Jaydeep Sen (IBM Research, India), Immanuel Trummer (Cornell University, USA) , Lei Zou (Peking University, China) Chair: Jayant Haritsa (Indian Institute of Science, Bangalore, India)

April 13, 2022 (Wednesday)
Timings	Hall 1	Hall 2	Hall 3	Hall 4	Hall 5
09.00-10.00	Keynote by Tirthankar Lahri (Oracle Corporation) (Hall 1) Title: Oracle Database In-Memory: The Enterprise, at Warp Speed! Chair: Masaru Kitsuregawa (University of Tokyo, Japan)
10.00-10.15	Tea break
10.15-12.30	Research Session #7 (Spatio-temporal Data) Chair: Uday Kiran, (The University of Aizu, Japan)	Research Session #8 (Recommendation-II) Chair: Vikram Goel (IIIT Delhi)	Research Session #9 (Applications of ML-II) Chair: Ranganath Kondapally (Microsoft)	Industry Session #2 (Data Management and Search) Chair: Rajasekar Krishnamurthy (Adobe)	Tutorial Session #3 (Title: A tutorial on biomedical image segmentation using deep learning) Chair: Sudipta Banerjee (IIIT Hyderabad, India)
12.30-13.30	Lunch break
13.30-14.30	Keynote by Ioana Manolescu (Inria and Institut Polytechnique de Paris) (Hall 1) Title: Teasing journalistic findings out of heterogeneous sources: a data/AI journey Chair: Srinath Srinivasa (IIIT Bangalore, India)
14.30-16.15	Research Session #10 (Algorithms-I) Chair: Anirban Mondal (Ashoka University)	Research Session #11 (Systems) Chair: Satish Narayana Srirama (University of Hyderabad)	Research Session #12 (Applications of ML-III) Chair: Ranganath Kondapally (Microsoft)		Tutorial Session #4 (Title: AI Meets NoSQL Database: Methods, Opportunities and Challenges) Chair: Kamal Karlapalem (IIIT Hyderabad)
16.15-16.30	Tea break
16.30-18.00	Research Session #13 (Security) Chair: Ashok Kumar Das (IIIT Hyderabad)	Research Session #14 (Text and Image-III) Chair: Anil Kumar Vuppala (IIIT Hyderabad, India)	Research Session #15 (Graphs-II) Chair: Tanmoy Chakraborty (IIIT Delhi)

April 14, 2022 (Thursday)
Timings	Hall 1	Hall 2	Hall 3	Hall 4	Hall 5
09.00-10.00	Keynote by Sunita Sarawagi (IIT Bombay) (Hall 1) Title: Modern AI for Age-old problems of Database Systems Chair: Sang-Won Lee (SKKU, Korea)
10.00-10.15	Tea break
10.15-12.30	Research Session #16 (Algorithms-II) Chair: Uday Kiran, (The University of Aizu, Japan)	Research Session #17 (Recommendation-III) Chair: Anirban Mondal (Ashoka University, India)	Research Session #18 (Applications of ML-IV) Chair: Vikram Goel (IIIT Delhi)	Industry Session #3 (Industrial Machine Learning Applications) Chair: Daxin Jiang, (Micorosoft)	Tutorial Session #5 (Title: Time Series Anomaly Detection Toolkit for AI Applications) Chair: Praveen Paruchuri (IIIT Hyderabad, India)
12.30-13.30	Lunch break
13.30-14.30	Keynote by Guoliang Li (Tsinghua University, Beijing, China) (Hall 1) Title: openGauss: An Autonomous Database System Chair: Arnab Bhattacharya (IIT Kanpur, India)
14.30-16.15	Research Session #19 (Knowledge Bases) Chair: V. Raghava Mutharaju (IIIT Delhi)	Research Session #20 (Text and Image-IV) Chair:	Research Session #21 (Recommendation-IV) Chair: V.Ravi (IDRBT, Hyderabad)
16.15-16.30	Concluding Remarks (Hall 1) Chair: P.Krishna Reddy (IIIT Hyderabad)

Program Details

1. Keynote Speakers

Prof. Gautam Das

Director of Center for Artificial Intelligence and Big Data (CARIDA) and Database Exploration Laboratory (DBXLAB),
University of Texas at Arlington

Title:

Fairness in Database Querying.
April 12, 2022 (Tuesday), 09.00-10.00, Hall 1

Abstract:

We are being constantly judged by automated decision systems that have been criticized for being sometimes discriminatory and unfair. In this talk, we focus on fairness issues that arise when users perform ad-hoc exploration of databases using commonly available querying mechanisms such as selection/range queries, ranking queries, top-k queries, etc. For example, a user may use such queries to retrieve suitable employment opportunities in a jobs database, dating partners in a matching website, or apartments to rent in a real estate database. We will discuss how such querying mechanisms can give sometimes give results that are discriminatory, and discuss approaches to detect, mitigate and prevent such scenarios from occurring. Our work represents some of the initial steps towards the broader goal of integrating fairness conditions into database query processing and data management.

Bio:

Dr. Das is the Associate Dean for Research, College of Engineering, a Distinguished University Chair Professor of Computer Science and Engineering, Director of the Center for Artificial Intelligence and Big Data (CARIDA), and Director of the Database Exploration Laboratory (DBXLAB) at UT-Arlington. Prior to joining UTA in 2004, he has held positions at Microsoft Research, Compaq Corporation and the University of Memphis. He graduated with a B.Tech in computer science from IIT Kanpur, India in 1983, and with a Ph.D in computer science from the University of Wisconsin, Madison in 1990. He is a Fellow of the IEEE and a member of the ACM.

Dr. Das has published over 200 papers, many of which have appeared in premier data mining, database and algorithms conferences and journals. His work has received several awards, including the Communications of the ACM Research Highlights in 2021, ACM SIGMOD Research Highlights in 2019, IEEE ICDE 10-Year Influential Paper Award in 2012, ACM SIGKDD Doctoral Dissertation Award (honorable mention) in 2014 for his former student, and numerous other awards. He has presented keynotes and invited lectures, tutorials and courses at various universities, research labs, and conferences. He has been on the Editorial Board of the journals ACM Transactions on Database Systems and IEEE Transactions on Knowledge and Data Engineering. He has served in the organization roles of several major conferences, including as General Chair of ACM SIGMOD/PODS 2018.

Tirthankar Lahiri

Senior Vice President,
Data and In-Memory Technologies,
Oracle Corporation.

Title:

Oracle Database In-Memory: The Enterprise, at Warp Speed !
April 13, 2022 (Wednesday), 09.00-10.00, Hall 1

Abstract:

In-memory computing is more than simply about speed. It enables a fundamental transformation in business processes. Just as air travel enabled more than just the ability to travel faster: It enabled a completely new global economy, reshaped politics, and transformed society. Oracle's Database In-Memory feature similarly enables not just faster analytics, but a fundamental rethinking and drastic simplification of the traditional analytic platform. Combined with Oracle's many converged database capabilities that bring together many data models and many workloads, and with Oracle's Autonomous Database platform that makes self-driving machine-learning powered databases a reality, Oracle Database In-Memory allows for the development of a new category of enterprise architectures, with significant reduction in cost and complexity, while providing unmatched performance for both transactional and analytic workloads.

Bio:

Tirthankar Lahiri is Senior Vice President of the Data and In-Memory Technologies area within Oracle Database. This includes the Oracle Database Engine (Transactions, Data formats, Indexes, Advanced Compression, Database In-Memory, the Database Filesystem, etc.), the Oracle TimesTen In-Memory Database, and Oracle NoSQLDB. Tirthankar has 26 years of experience in the database industry and has worked on a number of areas such as Performance, Scalability, Manageability and In-Memory architectures. He has 45 issued and several pending patents and a number of academic publications. He has a B.Tech in Computer Science from the Indian Institute of Technology (Kharagpur) and an MS in Electrical Engineering from Stanford University. He was in the PhD program at Stanford and his research included NUMA Operating Systems (the Hive project) and Semistrucured Data (the Ozone project) before his PhD was superceded by his industrial career.

Ioana Manolescu

Inria and Institut Polytechnique de Paris.

Title:

Teasing journalistic findings out of heterogeneous sources: a data/AI journey.
April 13, 2022 (Wednesday), 13.30-14.30, Hall 1

Abstract:

Freedom of the press is under thread worldwide, and the quality of information that people have access to is dangerously degraded, under the joint threat of non-democratic governments and fake information propagation. The press as an industry needs powerful data management tools to help them interpret the complex reality surrounding us.

Since 2018, I have been cooperating with journalists from Le Monde, France's leading newspaper, in devising tools for analyzing large and heterogeneuos data sources that they are interested in. This research has been embodied in ConnectionLens, a graph ETL tool capable of ingesting heterogeneous data sources into a graph, enriched (with the help of ML methods) with entities extracted from data of any type. On such integrated graphs, we devised novel algorithms for keyword search, and combine them in more recent research with structured querying. The talk describes the architecture and main algorithmic challenges in building and exploiting ConnectionLens graphs, illustrated in particular on an application where we study conflicts of interest in the biomedical domain. This is joint work with A. Anadiotis, O. Balalau, H. Galhardas and many others. ConnectionLens Web site (papers+code): https://team.inria.fr/cedar/connectionlens/

This research has been funded by Agence Nationale de la Recherche AI Chair SourcesSay (https://sourcessay.inria.fr)

Bio:

Ioana Manolescu is a senior researcher at Inria Saclay and a part-time professor at Ecole Polytechnique, France. She is the lead of the CEDAR INRIA team focusing on rich data analytics at cloud scale. She is also the scientific director of LabIA, a program ran by the French government whereas AI problems raised by branches of the local and national French public administration are tackled by French research teams. She is a member of the PVLDB Endowment Board of Trustees, and has been Associate Editor for PVLDB, president of the ACM SIGMOD PhD Award Committee, chair of the IEEE ICDE conference, and a program chair of EDBT, SSDBM, ICWE among others. She has co-authored more than 150 articles in international journals and conferences and co-authored books on "Web Data Management" and on "Cloud-based RDF Data Management". Her main research interests algebraic and storage optimizations for semistructured data, in particular Semantic Web graphs, novel data models and languages for complex data management, data models and algorithms for fact-checking and data journalism, a topic where she is collaborating with journalists from Le Monde. She is also a recipient of the ANR AI Chair titled "SourcesSay: Intelligent Analysis and Interconnexion of Heterogeneous Data in Digital Arenas" (2020-2024).

Prof. Sunita Sarawagi

Institute Chair Professor, Computer Science and Engineering, IIT Bombay.

Title:

Modern AI for Age-old problems of Database Systems.
April 14, 2022 (Thursday), 09.00-10.00, Hall 1

Abstract:

Modern deep learning methods are pushing the frontiers of many challenging problems in database systems. We will discuss state-of-the-art machine learning models that are providing record breaking accuracies on age-old tasks such as entity resolution, missing value imputation and natural language querying. We are also witnessing brand new capabilities that were not possible a few years back. We can perform entity resolution across heterogeneous, multilingual datasets via actively learned nearest neighbor indices, thereby eliminating the need for hand-designing blocking predicates. On multi-dimensional analytical datasets, we can now obtain joint distributions over thousands of interacting time series. Advances in pre-trained language models have significantly increased the capability of handling natural variations in parsing text input to SQL. In this talk we will go over the latest ML research that is enabling these capabilities, and present directions for future research.

Bio:

Sunita Sarawagi researches in the fields of databases and machine learning. She is institute chair professor at IIT Bombay. She got her PhD in databases from the University of California at Berkeley and a bachelors degree from IIT Kharagpur. She has also worked at Google Research (2014-2016), CMU (2004), and IBM Almaden Research Center (1996-1999). She was awarded the Infosys Prize in 2019 for Engineering and Computer Science, and the distinguished Alumnus award from IIT Kharagpur. She has several publications including best paper awards at ACM SIGMOD, VLDB, ICDM, NIPS, and ICML conferences. She has served on the board of directors of the ACM SIGKDD and VLDB foundation. She was program chair for the ACM SIGKDD 2008 conference, research track co-chair for the VLDB 2011 conference and has served as program committee member for SIGMOD, VLDB, SIGKDD, ICDE, and ICML conferences, and on the editorial boards of the ACM TODS and ACM TKDD journals.

Prof. Guoliang Li

Tsinghua University, Beijing, China.

Title:

openGauss: An Autonomous Database System.
April 14, 2022 (Thursday), 13.30-14.30, Hall 1

Abstract:

In this talk, I will present how to build an autonomous database system. I discuss how to integrate effective learning-based models into database systems to build learned optimizers (including learned query rewrite, learned cost/cardinality estimation, learned join order selection and physical operator selection) and learned database advisors (including self-monitoring, self-diagnosis, self-configuration, and self- optimization). I also propose an effective validation model to validate the effectiveness of learned models. I discuss effective training data management and model management platforms to easily deploy learned models. Finally I will introduce our autonomous database system openGauss.

Bio:

Guoliang Li is a full professor and the deputy head of Department of Computer Science, Tsinghua University, Beijing, China. His research interests include large-scale data integration and cleaning, human-in-the-loop data management, machine learning for database, and database for machine learning. He is a general co-chair of SIGMOD 2021, demo co-chair of VLDB 2021, industry co-chair of ICDE 2022, and PC co-chair of DASFAA 2019. He is also an associate editor of VLDB journal and IEEE TKDE. He is a steering committee member of IEEE TCDE and DASFAA. He received best paper awards (candidates) of VLDB 2020, ICDE 2018, KDD 2018, CIKM 2017 and DASFAA 2014. He received Early Research Contribution Award of VLDB and Early Career Award of IEEE TCDE.

2. Panel Session

Title: Futuristic Data Interfaces
Hall 1, April 12, 2022 (Tuesday), 17:00-18:30 India Time

Overview:

Database systems have been around for more than four decades and have been widely used in academia and industry across the globe. In this panel we discuss the following two questions from various perspectives with four different internationally renowned database experts.

What are the challenges of current database interfaces?
What solutions do you propose to solve these challenges?

Moderator:

Kurt Stockinger

Zurich University of Applied Sciences,
Switzerland.

Panelists:

Georgia Koutrika

Athena Research,
Greece.

Jaydeep Sen

IBM Research,
India.

Immanuel Trummer

Cornell University,
USA.

Lei Zou

Peking University,
China.

Moderator: Kurt Stockinger (Zurich University of Applied Sciences, Switzerland)

Bio: Prof. Dr. Kurt Stockinger is Professor of Computer Science, Director of Studies in Data Science at Zurich University of Applied Sciences (ZHAW) and Co-Head of the ZHAW Datalab. His research focuses on Data Science with emphasis on Big Data, Natural Language Query Processing, Query Optimization and Quantum Computing. Essentially, his research interests are at the intersection of databases, natural language processing and machine learning. He is also on the Advisory Board of Callista Group AG and the International AIQT Foundation. Previously Kurt Stockinger worked at Credit Suisse in Zurich, Switzerland, at Lawrence Berkeley National Laboratory in Berkeley, California, at California Institute of Technology, California as well as at CERN in Geneva, Switzerland. He holds a Ph.D. in computer science from CERN / University of Vienna.

(i) Georgia Koutrika (Athena Research, Greece)

Title: Intelligent Data Assistants

Abstract: Data is considered the 21st century’s most valuable commodity. Analysts exploring data sets for insight, scientists looking for patterns, and consumers looking for information are just a few examples of user groups that need to access and dig into data. Despite technological advances in the data exploration and data management domains, existing systems are falling behind in bridging the chasm between data and users, making data accessible and useful only to the few. A futuristic data interface would enable interaction with data using natural language, would understand the data as well as the user intent, would guide the user, and make suggestions, and altogether help the user leverage data for all sorts of purposes (from finding answers to questions to revealing patterns and finding solutions to problems) in a more natural way. These systems, which we call intelligent data assistants, require the synergy of several technologies and innovation in all these fronts, including natural language interfaces, data exploration, conversational AI, and data management.

Bio: Georgia Koutrika is a Research Director at Athena Research Center in Greece. She has more than 15 years of experience in multiple roles at HP Labs, IBM Almaden, and Stanford. Her work emerges at the intersection of data management, natural language processing and deep learning and focuses on intelligent and interactive data exploration, conversational data systems, and user-driven data management. Her work has been incorporated in commercial products, described in 14 granted patents and 26 patent applications in the US and worldwide, and published in more than 100 papers in top-tier conferences and journals. Georgia is an ACM Senior Member and IEEE Senior Member. She is a member of the VLDB Endowment Board of Trustees, member of the PVLDB Advisory Board, member of the ACM-RAISE Working Group, co-Editor-in-chief for VLDB Journal, PC co-chair for VLDB 2023, co-EiC of Proceedings of VLDB (PVLDB). She has been associate editor in top-tier conferences (such as ACM SIGMOD, VLDB) and journals (VLDB Journal, IEEE TKDE), and she has been in the organizing committee of several conferences including SIGMOD, ICDE, EDBT, among others. She has received a PhD and a diploma in Computer Science from the Department of Informatics and Telecommunications, University of Athens, Greece.

(ii) Jaydeep Sen (IBM Research, India)

Title: Evolution of NLIDB systems and their application in Industry setups

Abstract:: In this modern era of technology, multitude of business applications are rapidly moving towards data driven insights for intelligent decision making, analytics and more. As we continue to see heaps of digital exhaust being generated, access to data is still limited to technical users who can query the datastores with specific query languages. Natural Language Interface to Databases (NLIDB) systems have gained a lot of focus recently owing to its fascinating aim of democratizing data access to non-technical business users. The research space for NL interfaces for data has evolved a lot since its inception, starting from simple keyword based queries, all the way to machine learning based systems, also dubbed as text-to-sql challenge. While the appeal of NLIDB system is common across different persona and use-cases, deploying a NLIDB system for an industry application has its own set of challenges which are often closely coupled with the exact domain and use-case. With no "one size fits all" solution in place, it is the right time for the community to review how the different methodologies adopted for NLIDB systems correlate with their applicability across different use-cases seen in academia and industry.

Bio: Jaydeep Sen is a Research Staff Member in IBM Research AI, India Lab . His research interests include applications for natural language understanding, semantic reasoning, designing intelligent algorithms for “learning from small data” applications. His work at IBM has powered some of IBM's most prominent QA and NL application portfolio. He has publications at conferences like VLDB, SIGMOD, IJCAI, IEEE SCC, EMNLP, COLING etc. and has served as Program Committee members for AAAI, SIGMOD, ICDE etc. He has more than 20 patents (granted/filed) in USPTO as of Dec-2021.

(iii) Immanuel Trummer (Cornell University, USA)

Title: Voice Interfaces for Data Access

Abstract:The communication between user and user is shifting more and more towards voice interfaces. This trend is evidenced by devices and services such as Google Home, Amazon Alexa, or Apple’s Siri. For many users, speech is the most natural form of interaction. It enables computer use from a distance, even in scenarios where hands or the visual attention are bound (e.g., while driving). All those advantages motivate the question of how to leverage voice interfaces for convenient data access. Accessing data via voice query interfaces is challenging. First, noisy speech recognition adds uncertainty on top of the inherent difficulties of natural language understanding. Second, transferring query results to users via voice output is difficult. Verbose speech output risks overwhelming the listener. Hence, output needs to summarize and to focus on the most important trends in the data. Recent research tackles some of those challenges. Still, many research questions remain open and must be answered to make the vision of natural data access via voice query interfaces a reality.

Bio: Immanuel Trummer is assistant professor at Cornell University, working towards making data analysis more efficient and more user-friendly. His papers were selected for “Best of VLDB”, “Best of SIGMOD”, for the ACM SIGMOD Research Highlight Award, and for publication in CACM as CACM Research Highlight. His current research is funded by the NSF and by multiple Google Faculty Research Awards.

(iv) Lei Zou (Peking University, China)

Title: Natural Language Question Answering over Knowledge Graph

Abstract: AS more and more structured data become available on the web, the question of how end users can access this body of knowledge becomes of crucial importance. As a de facto standard of a knowledge base, RDF (Resource Description Framework) repository is a collection of triples, denoted as . Although SPARQL is a standard way to access RDF data, it remains tedious and difficult for end users because of the complexity of the SPARQL syntax and the RDF schema. An ideal system should allow end users to profit from the expressive power of Semantic Web standards (such as RDF and SPARQLs) while at the same time hiding their complexity behind an intuitive and easy-to-use interface.
Generally, there are two categories of existing methods on natural language question answering (Q/A) over RDF database---one is IR (Information Retrieval)-based and the other one is called semantic parsing method. In this panel, I will talk about our solution gAnswer, which is based on graph matching-based technique, to design an effective natural language interface to access KG database.

Bio: Lei Zou is a professor at Peking University, China, and his recent research interests include graph databases, knowledge graph, particularly in graph-based RDF data management, natural language question answering over knowledge graph, hardware assisted graph database systems. Lei Zou’s research is supported by multiple NSFC projects. Prof. Zou also obtained Newton Advanced Fellowships of UK Royal Society. Lei Zou has publications at conferences like VLDB, SIGMOD, ICDE etc, and has served as Program Committee members for SIGMOD, VLDB and ICDE. He served PC Area Chair of ICDE 2021 and PC Chair of WISE 2022. Now, he is an Associate Editor of IEEE Transactions on Knowledge and Data Engineering (TKDE).

3. Tutorial talks

Date and Time (Hall)	Tutorial talks
Tutorial 1 (April 12 (Tuesday), 2022, 14:45 to 16:30, Hall 5)	Title: Make Wise Decisions for Your DBMSs: Workload Forecasting and Performance Prediction Before Execution Speakers: Zhengtong Yan (University of Helsinki) Jiaheng Lu (University of Helsinki) Qingsong Guo (University of Helsinki) Gongsheng Yuan (University of Helsinki) Calvin Sun (Huawei Toronto) Steven Yang (Huawei Toronto) Brief outline of the tutorial: In this tutorial, we will focus on 1) how to forecast the future workloads (e.g., workload shift detection, arrival rate prediction, and next query prediction), and 2) how to analyze the behaviors of the workloads (e.g., execution time prediction and resource usage estimation). We will provide a comprehensive overview and detailed introduction of the two topics, from state-of-the-art methods, real-world applications, to open problems and future directions. Specifically, we will not only discuss traditional methods, such as time-series analysis, Markov modeling, analytical modeling, and experiment-driven methods, but also cover the state-of-the-art AI techniques, including machine learning, deep learning, reinforcement learning, and graph embedding. Speakers Bio Zhengtong Yan is a doctoral student at the University of Helsinki. His research topics include autonomous multi-model databases and cross-model query optimization. Jiaheng Lu is a professor at the University of Helsinki. His main research interests lie in database systems specifically in the challenge of efficient data processing from real-life, massive data repositories and the Web. He has written four books on Hadoop and NoSQL databases, and more than 100 papers published in SIGMOD, VLDB, TODS, and TKDE, etc. He has given several tutorials on multi-model data management and autonomous databases in VLDB, CIKM, and EDBT conferences. He frequently serves as a PC member for conferences including SIGMOD, VLDB, ICDE, EDBT, CIKM, etc. Qingsong Guo is a postdoctoral researcher at the University of Helsinki His research interests include multi-model databases and automatic management of big data with deep learning. Gongsheng Yuan is a doctoral student at the University of Helsinki. His research topics lie in databases with quantum theory or reinforcement learning. Calvin Sun is the Chief Database Architect at Huawei Cloud. He has 20+ years of experience in developing several database systems, ranging from embedded databases, large-scale distributed databases, to cloud-native databases. Steven Yuan is the Director of Huawei Toronto Distributed Scheduling and Data Engine Lab. He leads a research team in the big data and cloud domain, focusing on distributed scheduling and distributed database, from IaaS to PaaS.
Tutorial 2 (April 12 (Tuesday), 2022, 10:15 to 12.30, Hall 5)	Title: Reachability on Large-scale Graphs: Models, Techniques, and Trends Speakers: Xiaoshuang Chen (Guangzhou University) Long Yuan (Nanjing University of Science and Technology) Wenjie Zhang (University of New South Wales ) Ying Zhang (University of Technology Sydney) Brief outline of the tutorial: In this tutorial, we will first show the importance and challenges of studying reachability queries. Then, we will introduce the existing reachability models defined over different graphs. The computing algorithms regarding different models and environmental settings will also be presented. Finally, we will discuss future research directions in this important research area. Speakers Bio Xiaoshuang Chen is an Associate Professor in the Cyberspace Institute of Advanced Technology, Guangzhou University. Before that, she was a Postdoctoral Fellow in the School of Computer Science and Engineering, University of New South Wales. Her research interest lies in large-scale graph data analysis. She has published several papers in ICDE, VLDB and VLDBJ since 2017. Long Yuan is a Professor in the School of Computer Science and Engineering, Nanjing University of Science and Technology, China. His research focuses on graph data management and analysis. He has published papers in top venues such as VLDB, WWW, ICDE, VLDBJ, and TKDE. Wenjie Zhang is an Australian ARC Future Fellow (2021-2025) and Professor in the School of Computer Science and Engineering, University of New South Wales. Her research interests include spatial-temporal data analysis and graph data processing. She has published more than 100 papers in top venues such as TKDE, TODS, VLDBJ, SIGMOD, VLDB, and ICDE. She received the Discovery Early Career Researcher Award in 2011 and the prestigious Chris Wallace Award in 2019. Ying Zhang is an Australian ARC Future Fellow and Professor at the University of Technology, Sydney (UTS). He has been the head of the database group at the Centre for Artificial Intelligence (CAI) since 2014. His research focuses on efficient query processing and analytics on big data. He has published more than 80 papers in top venues. He had received seven ARC grants which are under the National Competitive Grants Programme (NCGP) including one ARC ADP fellowship, one ARC DECRA fellowship, one ARC future fellowship and four ARC discovery projects.
Tutorial 3 (April 13 (Wednesday), 2022, 10:15 to 12:30, Hall 5)	Title: A tutorial on biomedical image segmentation using deep learning Speakers: Sonali Agarwal (IIIT Allahabad) Krishna Pratap Singh (IIIT Allahabad) Sanjay Kumar Sonbhadra (ITER Bhubaneswar) Narinder Singh Punn (IIIT Allahabad) Brief outline of the tutorial: Deep learning algorithms, in particular convolutional networks, have rapidly become a methodology of choice for analyzing medical images. Most medical applications require identifying and localizing the objects or regions (damaged tissues, cells or nuclei) found in medical imaging such as CAT scans, X-Rays, Ultrasound, etc. for diagnosis, monitoring and treatment. This delineation is generally performed by expert clinicians or radiologists which is a complex and time-consuming task. In recent studies, the implication of transfer learning and U-Net based approaches have illustrated state-of-the-art performance in different applications for the development of computer-aided diagnosis systems to localize the infected or damaged tissues or cells in the body using various modalities for early diagnosis and treatment of diseases such as brain tumor, lung cancer, Alzheimer, breast cancer, etc. With this motivation, this tutorial focuses on state-of-the-art deep learning approaches, a critical discussion of open challenges and directions for future research in the area of biomedical image segmentation. Speakers Bio Sonali Agarwal is working as an Associate Professor in the Information Technology Department of Indian Institute of Information Technology (IIIT), Allahabad, India. She received her Ph. D. Degree at IIIT Allahabad and joined as faculty at IIIT Allahabad, where she has been teaching since October 2009. She holds Bachelor of Engineering (B.E.) degree in Electrical Engineering from Bhilai Institute of Technology, Bhilai, (C.G.) India and Masters of Engineering (M.E.) degree in Computer Science from Motilal Nehru National Institute of Technology (MNNIT), Allahabad, India Her main research interests are in the areas of Artificial Intelligence and Big Data. She is the head of Big Data Analytics Lab at IIIT Allahabad, India. Krishna Pratap Singh is working as an Associate professor in the Information Technology Department at IIIT Allahabad, India. He received his PhD and Master from IIT Roorkee. He has been working at IIIT Allahabad since 2009. His main research areas are Machine Learning, Transfer Learning and Optimization. Sanjay Kumar Sonbhadra is presently working as Assistant Professor in the Computer Science and Engineering Department of ITER, Shiksha ‘O’ Anusandhan, Bhubaneswar, Odisha, India. He is mainly working on One-class classification, Anomaly detection, Target class guided dimensionality reduction and training sample selection techniques and Big data analytics. During 2017-2021, he worked as a senior member of “Big Data Analytics Lab” at IIIT Allahabad, India. He has published many articles in the area of machine learning applications to address recent challenges of COVID-19. He has working experience of machine learning algorithms to address the challenging problem of target specific learning with limited target samples. Narinder Singh Punn is working as a Teaching Research Assistant (TRA) in the Information Technology Department of Indian Institute of Information Technology (IIIT), Allahabad, India. Narinder’s main research includes Medical Imaging segmentation, Deep learning and Artificial Intelligence techniques in healthcare. He is a senior member of “Big Data Analytics Lab” at IIIT Allahabad, India. His recent publications cover applications of deep learning in the detection and prevention of COVID-19, while also exploiting the potential of self-supervised learning in biomedical image segmentation.
Tutorial 4 (April 13 (Wednesday), 2022, 14:30 to 16:15, Hall 5)	Title: AI Meets NoSQL Database: Methods, Opportunities and Challenges Speakers: Hongzhi Wang (Harbin Institute of Technology) Zhixin Qi (Harbin Institute of Technology) Yu Yan (Harbin Institute of Technology) Brief outline of the tutorial: This tutorial is planned for 1.5 hours and consists of the following parts. (1) Background and Motivation (10’): We introduce the background of AI for database and motivate the need for applying AI techniques on NoSQL database with several scenarios. (2) Cost Estimation for Graph and Document Databases (20’): We discuss how AI techniques estimate query costs for graph and document databases. (3) Physical Design for Key-Value and Graph Databases (20’): We introduce physical design methods for key-value and graph databases based on AI techniques. (4) Index Recommendation for Key-Value and Document Databases (15’): We discuss existing AI-based index recommendation approaches for key-value and document databases. (5) Opportunities and Challenges (20’): We present the research opportunities and challenges for NoSQL database management based on AI techniques. (6) Summary (5’): We summarize this tutorial and give our critical thoughts to AI for NoSQL database. Speakers Bio Hongzhi Wang, Professor, PHD supervisor, the head of massive data computing center and the vice dean of the honors school of Harbin Institute of Technology, the secretary general of ACM SIGMOD China, outstanding CCF member, a standing committee member CCF databases and a member of CCF big data committee. Research Fields include big data management and analysis, database systems, knowledge engineering and data quality. He was “starring track” visiting professor at MSRA and postdoctoral fellow at University of California, Irvine. Prof. Wang has been PI for more than 10 national or international projects including NSFC key project, NSFC projects and National Technical support project, and co-PI for more than 10 national projects include 973 project, 863 project and NSFC key projects. He also serves as a member of ACM Data Science Task Force. He has won First natural science prize of Heilongjiang Province, MOE technological First award, Microsoft Fellowship, IBM PHD Fellowship and Chinese excellent database engineer. His publications include over 300 papers in the journals and conferences such as VLDB Journal, IEEE TKDE, VLDB, SIGMOD, ICDE and SIGIR, 6 books and 6 book chapters. His PHD thesis was elected to be outstanding PHD dissertation of CCF and Harbin Institute of Technology. He severs as the reviewer of more than 20 international journal including VLDB Journal, IEEE TKDE, and PC members of over 50 international conferences including SIGMOD, VLDB, KDD, ICML, NeurpIS, ICDE, etc. His papers were cited more than 3000 times. His personal website is http://homepage.hit.edu.cn/wang. Zhixin Qi is currently a PhD student in School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China. She received her bachelor degree from Harbin Engineering University in 2016, and received her master degree from Harbin Institute of Technology in 2018. Her research interests include AI4DB, knowledge graph, and graph data management. She was awarded National Scholarship for PhD students in 2021, National Scholarship for master students in 2017, and National Scholarship for undergraduates in 2014. She has published more than 10 papers in international journals and conferences, including TKDE, KAIS, KBS, Neurocomputing, JCST, CIKM, and DASFAA. Yu Yan is currently a PhD student in School of Computer Science and Technology, Harbin Institute of Technology, Harbin, China. She received her master degree from Harbin Institute of Technology in 2021. She committed to the research of database tuning, multi-model database, and database auto-management. She got National Scholarship for master students in 2020. She has published many papers in international conferences and journals, such as Information Sciences, ApWeb and etc.
Tutorial 5 (April 14 (Thursday), 2022, 10:15 to 12:30, Hall 5)	Title: Time Series Anomaly Detection Toolkit for AI Applications [Slides] Speakers: Dhaval Patel (IBM Research) Dzung Phan (IBM Research) Markus Mueller (IBM Germany) Brief outline of the tutorial: The tutorial is organized in a sequence of three sections: Introduction, Theory and Hands-on-demo. In part one, we will briefly discuss foundations of time series dataset with the help of real-world examples. We will also present a broad taxonomy of time series dataset. We will also present general definition of anomalies in time series data and discuss three common variants of Anomaly/Outlier Detection problems. Next, we discuss basic machine learning primitives such as Estimator, Transformer, Data Stationarizer, etc that are useful for building anomaly pipeline. In machine learning field, these components become a backbone for building a complex model learning pipelines. We will formally introduce the key API such as ``fit'', ``predict'', ``decision\_function'', to the participant with the help of 30+ different anomaly detection algorithms. Apart from provide the categorization of these algorithms, we will also discuss one algorithm namely Gaussian Graphical Model for interpretable anomaly detection. The access to the toolkit is made available via IBM API Hub Platform (https://developer.ibm.com/apis/catalog/ai4industry--anomaly-detection-product/Introduction). The example notebooks are accessible at IBM's public github (https://github.com/IBM/anomaly-detection-code-pattern/). The tutorial finally analyzes open issues and future directions in this vibrant and rapidly evolving research area. Speakers Bio Dhaval Patel work at IBM Research since 2016. Dr. Dhaval Patel hold PhD in Computer Science from National University of Singapore and master's degree in Information Technology from Indian Institute of Technology – Kharagpur. Dr. Patel is an expert in Data Mining, Machine Learning, Time Series Data Analysis, etc. The significance of his research contributions has been demonstrated in 60+ published papers (10 journal papers and 50+ conference papers) in high impact, refereed, top-notch venues in Database, Data Mining, Big Data and Machine Learning, 1 issued US patent and 18 patent applications. He is recipients of 9 outstanding technical/research accomplishments awards from IBM for advancing AI technology to solve several real-world industrial problems. He is key contributor in many Flagship IBM Product including AutoAI-TS, Maximo Application Suites for Anomaly Detection at Scale, etc. Dzung Phan is a Research Staff Member at IBM Research, New York, USA since 2010. He received a Ph.D. degree in applied mathematics from the University of Florida in 2010, a M.S. degree in computational engineering from National University of Singapore in 2004, and a B.S. degree in mathematics from Vietnam National University, Hanoi in 2001. His research interests include optimization theory and algorithms, machine learning, and operations research. In particular, he is currently working on anomaly and change detection, sparse learning, and data-driven decision making. He has published more than 40 technical papers in refereed conferences and journals including top machine learning/data mining conferences such as ICML, NeurIPS, IJCAI, and ICDM. He has also filed about 40 U.S. patents. He received the 2012 Pat Goldberg Best Paper Award and a 2020 INFORMS Wagner Prize semi-finalist. Markus Muller studied Math, Operations Research and Computer Science and has over 2 decades of experience in IT in different roles, mainly as software architect. He has shifted towards Machine Learning in 2018, first as an architect for an NLU related offering, then as data scientist for an offering in the IIoT space. Markus worked at IBM Watson Center, Munich.

4. Workshops (April 11, 2022, Monday, 09.00- 18.00)

Workshop Name	Program
International Workshop on Blockchain Technologies (IWBT2022)	https://sites.google.com/ahduni.edu.in/iwbt2022/home 9:00 – 9:15 (IST) "Welcome and Introduction" by Dr. Sanjay Chaudhary, Workshop co-chair, Dean and Professor, Ahmedabad University 9:15 - 10:15 (IST) Keynote Talk: "From Impossible Triangles to innovations through Blockchains" by Professor C. Pandu Rangan, Sathish Dhawan Chair professor, Computer science and Automation, IISc, Bangalore, India 10:15 to 11:00 (IST) Invited Talk: "SnarkWars: The promise of Zero Knowledge Proofs for Scaling Blockchains" by Dr. Nitin Singh, Senior Research Scientist, IBM Research India 11:00 - 11:15 (IST) Tea / Coffee Break 11:15 - 12:00 (IST) Invited Talk: "Proofs of Storage and Applications" by Dr. Sushmita Ruj, Senior Lecturer, School of Computer Science and Engineering, UNSW, Sydney 12:00 - 12:30 (IST) Research Paper Presentation "Securing Cookies/Sessions through Non-Fungible Tokens" by Dr. Kaushal Shah, Uday Khokhariya, Nidhay Pancholi, Shambhavi Kumar and Keyur Parmar, Pandit Deendayal Energy University (PDEU), Gandhinagar, India 12:30 - 14:00 (IST) Lunch Break 14:00 - 14:45 (IST) Invited Talk “Next-Generation Blockchains -- Potential and Challenges” by Dr. Sujit Prakash Gujar, Assistant Professor, International Institute of Information Technology, Hyderabad, India 14:45 - 16:15(IST) Research Paper Presentations "Towards a Blockchain Solution for Customs Duty-Related Fraud" by Dr. Christopher G. Harris, Associate Professor, Computer Science, School of Mathematical Sciences, University of Northern Colorado, USA (14:45 - 15:15) "Model-Driven Development of Distributed Ledger Applications" by Piero Fraternali, Dr. Sergio Luis Herrera Gonzalez, Matteo Frigerio and Mattia Righetti, Dipartimento di Elettronica, Informazione e Bioingegneria, Politecnico di Milano, Piazza Leonardo da Vinci 32, Milan, Italy (15:15 - 15:45) "Collaborative Blockchain based Distributed Denial of Service Attack Mitigation approach with IP Reputation System" by Darshi Patel and Dr. Dhiren Patel, Gujarat Vidyapith, India (15:45 - 16:15) 16:15 - 16:30(IST) Tea / Coffee Break 16:30 - 17:30(IST) Panel Discussion on "Trends and Challenges in the adoption of blockchain for real-world problems" by Dr. Raghavendra Deshmukh, Engineering Leader, Google Dr. Sandip Chakraborty, Associate Professor, Indian Institute of Technology, Kharagpur, West Bengal, India Dr. Anu Singh, Blockchain Supply Chain Transformation Leader, IBM 17:30 - 17:45(IST) Concluding Remarks by Dr. Krishnasuri Narayanam, Workshop co-chair, Senior Research Software Engineer, IBM Research, India
Managing Air Quality Through Data Science	https://cs.ashoka.edu.in/maqtds/ 9:30 – 9:45 AM(IST) Welcome and Introduction to Workshop 9:45 – 10:00 AM(IST) Opening address by "Prof. Gautam Menon, Director, Centre for Climate Change and Sustainability (3CS) at Ashoka University" 10:00 – 11:00 AM(IST) Keynote 1: Big Data Resources to Support Research Opportunities on Air Pollution Analysis in India by "Dr. Sarath Guttikunda, Adjunct Faculty, TRIP-C, IIT Delhi, and Founder/ Director, Urban Emissions" 11:00 – 11:15 AM(IST) Break 11:15 – 11:45 AM(IST) Presentation 1: Leveraging geospatial technology for air quality monitoring by "Tejasvini Puri, BlueSky Analytics" 11:45 AM – 12:15 PM(IST) Presentation 2: Visualizing Spatio-Temporal Variation of Ambient air pollution in selected small cities in India by "Prof. Girish Agrawal & Hifzur Rahman, OP Jindal Global University" 12:15 – 1:15 PM(IST) Lunch Break 1:15 – 1:45 PM(IST) Presentation 3: Air Quality Data Collection in Hyderabad using Low Cost Sensors: Initial Experiences by "Srinivas Annappalli & Chandra Shekar N, IIIT Hyderabad" 1:45 – 2:45 PM(IST) Keynote 2: Making air pollution legible! Data and governance in Indian cities by "Dr. Anant Maringanti, Director, Hyderabad Urban Lab" 2:45 – 3:15 PM(IST) Presentation 4: AIR-Tree: Air-quality Indices based R-Tree for Air-Quality Data Management by "Prof. Anirban Mondal & Mr. Raghav Mittal, Ashoka University" 3:15 – 4:00 PM(IST) Panel Discussion moderated by Prof. Anirban Mondal: Collaboration across a range of STEM and social sciences disciplines towards air quality improvement Dr. Jai Ganesh (Mphasis) Mr. Vikas Sahni (Ashoka University) Dr. Anant Maringanti (Hyderabad Urban Lab) Prof. Girish Agrawal (OP Jindal Global University) Prof. P. Krishna Reddy (IIIT Hyderabad) 4:00 – 4:15 PM(IST) Closing Remarks
1^st Workshop on Pattern mining and Machine learning in Big complex Databases (PMBD 2022)	https://philippe-fournier-viger.com/PMDB_2021/program.html 8:30 AM(IST) Opening of the workshop 8:45 AM(IST) Keynote Talk: Approximate Computing for Big Data Analysis by Distinguished Prof. Joshua Zhexue Huang 9:45 AM(IST) Paper #2 An Algorithm for Mining Fixed-Length High Utility Itemsets by "Le Wang" 10:05 AM(IST) Paper #3 A Novel Method to Create Synthetic Samples with Autoencoder Multi-layer Extreme Learning Machine by "Qihang Huang, Yulin He, Shengsheng Xu and Joshua Zhexue Huang" 10:25 AM(IST) Paper #4 Pattern Mining: Current Challenges and Opportunities by "Philippe Fournier-Viger, Wensheng Gan, Youxi Wu, Mourad Nouioua, Wei Song, Tin Truong and Hai Duong Van" 10:45 AM(IST) Paper #7 Why not to Trust Big Data: Identifying Existence of Simpson’s Paradox by "Rahul Sharma, Minakshi Kaushik, Sijo Arakkal Peious, Mahtab Shahin, Ankit Vidhyarthi and Dirk Draheim" 11:05 AM(IST) Paper #8 Localized Metric Learning for Large Multi-Class Extremely Imbalanced Face Database by "Seba Susan and Ashu Kaushik" 11:25 AM(IST) Paper #9 Top-k dominating queries on incremental datasets by "Jimmy Ming-Tai Wu, Ke Wang and Jerry Chun-Wei Lin" 11:45 AM(IST) Closing of the workshop
The 8^th International Workshop on Big Data Management and Service (BDMS 2022)	https://zheng-kai.com/cfp 9.00-12.30 AM(IST) Morning session Session Chair: Kevin Zheng, An Liu, Xiaoling Wang BDMS001: H-V：An Improved Coding Layout based on Erasure Coded Storage System (9:00-9:20) BDMS002: An Autoencoder-based Model for Pedestrian Trajectory Prediction of Variable-Length (9:20-9:40) BDMS003: A Survey on Spatiotemporal Data Processing Techniques in Smart Urban Rail (9:40-10:00) BDMS004: Fast Vehicle Track Counting in Traffic Video (10:00-10:20) BDMS005: TSummary A Traffic Summarization System using Semantic Words (10:20-10:40) BDMS006: Attention Cooperated Reinforcement Learning for Multi agent Path Planning (10:40-11:00) BDMS007: Big Data-driven Stable Task Allocation in Ride-hailing Services (11:00-11:20) BDMS008: Weighted Mean Field Multi Agent Reinforcement Learning via Reward Attribution Decomposition (11:20- 11:40) Keynotes： Zero-Shot Event Extraction Based on Contrastive Learning (11:40-12:30) by Wendi Ji
The 7^th International Workshop on Big Data Quality Management (BDQM 2022)	https://bdqm2022.github.io/ 9:00-10:10 AM(IST) Invited Talk: Title: Big Data Management and Analytics via Direct Computing on Compression by "Feng Zhang, Renmin University of China". 10:10-11:20 AM(IST) Invited Talk: Title: Data Quality Management: Turn Waste into Wealth by "Shaoxu Song, Tsinghua University, China". 11:20-12:50 AM(IST) Paper presentations Evaluating Presto and SparkSQL with TPC-DS by "Yinhao Hong, Sheng Du, Jianquan Leng (Renmin University of China, Beijing Kingbase Information Technology Co., Ltd, Beijing, China)" Optimizing the Age of Sensed Information in Cyber-Physical Systems by "Yinlong Li，Siyao Cheng，Feng Li, Jie Liu, Hanling Wu (Harbin Institute of Technology, Harbin Institute of Technology(shenzen), Beijing Institute of Astronautical Systems Engineering, China)" Aggregate Query Result Correctness using Pattern Tables by "Nitish Yadav, Ayushi Malhotra, Sakshee Patel, Minal Bhise (Distributed Databases Group, DAIICT, Gandhinagar, India)" Time Series Data Quality Enhancing based on Pattern Alignment by "Jianping Huang, Hao Chen, Hongkai Wang, Jun Feng, Liangying Peng， Zheng Liang, Hongzhi Wang, Tianlan Fan, Tianren Yu (State Grid Zhejiang Electric Power Co., Ltd, State Grid Zhejiang Information and Telecommunication Branch, China, Harbin Institute of Technology)" Research on feature extraction method of data quality intelligent detection by "Weiwei Liu，Shuya Lei, Xiaokun Zheng, Xiao Liang (Artificial Intelligence on Electric Power System State Grid Corporation Joint Laboratory(GEIRI) China)"
The 6^th International Workshop on Graph Data Management and Analysis (GDMA 2022)	https://gdma2022.github.io/ 14:00-14:35 (IST) Keynote by Tieyun Qian (Wuhan University) Title: Diversity and Novelty Oriented Recommendation for Long-tailed Items 14:35-15:10 (IST) Keynote by Peng Peng (Hunan University) Title: Distributed RDF Graph Management 15:20-16:45 (IST) (Oral Paper) Chinese Spelling Error Detection and Correction Based on Knowledge Graph 16:45-17:10 (IST) (Oral Paper) Construction and Application of Event Logic Graph: A Survey 17:10-17:35 (IST) (Oral Paper) Enhancing Low-resource Languages Question Answering with Syntactic Graph 17:35-18:00 (IST) (Oral Paper) Profile Consistency Discrimination

5. Research Sessions

Research Session Id	Session Name	Paper ID	Paper Title
	(Date, Time, and Hall)
1	Research Session #1 (Queries) (April 12 (Tuesday), 2022, 10.15-12.30, Hall 1)	20	Approximate Continuous Top-K Queries Over Memory Limitation-based Streaming Data [Abstract] Abstract Continuous top-k query over sliding window is a fundamental prob- lem over data stream. It retrieves k objects with the highest scores when the window slides. Existing efforts include exact-based algorithms and approximate- based algorithms. Their common idea is maintaining a small subset of objects in the window. When the window slides, query results could be found from this set as much as possible. However, the space cost of all existing efforts is high ,i.e., linear to the scale of objects in the window, cannot work under memory limitation-based streaming data, i.e., a general environment in real applications. In this paper, we define a novel query named ρ−approximate continuous top-k query, which returns error-bounded answers to the system. Here, ρ is a thresh- old, used for bounding the score ratio between approximate and exact results. In order to support ρ−approximate continuous top-k query, we propose a novel framework named ρ−TOPK. It can self-adaptively adjust ρ based on the dis- tribution of streaming data, and achieve the goal of supporting ρ−approximate continuous top-k query over memory limitation-based streaming data. Theoret- ical analysis indicates that even in the worsst case, both running cost and space cost of ρ−TOPK are all unrelated with data scale.
			Rui Zhu (School of Computer Science, Shenyang Aerospace University, China); Liu Meng (School of Computer Science, Shenyang Aerospace University, China); Bin Wang (College of Computer Science and Engineering, Northeastern University, China); Xiaochun Yang (College of Computer Science and Engineering, Northeastern University, China); Xiufeng Xia (School of Computer Science, Shenyang Aerospace University, China);
		22	Cross-Model Conjunctive Queries over Relation and Tree-structured Data [Abstract] Abstract Conjunctive queries are the most basic and central class of database queries. With the continued growth of demands to manage and process the massive volume of different types of data, there is little research to study the conjunctive queries between relation and tree data. In this paper, we study Cross-Model Conjunctive Queries (CMCQs) over relation and tree-structured data (XML and JSON). To efficiently process CMCQs with bounded intermediate results we first encode tree nodes with position information. With tree node original label values and encoded position values, it allows our proposed algorithm CMJoin to join relations and tree data simultaneously, avoiding massive intermediate results. CMJoin achieves worst-case optimality in terms of the total result of label values and encoded position values. Experimental results demonstrate the efficiency and scalability of the proposed techniques to answer a CMCQ in terms of running time and intermediate result size.
			Yuxing Chen (University of Helsinki); Valter Uotila (University of Helsinki); Jiaheng Lu (University of Helsinki); Zhen Hua Liu (Oracle); Souripriya Das (Oracle);
		178	Leveraging Search History for Improving Person-Job Fit [Abstract] Abstract As the core technique of online recruitment platforms, personjob fit can improve hiring efficiency by accurately matching job positions with qualified candidates. However, existing studies mainly focus on the recommendation scenario, while neglecting another important channel for linking positions with job seekers, i.e., search. Intuitively, search history contains rich user behavior in job seeking, reflecting important evidence for job intention of users. In this paper, we present a novel Search History enhanced PersonJob Fit model, named as SHPJF. To utilize both text content from jobs/resumes and search histories from users, we propose two components with different purposes. For text matching component, we design a BERT-based text encoder for capturing the semantic interaction between resumes and job descriptions. For intention modeling component, we design two kinds of intention modeling approaches based on the Transformer architecture, either based on the click sequence or query text sequence. To capture underlying job intentions, we further propose an intention clustering technique to identify and summarize the major intentions from search logs. Extensive experiments on a large real-world recruitment dataset have demonstrated the effectiveness of our approach.
			Yupeng Hou (Gaoling School of Artificial Intelligence, Renmin University of China, China); Xingyu Pan (School of Information, Renmin University of China, China); Wayne Xin Zhao (Gaoling School of Artificial Intelligence, Renmin University of China, China); Shuqing Bian (School of Information, Renmin University of China, China); Yang Song (BOSS Zhipin, Beijing, China); Tao Zhang (BOSS Zhipin, Beijing, China); Ji-Rong Wen (Gaoling School of Artificial Intelligence, Renmin University of China, China);
		503	Efficient In-Memory Evaluation of Reachability Graph Pattern Queries on Data Graphs [Abstract] Abstract Graphs are a widely used data model in modern data-intensive applications. Graph pattern matching is a fundamental operation for the exploration and analysis of large data graphs. In this paper, we present a novel approach for efficiently finding homomorphic matches of graph pattern queries, where pattern edges denote reachability rela- tionships between nodes in the data graph. We first propose the concept of query reachability graph to compactly encode all the possible homo- morphisms from a query pattern to the data graph. Then, we design a graph traversal-based filtering method to prune nodes from the data graph which violate reachability conditions induced by the pattern edges. We use the pruned data graph to generate a refined query reachability graph which serves as a compact search space for the pattern query an- swer. Finally, we design a multiway join algorithm to enumerate answer tuples from the query reachability graph without generating an excessive number of redundant intermediate results (a drawback of previous ap- proaches). We experimentally verify the efficiency of our approach and demonstrate that it outperforms by far existing approaches and a recent graph DBMS on evaluating reachability graph pattern queries.
			Xiaoying Wu (Wuhan University, China); Dimitri Theodoratos (New Jersey Institute of Technology, USA); Dimitrios Skoutas (R.C. Athena, Athens, Greece); Michael Lan (New Jersey Institute of Technology, USA);
		523	Revisiting Approximate Query Processing and Bootstrap Error Estimation on GPU [Abstract] Abstract Sampling-based Approximate Query Processing (AQP) is one of the promising approaches for timely and cost-effective analytics over big data. There are mainly two methods to estimate errors of approximate query results, namely analytical method and bootstrap method. Although the bootstrap method is much more general than the first method, it is rarely used in the existing AQP system due to its high computation overhead. In this paper, we propose to use the powerful GPU and a series of advanced optimization mechanisms to accelerate bootstrap, thus make it feasible to address the essential error estimation problem for AQP by utilizing bootstrap. Besides, since modern GPUs have bigger and bigger memory capacity, we can store samples in the GPU memory and use GPU to accelerate the execution of AQP queries in addition to using GPU to accelerate the bootstrap-based error estimation. Extensive experiments on the SSB benchmark show that our GPU-accelerated method is at most about two orders of magnitude faster than the CPU method.
			Hang Zhao (Fudan University, Shanghai, China); Hanbing Zhang (Fudan University, Shanghai, China); Yinan Jing (Fudan University, Shanghai, China); Kai Zhang (Fudan University, Shanghai, China); Zhenying He (Fudan University, Shanghai, China); X. Sean Wang (Fudan University, Shanghai, China);
		257	μ-join: Efficient Join with Versioned Dimension Tables [Abstract] Abstract The star schema is composed of two types of tables; fact tables record business process, whereas dimension tables store descrip- tion of business resources and contexts that are referenced from the fact tables. Analytical business queries join both tables to provide and ver- ify business findings. Business resources and contexts are not necessarily constant; the dimension table may be updated at times. Versioning pre- serves every version of a dimension record and allows a fact record to reference its associated version of the dimension record correctly. How- ever, major existing versioning practices (utilizing a binary join operator and a union operator) cause processing redundancy in queries joining a fact table and a dimension table. This paper proposes μ-join, an ex- tended join operator that directly accepts a fact table and an arbitrary number of dimension tables, and presents that this operator reduces the redundancy and speeds up fact-dimension joins queries. Our experiment demonstrates that μ-join offers speedup using the synthetic dataset (up to 71.7%).
			Mika Takata (The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo, Japan); Kazuo Goda (The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo, Japan); Masaru Kitsuregawa (The University of Tokyo, 4-6-1 Komaba, Meguro-ku, Tokyo, Japan);
		532	Learning-based Optimization for Online Approximate Query Processing [Abstract] Abstract Approximate query processing (AQP) technique speeds up query execution by reducing the amount of data that needs to be processed, while sacrificing the accuracy of the query result to some extent. AQP is essentially a trade-off between the accuracy of the query result and the query latency. However, the heuristic AQP optimization and error control mechanism used by the existing AQP system fails to meet the accuracy requirements of users. This paper proposes a deep learningbased error prediction model to guide AQP query optimization. By using this model, we can estimate the errors of candidate query plans and select the appropriate plans that can meet the accuracy requirement with high probability. Extensive experiments show that the AQP system proposed in this paper can outperform the state-of-the-art online sampling-based AQP approach.
			Wenyuan Bi (Fudan University, Shanghai, China); Hanbing Zhang (Fudan University, Shanghai, China); Yinan Jing (Fudan University, Shanghai, China); Zhenying He (Fudan University, Shanghai, China); Kai Zhang (Fudan University, Shanghai, China); X. Sean Wang (Fudan University, Shanghai, China);
2	Research Session #2 (Text and Image-I), (April 12 (Tuesday), 2022, 10.15-12.30, Hall 2)	65	Emotion-aware Multimodal Pre-training for Image-grounded Emotional Response Generation [Abstract] Abstract Face-to-face communication leads to better interactions between speakers than text-to-text conversations since the speakers can capture both textual and visual signals. Image-grounded emotional response generation (IgERG) tasks requires chatbots to generate a response with the understanding of both textual contexts and speakers’ emotions in visual signals. Pre-training models enhance many NLP and CV tasks and image-text pre-training also helps multimodal tasks. However, existing image-text pre-training methods typically pre-train on images by recognizing or modeling objects, but ignore the emotions expressed in the images. In this paper, we propose several pre-training tasks in a unified framework that not only captures emotions from images but also learns to incorporate the emotion into text generation. The pre-training involves single-modal learning to strengthen the ability to understand images and generate texts. It also involves cross-modal learning to enhance interactions between images and texts. The experiments verify our method in appropriateness, informativeness, and emotion consistency.
			Zhiliang Tian (The Hong Kong University of Science and Technology, Hong Kong SAR, China); Zhihua Wen (Science and Technology on Parallel and Distributed Laboratory, National University of Defense Technology, Hunan, China); Zhenghao Wu (The Hong Kong University of Science and Technology, Hong Kong SAR, China); Yiping Song (National University of Defense Technology, Hunan, China); Jintao Tang (National University of Defense Technology, Hunan, China); Dongsheng Li (Science and Technology on Parallel and Distributed Laboratory, National University of Defense Technology, Hunan, China); Nevin L. Zhang (The Hong Kong University of Science and Technology, Hong Kong SAR, China);
		416	KdTNet: Medical Image Report Generation via Knowledge-driven Transformer [Abstract] Abstract Writing medical image reports is an inefficient and timeconsuming task for doctors. Automatically generating medical reports is an essential task of medical data mining, which can alleviate the workload of doctors and improve the standardization of reports. However, the existing methods mainly adopt the CNN-RNN structure to align image features with text features. This structure has difficulty dealing with the dependencies between distant text locations, leading to inconsistent context and semantics in the generated report. In this paper, we propose a knowledge-driven transformer (KdTNet) model for generating coherent medical reports. First, the visual grid and graph convolutional modules are devised to extract fine-grained visual features. Second, we adopt the transformer decoder to generate the hidden semantic states. Subsequently, a BERT-based auxiliary language module is employed to obtain the context language features of reports from the pre-defined medical term knowledge. We design a multimodal information fusion module to adaptively calculate the contribution of visual and linguistic features to the report for report generation. Extensive experiments on two real datasets explicate that our KdTNet model has achieved superior performance in captioning metrics and human evaluation compared with the state-of-the-art methods.
			Yiming Cao (School of Software, Shandong University, China); Lizhen Cui (Joint SDU-NTU Centre for Artificial Intelligence Research (C-FAIR), Shandong University, China); Fuqiang Yu (School of Software, Shandong University, China); Lei Zhang (School of Software, Shandong University, China); Zhen Li (Department of Gastroenterology, Qilu Hospital of Shandong University, China); Ning Liu (School of Software, Shandong University, China); Yonghui Xu (Joint SDU-NTU Centre for Artificial Intelligence Research (C-FAIR), Shandong University, China);
		464	AdCSE: An Adversarial Method for Contrastive Learning of Sentence Embeddings [Abstract] Abstract Due to the impressive results on semantic textual similarity (STS) tasks, unsupervised sentence embedding methods based on contrastive learning have attracted much attention from researchers. Most of these approaches focus on constructing high-quality positives, while only using other in-batch sentences for negatives which are insuﬃcient for training accurate discriminative boundaries. In this paper, we demonstrate that high-quality negative representations introduced by adversarial training help to learn powerful sentence embeddings. We design a novel method named AdCSE for unsupervised sentence embedding. It consists of an untied dual-encoder backbone network for embedding positive sentence pairs and a group of negative adversaries for training hard negatives. These two parts of AdCSE compete against each other mutually in an adversarial way for contrastive learning, obtaining the most expressive sentence representations while achieving an equilibrium. Experiments on 7 STS tasks show the eﬀectiveness of AdCSE. The superiority of AdCSE in constructing high-quality sentence embeddings is also validated by ablation studies and quality analysis of representations.
			Renhao Li (School of Computer Science, Sichuan University, Chengdu, China); Lei Duan (School of Computer Science, Sichuan University, Chengdu, China); Guicai Xie (School of Computer Science, Sichuan University, Chengdu, China); Shan Xiao (School of Computer Science, Sichuan University, Chengdu, China); Weipeng Jiang (School of Computer Science, Sichuan University, Chengdu, China);
		502	HRG: A Hybrid Retrieval and Generation model in Multi-turn Dialogue Generation [Abstract] Abstract In multi-turn dialogue generation, the generated response should consider the content before the current turn of dialogue. Due to multiple turns, it is difficult to maintain the context consistency by using only a few previous turns of the dialogue indiscriminately. Except for the context information, we can retrieve additional candidates from histor- ical contexts, according to semantic similarity. Therefore, in this paper, we integrate the historical information into the generative model called HRG. The HRG model can generate a response by using both context information and retrieved historical candidates, which contain richer in- formation such as theme and latent information. We encode contexts, current turn and historical information separately to find the most im- portant turns and give the current turn a higher level of attention. Then we propose a hierarchical fusion encoder to integrate the retrieval infor- mation through a KL divergence gate dynamically. Finally, we conduct experiments on the Ubuntu large-scale English multi-turn dialogue com- munity dataset and Daily dialogue dataset. The results show that our hybrid model performs well on both automatic evaluation and human evaluation compared with the existing baseline models.
			Deji Zhao (School of Information Science and Technology, Dalian Maritime University, Dalian, China); Xinyi Liu (School of Information Science and Technology, Dalian Maritime University, Dalian, China); Bo Ning (School of Information Science and Technology, Dalian Maritime University, Dalian, China); Chengfei Liu (Swinburne University of Technology, Australia);
		122	SimEmotion: A Simple Knowledgeable Prompt Tuning Method for Image Emotion Classification [Abstract] Abstract Image emotion classification is an important computer vi- sion task to extract emotions from images. The state-of-the-art methods for image emotion classification are primarily based on proposing new architectures and fine-tuning them on pre-trained Convolutional Neu- ral Networks. Recently, learning transferable visual models from natural language supervision has shown great success in zero-shot settings due to the easily accessible web-scale training data, i.e., CLIP. In this paper, we present a conceptually simple while empirically powerful framework for supervised image emotion classification, SimEmotion, to effectively lever- age the rich image and text semantics entailed in CLIP. Specifically, we propose a prompt-based fine-tuning strategy to learn task-specific rep- resentations while preserving knowledge contained in CLIP. As image emotion classification tasks lack text descriptions, sentiment-level con- cept and entity-level information are introduced to enrich text semantics, forming knowledgeable prompts and avoiding considerable bias intro- duced by fixed designed prompts, further improving the model’s ability to distinguish emotion categories. Evaluations on four widely-used affec- tive datasets, namely, Flickr and Instagram (FI), EmotionROI, Twitter I, and Twitter II, demonstrate that the proposed algorithm outperforms the state-of-the-art methods to a large margin (i.e., 5.27% absolute ac- curacy gain on FI) on image emotion classification tasks.
			Sinuo Deng (Faculty of Information Technology, Beijing University of Technology, Beijing, China); Ge Shi (Faculty of Information Technology, Beijing University of Technology, Beijing, China); Lifang Wu( (Faculty of Information Technology, Beijing University of Technology, Beijing, China); Lehao Xing (Faculty of Information Technology, Beijing University of Technology, Beijing, China); Wenjin Hu (Faculty of Information Technology, Beijing University of Technology, Beijing, China); Heng Zhang (Faculty of Information Technology, Beijing University of Technology, Beijing, China); Ye Xiang (Faculty of Information Technology, Beijing University of Technology, Beijing, China);
		196	E-commerce Knowledge Extraction via Multi-modal Machine Reading Comprehension [Abstract] Abstract Extracting commodity attributes from unstructured data is an essential information extraction task in the e-commerce domain. It plays an important role in tasks such as commodity recommendation and commodity knowledge base extension. Traditional models for these tasks often only used text modal information, which is not enough to describe the commodity. In recent years, we increasingly see using multi-modal data to describe commodities such as text, images, and videos, which provides the possibility for better product attribute extraction. To this end, we propose a novel model named E-commerce Knowledge Extrac- tion via Multi-modal Machine Reading Comprehension(EKE-MMRC). Specifically, it finds the missing attributes from the existing knowledge base and generates problems. Then packages them with multi-modal de- scriptions and encode as fusion vector. And then, decode the correlation of description with the attribute and generate an answer from the fusion vector. Finally, use the correlation as the weight of voting for the answer. At the same time, we contributed a dataset for this task based on public e-commerce data: E-commence Multi-modal Commodity Attributes Ex- traction Dataset (E-MCAE). We also conducted experiments on a public dataset. The experimental results show that our proposed method is ef- fective, and it achieves more than 15% improvement compared with the SOTA single-modal extraction method.
			Chaoyu Bai (School of Cyber Science and Engineering, Southeast University, Nanjing, China);
		403	Concurrent Transformer for Spatial-Temporal Graph Modeling [Abstract] Abstract Previous studies have shown that concurrently extracting spatial and temporal information is a better way to model spatial-temporal data. However, in these studies, the receptive ﬁeld has been ﬁxed to construct the carrier of concurrent extraction, resulting in the lack of ﬂexibility in selecting receptive ﬁelds, and the loss of scalability for capturing long-range temporal dependencies. Furthermore, these studies will result in static weights that are insuﬃcient to describe the complex spatial and temporal dependencies. In this paper, we propose the Concurrent Spatial-Temporal Transformer (CSTT), which ensures the denseness of the carrier of concurrent extraction, so that messages under diﬀerent receptive ﬁelds can be passed more eﬃciently, thus making the selection of receptive ﬁeld more ﬂexible. Therefore, the scalability of capturing long-range temporal dependencies can also be promised. Additionally, a uniﬁed self-attention mechanism will be applied on the carrier of concurrent extraction to capture spatial and temporal information, so that the dependencies between both dimensions under diﬀerent contextual information can be preserved. On these bases, we design an iterative strategy to further tackle long sequences. Experiments on four traﬃc real-world datasets illustrate that our algorithm can achieve signiﬁcant improvement on the classical spatial-temporal modeling task.
			Yi Xie (School of Computer Science, Fudan University, Shanghai, China); Yun Xiong (School of Computer Science, Fudan University, Shanghai, China); Yangyong Zhu (School of Computer Science, Fudan University, Shanghai, China); Philip S. Yu (University of Illinois at Chicago, Chicago, U.S.A); Cheng Jin (Shanghai Meteorological Disaster Prevention Technology Center, Shanghai, China); Qiang Wang (Shanghai Meteorological Disaster Prevention Technology Center, Shanghai, China); Haihong Li (Shanghai Meteorological Disaster Prevention Technology Center, Shanghai, China);
		519	Modeling Uncertainty in Neural Relation Extraction [Abstract] Abstract Previous work on neural relation extraction mainly focus on point-based methods and ignores uncertainty within the bag, thus making poor predictions when there are insufficient instances of the bag. To solve this problem, in this paper, we propose two density-based methods. Specifically, we assume each bag is a Gaussian distribution and sentences in the bag are drawn from it. We use predicted variance, capturing bag’s uncertainty, as well as predicted mean to draw more samples to enrich one-instance bags. We also use predicted variance to vote for good representation and temper the loss. To the best of our knowledge, this is the first paper to model uncertainty in neural relation extraction. Experiment results on NYT-10 show significant improvements over baselines.
			Hang Zhao (School of Computer Science, Fudan University, Shanghai, China); Hanbing Zhang (School of Computer Science, Fudan University, Shanghai, China); Yinan Jing (School of Computer Science, Fudan University, Shanghai, China); Kai Zhang (School of Computer Science, Fudan University, Shanghai, China); Zhenying He (School of Computer Science, Fudan University, Shanghai, China); X. Sean Wang (School of Computer Science, Fudan University, Shanghai, China);
3	Research Session #3 (Applications of ML-I), (April 12 (Tuesday), 2022, 10.15-12.30, Hall 3)	174	Similarity-Aware Collaborative Learning for Patient Outcome Predcition [Abstract] Abstract The rapid growth in the use of electronic health records (EHR) offers promises for predicting patient outcomes. Previous works on this task focus on exploiting temporal patterns from sequential EHR data. Nevertheless, such approaches model patients independently, miss- ing out on the similarities between patients, which are crucial for pa- tients’ health risk assessment. Moreover, they fail to capture the fine- grained progression of patients’ status, which assist in inferring the pa- tients’ future status. In this work, we propose a similarity-aware col- laborative learning model SiaCo for patient outcome prediction. In par- ticular, we design two similarity measurers and two global knowledge matrices to separately calculate the similarity of patients with different information levels and support collaborative learning between patients. To capture the more fine-grained progression of patients’ status, we de- sign a parallelized LSTM to model the temporal-dependent patterns of patient status. Finally, SiaCo integrates the information learned from two measures and the parallelized LSTM to predict patient outcomes. Extensive experiments are conducted on two real disease datasets. The experimental results demonstrate that SiaCo outperforms the state-of- the-art models for two typical tasks of patient outcome prediction.
			Fuqiang Yu (School of Software, Shandong University, China); Lizhen Cui (Joint SDU-NTU Centre for Artificial Intelligence Research (C-FAIR), Shandong University, China); Yiming Cao (School of Software, Shandong University, China); Ning Liu (School of Software, Shandong University, China); Weiming Huang (School of Software, Shandong University, China); Yonghui Xu (Joint SDU-NTU Centre for Artificial Intelligence Research (C-FAIR), Shandong University, China);
		252	Heterogeneous Federated Learning via Grouped Sequential-to-Parallel Training [Abstract] Abstract Federated learning (FL) is a rapidly growing privacy preserving collaborative machine learning paradigm. In practical FL applications, local data from each data silo reﬂect local usage patterns. Therefore, there exists heterogeneity of data distributions among data owners (a.k.a. FL clients). If not handled properly, this can lead to model performance degradation. This challenge has inspired the research ﬁeld of heterogeneous federated learning, which currently remains open. In this paper, we propose a data heterogeneity-robust FL approach, FedGSP, to address this challenge by leveraging on a novel concept of dynamic Sequential-to-Parallel (STP) collaborative training. FedGSP assigns FL clients to homogeneous groups to minimize the overall distribution divergence among groups, and increases the degree of parallelism by reassigning more groups in each round. It is also incorporated with a novel Inter-Cluster Grouping (ICG) algorithm to assist in group assignment, which uses the centroid equivalence theorem to simplify the NP-hard grouping problem to make it solvable. Extensive experiments have been conducted on the non-i.i.d. FEMNIST dataset. The results show that FedGSP improves the accuracy by 3.7% on average compared with seven state-of-the-art approaches, and reduces the training time and communication overhead by more than 90%.
			Shenglai Zeng (Sch of Info & Comm Engin, University of Electronic Science and Technology of China, Chengdu, China); Zonghang Li (Sch of Info & Comm Engin, University of Electronic Science and Technology of China, Chengdu, China); Hongfang Yu (Sch of Info & Comm Engin, University of Electronic Science and Technology of China, Chengdu, China); Yihong He (Sch of Info & Comm Engin, University of Electronic Science and Technology of China, Chengdu, China); Zenglin Xu (Sch of Comp Sci & Tech, Harbin Institute of Technology, Shenzhen, China); Dusit Niyato (Sch of Comp Sci & Engin, Nanyang Technological University, Singapore); Han Yu (Sch of Comp Sci & Engin, Nanyang Technological University, Singapore);
		335	Peripheral Instance Augmentation for End-to-End Anomaly Detection using Weighted Adversarial Learning [Abstract] Abstract Anomaly detection has been a lasting yet active research area for decades. However, the existing methods are generally biased towards capturing the regularities of high-density normal instances with insufficient learning of peripheral instances. This may cause a failure in finding a representative description of the normal class, leading to high false positives. Thus, we introduce a novel anomaly detection model that utilizes a small number of labelled anomalies to guide the adversarial training. In particular, a weighted generative model is applied to generate peripheral normal instances as supplements to better learn the characteristics of the normal class, while reducing false positives. Additionally, with the help of generated peripheral instances and labelled anomalies, an anomaly score learner simultaneously learns (1) latent representations of instances and (2) anomaly scores, in an end-to-end manner. The experimental results show that our model outperforms the state-of-the-art anomaly detection methods on four publicly available datasets, achieving improvements of 6.15%-44.35% in AUPRC and 2.27%-22.3% in AUROC, on average. Furthermore, we applied the proposed model to a real merchant fraud detection application, which further demonstrates its effectiveness in a real-world setting. 1
			Weixian Zong (School of Data Science and Engineering, East China Normal University, Shanghai, China); Fang Zhou (School of Data Science and Engineering, East China Normal University, Shanghai, China); Martin Pavlovski (Temple University, Philadelphia, United States); Weining Qian (School of Data Science and Engineering, East China Normal University, Shanghai, China);
		342	HieNet: Bidirectional Hierarchy Framework for Automating ICD Coding [Abstract] Abstract International Classification of Diseases (ICD) is a set of classification codes for medical records. Automated ICD coding, which assigns unique Interna- tional Classification of Diseases codes with each medical record, is widely used recently for its efficiency and error-prone avoidance. However, there are challenges that remain such as heterogeneity, label unbalance, and complex relationships between ICD codes. In this work, we proposed a novel Bidirectional Hierar- chy Framework(HieNet) to address the challenges. Specifically, a personalized PageRank routine is developed to capture the co-relation of codes, a bidirectional hierarchy passage encoder to capture the codes’ hierarchical representations, and a progressive predicting method is then proposed to narrow down the semantic searching space of prediction. We validate our method on two widely used datasets. Experimental results on two authoritative public datasets demonstrate that our proposed method boosts the state-of-the-art performance by a large margin.
			Shi Wang (Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, China); Daniel Tang (University of Luxembourg, Interdisciplinary Centre for Security, Reliability and Trust (SNT), TruX, Luxembourg); Luchen Zhang (National Computer Network Emergency Response Technical Team/Coordination Center of China); Huilin Li (Department of Civil Engineering, Technical University of Denmark, 2800 Lyngby, Denmark); Ding Han (Huazhong University of Science and Technology, China);
		2	Data Source Selection in Federated Learning: A Submodular Optimization Approach [Abstract] Abstract Federated learning is a new learning paradigm that jointly trains a model from multiple data sources without sharing raw data. For the practical deployment of federated learning, data source selection is compulsory due to the limited communication cost and budget in real-world applications. The necessity of data source selection is further amplified in presence of data heterogeneity among clients. Prior solutions are either low in efficiency with exponential time cost or lack theoretical guarantees. Inspired by the diminishing marginal accuracy phenomenon in federated learning, we study the problem from the perspective of submodular optimization. In this paper, we aim at efficient data source selection with theoretical guarantees. We prove that data source selection in federated learning is a monotone submodular maximization problem and propose FDSS, an efficient algorithm with a constant approximate ratio. Furthermore, we extend FDSS to FDSS-d for dynamic data source selection. Extensive experiments on CIFAR10 and CIFAR100 validate the efficiency and effectiveness of our algorithms.
			Ruisheng Zhang (State Key Laboratory of Software Development Environment, Beihang University, China); Yansheng Wang (State Key Laboratory of Software Development Environment, Beihang University, China); Zimu Zhou (Singapore Management University, Singapore, Singapore); Ziyao Ren (State Key Laboratory of Software Development Environment, Beihang University, China); Yongxin Tong (State Key Laboratory of Software Development Environment, Beihang University, China); Ke Xu (State Key Laboratory of Software Development Environment, Beihang University, China);
		41	CLZT:A Contrastive Learning Based Framework for Zero-Shot Text Classification [Abstract] Abstract Zero-shot text classification aims to predict classes which never been seen in training stage. The lack of annotated data and huge semantic gap between seen and unseen classes make this task extremely hard. Most of existing methods employ binary classifier-based framework, and regard it as a relatedness (yes/no) prediction problem between in- stances and every candidate class. However, these methods only consider the similarities between one instance and one class at a time,and ignore semantic relations between candidate classes. To alleviate this problem, we propose a novel Contrastive Learning based Zero-shot Text classifica- tion framework (CLZT). With the contrastive optimized objects, we can capture the semantic relations between classes that need to be predicted and build more discriminative embeddings. Main experiment shows that our method achieves the best overall f1 score compared with baselines in three different datasets.
			Kun Li (School of Cyber Security, University of Chinese Academy of Sciences, China); Meng Lin (School of Cyber Security, University of Chinese Academy of Sciences, China); Songlin Hu (School of Cyber Security, University of Chinese Academy of Sciences, China); Ruixuan Li (Institute of Information Engineering, Chinese Academy of Sciences, China);
		248	An Interpretable Time Series Classification Approach Based on Feature Clustering [Abstract] Abstract The time series classification problem has been an important mining task and applied in many real-life applications. A large number of approaches have been proposed, including shape-based approaches, dictionary-based ones, ensemble-based ones and some deep-learning ap- proaches. However, these approaches either suffer from low accuracy or need massive features which hinder the interpretability. To overcome these challenges, in this paper, we propose a novel approach, FCCA, based on the feature clustering. We first present the formal definition features of various types. Then we propose the approaches of feature candidates generation, feature filtering and feature clustering. With a small number of representative features, FCCA not only achieves high accuracy, but also improves the interpretability greatly. Extensive exper- iments are conducted on UCR benchmark to verify the effectiveness and efficiency of the proposed approach.
			Fan Qiao (Fudan University, Shanghai, China); Peng Wang (Fudan University, Shanghai, China); Wei Wang (Fudan University, Shanghai, China); Binjie Wang (Beijing Jiaotong University, Beijing, China);
		332	Supervised Multi-view Latent Space Learning by Jointly Preserving Similarities across Views and Samples [Abstract] Abstract In multi-view learning, leveraging features from various views in an optimal manner to improve the performance on predictive tasks is a challenging objective. For this purpose, a broad range of approaches have been proposed. However, existing works focus either on capturing (1) the common and complementary information across views, or (2) the underlying between-view relationships by exploiting view pair similarities. Besides, for the latter, we find that the obtained similarities cannot representatively reflect the differences among views. Towards addressing these issues, we propose a novel approach called MELTS (Multi-viEw LatenT space learning with Similarity preservation) for multi-view classification. MELTS first utilizes distance correlation to explore hidden between-view relationships. Furthermore, by assuming that different views share certain common information and each view carries its unique information, the method leverages both (1) the similarity information of different view pairs and (2) the label information of distinct sample pairs, to learn a latent representation among multiple views. The experimental results on both synthetic and real-world datasets demonstrate that MELTS considerably improves classification accuracy compared to other alternative methods.
			Xiaoyang Li (School of Data Science and Engineering, East China Normal University, Shanghai, China); Martin Pavlovski (Temple University, Philadelphia, Pennsylvania, United States); Fang Zhou (School of Data Science and Engineering, East China Normal University, Shanghai, China); Qiwen Dong (School of Data Science and Engineering, East China Normal University, Shanghai, China); Weining Qian (School of Data Science and Engineering, East China Normal University, Shanghai, China); Zoran Obradovic (Temple University, Philadelphia, Pennsylvania, United States);
4	Research Session #4 (Graphs-I), (April 12 (Tuesday), 2022, 14.45-16.30 , Hall 1)	11	Cascade-Enhanced Graph Convolutional Network for Information Diffusion Prediction [Abstract] Abstract Information diffusion prediction aims to estimate the probability of an inactive user to be activated next in an information diffusion cascade. Existing works predict future user activation either by capturing sequential dependencies within the cascade or leveraging rich graph connections among users. However, most of them perform prediction based on user correlations within the current cascade without fully exploiting diffusion properties from other cascades, which may contain beneficial collaborative patterns for the current cascade. In this paper, we propose a novel Cascade-Enhanced Graph Convolutional Networks (CE-GCN), effectively exploiting collaborative patterns over cascades to enhance the prediction of future infections in the target cascade. Specifically, we explicitly integrate cascades into diffusion process modeling via a heterogeneous graph. Then, the collaborative patterns are explicitly injected into unified user embedding by message passing. Besides, we design a cascade-specific aggregator to adaptively refine user embeddings by modeling different effects of collaborative features from other cascades with the guidance of user context and time context in the current cascade. Extensive experiments on three public datasets demonstrate the effectiveness of the proposed model.
			Ding Wang (School of Cyber Security, University of Chinese Academy of Sciences); Lingwei Wei (School of Cyber Security, University of Chinese Academy of Sciences); Chunyuan Yuan (JD.com, Beijing, China); Yinan Bao (School of Cyber Security, University of Chinese Academy of Sciences); Wei Zhou (School of Cyber Security, University of Chinese Academy of Sciences); Xian Zhu (School of Cyber Security, University of Chinese Academy of Sciences); Songlin hu (School of Cyber Security, University of Chinese Academy of Sciences);
		15	Diversify Search Results through Graph Attentive Document Interaction [Abstract] Abstract The goal of search result diversiﬁcation is to retrieve diverse documents to meet as many diﬀerent information needs as possible. Graph neural networks provide a feasible way to capture the sophisticated relationship between candidate documents, while existing graphbased diversiﬁcation methods require an extra model to construct the graph, which will bring about the problem of error accumulation. In this paper, we propose a novel model to address this problem. Speciﬁcally, we maintain a document interaction graph for the candidate documents of each query to model the diverse information interactions between them. To extract latent diversity features, we adopt graph attention networks (GATs) to update the representation of each document by aggregating its neighbors with learnable weights, which enables our model not dependent on knowing the graph structure in advance. Finally, we simultaneously compute the ranking score of each candidate document with the extracted latent diversity features and the traditional relevance features, and the ranking can be acquired by sorting the scores. Experimental results on TREC Web Track benchmark datasets show that the proposed model outperforms existing state-of-the-art models.
			Xianghong Xu (Shenzhen International Graduate School, Tsinghua University, China); Kai Ouyang (Shenzhen International Graduate School, Tsinghua University, China); Yin Zheng (Department of Search and Application, Weixin Group, Tencent, China); Yanxiong Lu (Department of Search and Application, Weixin Group, Tencent, China); Hai-Tao Zheng (Shenzhen International Graduate School, Tsinghua University, China); Hong-Gee Kim (Seoul National University, South Korea);
		365	Learning Robust Representation through Graph Adversarial Contrastive Learning [Abstract] Abstract Existing studies show that node representations generated by graph neural networks (GNNs) are vulnerable to adversarial attacks, such as unnoticeable perturbations of adjacent matrix and node features. Thus, it is requisite to learn robust representations in graph neural networks. To improve the robustness of graph representation learning, we propose a novel Graph Adversarial Contrastive Learning framework (GraphACL) by introducing adversarial augmentations into graph selfsupervised learning. In this framework, we maximize the mutual information between local and global representations of a perturbed graph and its adversarial augmentations, where the adversarial graphs can be generated in either supervised or unsupervised approaches. Based on the Information Bottleneck Principle, we theoretically prove that our method could obtain a much tighter bound, thus improving the robustness of graph representation learning. Empirically, we evaluate several methods on a range of node classification benchmarks and the results demonstrate GraphACL could achieve comparable accuracy over previous supervised methods.
			Jiayan Guo (School of Artificial Intelligence, Peking University, Beijing, China); , Shangyang Li (Peking-Tsinghua Center for Life Sciences, IDG/McGovern Institute for Brain Research, Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China); , Yue Zhao (Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China); Yan Zhang (School of Artificial Intelligence, Peking University, Beijing, China);
		197	Contrastive Disentangled Graph Convolutional Network for Weakly-supervised Node Classification [Abstract] Abstract Node classification on graph-structured data plays an impor- tant role in many machine learning applications. Recently, Graph Con- volutional Networks (GCNs) have shown remarkable success in the node classification task, due to the ability to aggregate neighborhood infor- mation and propagate supervised signals over the graph. However, most GCN-style models require relatively sufficient labeled data, which are not available in many real-world applications. Therefore, we in this paper study the problem of weakly-supervised node classification and propose a C ontrastive Disentangled Graph C onvolutional N etwork (CDGCN ) to learn disentangled node representations based on the contrastive learn- ing mechanism. Extensive experimental results show that CDGCN sig- nificantly outperforms all baselines on different label sparsities. The code is available at https://github.com/ChuXiaokai/CDGCN.
			Xiaokai Chu (Institute of Computing Technology, Chinese Academy of Sciences, China); Jiashu Zhao (Wilfrid Laurier University, Canada); Xinxin Fan (Institute of Computing Technology, Chinese Academy of Sciences, China); Di Yao (Institute of Computing Technology, Chinese Academy of Sciences, China); Zhihua Zhu (Institute of Computing Technology, Chinese Academy of Sciences, China); Lixin Zou (Baidu Inc., China); Dawei Yin (Baidu Inc., China); Jingping Bi (Institute of Computing Technology, Chinese Academy of Sciences, China);
		353	CSGNN: Improving Graph Neural Networks with Contrastive Semi-Supervised Learning [Abstract] Abstract The Graph Neural Network (GNN) is a rising graph analysis model family that encodes node features into low-dimensional representation vectors by aggregating local neighbor information. Nevertheless, the performance of GNNs is limited since GNNs are trained only over predictions of the labeled data. Hence, eﬀectively incorporating a great number of unlabeled nodes into GNNs will upgrade the performance of GNNs. To address this issue, we propose a Contrastive Semi-supervised learning based GNN (CSGNN) that improves the GNN from extra supervision predicted by contrastive learning. Firstly, CSGNN utilizes multi-loss contrast to learn node representations via maximizing the agreement between nodes, edges and labels of diﬀerent views. Then, a semi-supervised ﬁne-tuner learns from few labeled examples while making the best use of unlabeled nodes. Finally, we introduce the knowledge distillation based on label reliability, which further distills the node labels predicted by contrastive learning into the GNN. Experimentally, CSGNN eﬀectively improves the classiﬁcation performance of GNNs and outperforms other state-of-the-art methods in accuracy over a variety of real-world datasets.
			Yumeng Song (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China); Yu Gu (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China); Xiaohua Li (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China); Chuanwen Li (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China); Ge Yu (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China);
5	Research Session #5 (Text and Image-II), (April 12 (Tuesday), 2022, 14.45-16.30 , Hall 2)	271	Tracking the Evolution: Discovering and Visualizing the Evolution of Literature [Abstract] Abstract A common task in research preparation is to survey related work in bibliographic databases. Scientists are finding the survey task notably difficult as the volume of the databases has been increased con- siderably over the past few decades. Making a good use of a survey paper of the research topic can vastly lower the difficulty but there may be no survey paper in some emerging research topics due to the rapid develop- ment. In this work, we propose a novel Literature Evolution Discovery (LED) process that aims to provide an explainable evolution structure of literature. The explainability is based on co-citation analysis and la- tent relationship extraction, which are done by Steiner tree algorithm and context-consistent factor graph model, respectively. The experiments show the superiority of our context-consistent factor graph model, com- pared with the state-of-the-art baselines. Our case studies and visual- ization results demonstrate the effectiveness and interpretability of our proposed algorithms in practice.
			Siyuan Wu (State Key Laboratory of Internet of Things for Smart City Department of Computer and Information Science University of Macau, Macau SAR, China); Leong Hou U (State Key Laboratory of Internet of Things for Smart City Department of Computer and Information Science University of Macau, Macau SAR, China);
		323	Incorporating Commonsense Knowledge into Story Ending Generation via Heterogeneous Graph Networks [Abstract] Abstract Story ending generation is an interesting and challenging task, which aims to generate a coherent and reasonable ending given a story context. The key challenges of the task lie in how to comprehend the story context suﬃciently and handle the implicit knowledge behind story clues eﬀectively, which are still under-explored by previous work. In this paper, we propose a Story Heterogeneous Graph Network (SHGN) to explicitly model both the information of story context at diﬀerent granularity levels and the multi-grained interactive relations among them. In detail, we consider commonsense knowledge, words and sentences as three types of nodes. To aggregate non-local information, a global node is also introduced. Given this heterogeneous graph network, the node representations are updated through graph propagation, which adequately utilizes commonsense knowledge to facilitate story comprehension. Moreover, we design two auxiliary tasks to implicitly capture the sentiment trend and key events lie in the context. The auxiliary tasks are jointly optimized with the primary story ending generation task in a multi-task learning strategy. Extensive experiments on the ROCStories Corpus show that the developed model achieves new state-of-the-art performances. Human study further demonstrates that our model generates more reasonable story endings.
			Jiaan Wang (School of Computer Science and Technology, Soochow University, Suzhou, China); Beiqi Zou (Department of Computer Science, Princeton University, USA); Zhixu Li (Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University, Shanghai, China); Jianfeng Qu (School of Computer Science and Technology, Soochow University, Suzhou, China); Pengpeng Zhao (School of Computer Science and Technology, Soochow University, Suzhou, China); An Liu (School of Computer Science and Technology, Soochow University, Suzhou, China); Lei Zhao (School of Computer Science and Technology, Soochow University, Suzhou, China);
		400	Open-domain Dialogue Generation Grounded with Dynamic Multi-form Knowledge Fusion [Abstract] Abstract Open-domain multi-turn conversations normally face the challenges of how to enrich and expand the content of the conversation. Recently, many approaches based on external knowledge are proposed to generate rich semantic and information conversation. Two types of knowledge have been studied for knowledge-aware open-domain dialogue generation: structured triples from knowledge graphs and unstructured texts from documents. To take both advantages of abundant unstructured latent knowledge in the documents and the information expansion capabilities of the structured knowledge graph, this paper presents a new dialogue generation model, Dynamic Multi-form Knowledge Fusion based Open-domain Chatt-ing Machine (DMKCM). In particular, DMKCM applies an indexed text (a virtual Knowledge Base) to locate relevant documents as 1st hop and then expands the content of the dialogue and its 1st hop using a commonsense knowledge graph to get apposite triples as 2nd hop. To merge these two forms of knowledge into the dialogue effectively, we design a dynamic virtual knowledge selector and a controller that help to enrich and expand knowledge space. Moreover, DMKCM adopts a novel dynamic knowledge memory module that effectively uses historical reasoning knowledge to generate better responses. Experimental results indicate the effectiveness of our method in terms of dialogue coherence and informativeness.
			Feifei Xu (School of Computer Science and Technology, Shanghai University of Electric Power 201306, China); Shanlin Zhou (School of Computer Science and Technology, Shanghai University of Electric Power 201306, China); Yunpu Ma (Chair of Database Systems and Data Mining, University of Munich 80538, Munich); Xinpeng Wang (School of Electronics and Information Engineering, Tongji University 201804, China); Wenkai Zhang (School of Computer Science and Technology, Shanghai University of Electric Power 201306, China); Zhisong Li (Data Intelligence Center, Alibaba Local Life Service Co., Ltd 200062, China);
		158	Knowing What I Don’t Know: A Generation Assisted Rejection Framework in Knowledge Base Question Answering [Abstract] Abstract Existing Knowledge Base Question Answering (KBQA) systems suffer from the sparsity issue of Knowledge Graphs (KG). To alleviate the issue of KG sparsity, some recent research works introduce external text for KBQA. However, such external information is not always readily available. We argue that it is critical for a KBQA system to know whether it lacks the knowledge to answer a given question. In this paper, we present a novel Generation Assisted Rejection (Gear) framework that identifies unanswerable questions well. Gear can be applied to almost all KBQA systems as an add-on component. Specifically, the backbone of Gear is a sequence-to-sequence model that generates candidate predicates to rerank the original results of a KBQA system. Furthermore, we devise a Probability Distribution Reranking algorithm to ensemble Gear and KBQA since the architectural distinctions between Gear and KBQA are vast. Empirical results and case study demonstrates the effectiveness of our framework in improving the performance of KBQA, particularly in identifying unanswerable questions.
			Junyang Huang (Fudan University, Shanghai, China); Xuantao Lu (Fudan University, Shanghai, China); Jiaqing Liang (Fudan University, Shanghai, China); Qiaoben Bao (Fudan University, Shanghai, China); Chen Huang (CES Finance Co., Ltd., Shanghai, China); Yanghua Xiao (Fudan University, Shanghai, China); Bang Liu (Universit´e de Montr´eal, Montr´eal, Qu´ebec, Canada); Yunwen Chen (DataGrand Inc., Shanghai, China);
		348	PromptMNER: Prompt-based Entity-related Visual Clue Extraction and Integration for Multimodal Named Entity Recognition [Abstract] Abstract Multimodal named entity recognition (MNER) is an emerging task that incorporates visual and textual inputs to detect named entities and predicts their corresponding entity types. However, existing MNER methods often fail to capture certain entity-related but textloosely-related visual clues from the image, which may introduce taskirrelevant noises or even errors. To address this problem, we propose to utilize entity-related prompts for extracting proper visual clues with a pre-trained vision-language model. To better integrate different modalities and address the popular semantic gap problem, we further propose a modality-aware attention mechanism for better cross-modal fusion. Experimental results on two benchmarks show that our MNER approach outperforms the state-of-the-art MNER approaches with a large margin.
			Xuwu Wang (Shanghai Key Lab. of Data Science, School of Computer Science, Fudan University); Junfeng Tian (Alibaba DAMO Academy, Hangzhou, China); Min Gui (Shopee, Singapore, Singapore); Zhixu Li (Shanghai Key Lab. of Data Science, School of Computer Science, Fudan University); Jiabo Ye (East China Normal University, Shanghai, China); Ming Yan (Alibaba DAMO Academy, Hangzhou, China); Yanghua Xiao (Shanghai Key Lab. of Data Science, School of Computer Science, Fudan University);
		423	Towards Personalized Review Generation with Gated Multi-source Fusion Network [Abstract] Abstract Product review generation is an important task in recommender systems, which could provide explanation and persuasiveness for the recommendation. The key challenge of review generation is a lack of semantic information that can be used, resulting in more uncertainty in the generation processing. In this paper we propose a novel encoderdecoder personalized review generation model which aims to effectively and selectively fuse multiple resources including the user/product attributes and their textual information (e.g., historical reviews, short phrases). In the encoder part, we first design an attention layer to accordingly retrieve useful information from the users/product historical reviews since not all reviews are relevant to the target review generation. Note that the attention vector is derived from the semantic features of the short phrases (e.g., review summaries, product titles) which could indicate user preference and product characteristics. In the decoder part, we design a hierarchical gated multi-source network from both intra- and inter-gate perspectives to effectively fuse the multiple context features from encoders. Experimental results on the real-world dataset show that our proposed model could achieve better review generation performance than the baselines, in terms of fluency, readability, and personalization.
			Hongtao Liu (State Key Laboratory of Communication Content Cognition, Beijing, China); Wenjun Wang (State Key Laboratory of Communication Content Cognition, Beijing, China); Hongyan Xu (State Key Laboratory of Communication Content Cognition, Beijing, China); Qiyao Peng (State Key Laboratory of Communication Content Cognition, Beijing, China); Pengfei Jiao (School of Cyberspace, Hangzhou Dianzi University, Hangzhou, China); Yueheng Sun (State Key Laboratory of Communication Content Cognition, Beijing, China);
		506	Definition-Augmented Jointly Training Framework for Intention Phrase Mining [Abstract] Abstract We propose to mine intention phrases from large numbers of queries, for enabling rich query interpretation that identifies both query intentions and associated intention types. We formalize the notion of intention phrase as a sequence of keywords and an intention type, propose its three criteria (relevance, completeness, and clarity), and identify two key challenges. To handle the criterion modeling challenge, we design a jointly training framework with a sequence labeling model and a clarity classification model. To untangle the data sparsity challenge, we are the first to leverage definitions to learn the embeddings of words, discover a new data source (dictionary), and develop the definition-augmented encoder to generate good semantic representations for words and sentences. Our experiments over three large corpora (hotel, tourism, and product domains) verify the advantage of our model over baselines.
			Denghao Ma (Meituan, Beijing, China); Yueguo Chen (DEKE Lab, Renmin University of China, Beijing, China); Changyu Wang (JD.com, Inc., Beijing, China); Hongbin Pei (Xi’an Jiaotong University, Xian, China); Yitao Zhai (Meituan, Beijing, China); Gang Zheng (Meituan, Beijing, China); Qi Chen (Meituan, Beijing, China);
6	Research Session #6 (Text and Image-II), (April 12 (Tuesday), 2022, 14.45-16.30 , Hall 3)	99	Inter- and Intra-Domain Relation-aware Heterogeneous Graph Convolutional Networks for Cross-Domain Recommendation [Abstract] Abstract Cross-domain recommendation aims to model representations of users and items with the incorporation of additional knowledge from other domains, so as to alleviate the data sparsity issue. While recent studies demonstrate the eﬀectiveness of cross-domain recommendation systems, there exist two unsolved challenges: (1) existing methods focus on transferring knowledge to generate shared factors implicitly, which fail to distill domain-shared features from explicit cross-domain correlations; (2) The majority of solutions are unable to eﬀectively fuse domainshared and domain-speciﬁc features. To this end, we propose Inter- and Intra-domain Relation-aware Cross-Domain Recommendation framework (I 2RCDR) to explicitly learn domain-shared representations by capturing high-order inter-domain relations. Speciﬁcally, we ﬁrst construct a cross-domain heterogeneous graph and two single-domain heterogeneous graphs from ratings and reviews to preserve inter- and intradomain relations. Then, a relation-aware graph convolutional network is designed to simultaneously distill domain-shared and domain-speciﬁc features, by exploring the multi-hop heterogeneous connections across diﬀerent graphs. Moreover, we introduce a gating fusion mechanism to combine domain-shared and domain-speciﬁc features to achieve dualtarget recommendation. Experimental results on public datasets show that the eﬀectiveness of the proposed framework against many strong state-of-the-art methods.
			Ke Wang (Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China); Yanmin Zhu (Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China); Haobing Liu (Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China); Tianzi Zang (Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China); Chunyang Wang (Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China); Kuan Liu (Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China);
		294	Knowledge-Enhanced Multi-task Learning for Course Recommendation [Abstract] Abstract Knowledge tracing (KT) aims to model learners’ knowledge level and predict future performance given their past interactions in learning applications. Adaptive learning systems mainly generate course recommendations based on learner’s knowledge level acquired by KT. However, for KT tasks, learners’ forgetting has not been well modeled. In addition, learner’s individual diﬀerences also inﬂuence the accuracy of knowledge level prediction. While for recommendation tasks, most of methods are conducted separately from KT tasks, ignoring the deep connection between them. In this paper, we are motivated to propose a Knowledge-Enhanced Multi-task Learning model for Course Recommendation (KMCR), which regards the improved knowledge tracing task (IKTT) as an auxiliary task to assist the primary course recommendation task (CRT). Speciﬁcally, in IKTT, for assessing dynamic evolving knowledge level, we not only design a personalized controller to enhance the deep knowledge tracing model for modeling learner’s forgetting behavior, but also use personality to model the individual diﬀerences based on the theory of cognitive psychology. In CRT, we adaptively combine learner’s knowledge level obtained by IKTT with their sequential behavior to generate learners’ representation. The experimental results on real-world datasets demonstrate that our approach outperforms related methods in terms of recommendation accuracy.
			Qimin Ban (School of Computer Science and Technology, East China Normal University, Shanghai, China); Wen Wu (School of Computer Science and Technology, East China Normal University, Shanghai, China); Wenxin Hu (School of Data Science and Engineering, East China Normal University, Shanghai, China); Hui Lin (Shanghai Liulishuo Information Technology Co., Ltd., Shanghai, China); Wei Zheng (Information Technology Services, East China Normal University, Shanghai, China); Liang He (School of Computer Science and Technology, East China Normal University, Shanghai, China);
		406	Joint Locality Preservation and Adaptive Combination for Graph Collaborative Filtering [Abstract] Abstract Due to its powerful representation ability, Graph Convolutional Network (GCN) based collaborative ﬁltering (CF), which treats the interaction of user-items as a bipartite graph, has become the upstart in recommender systems. Nevertheless, existing GCNs based recommendation model only compromisingly exploits the shallow relationship (generally less than 4 layers) to represent the user and item with diﬀerent number of interactions, which limits their performance. To address this problem, we propose a novel recommendation framework named joint Locality preservation and Adaptive combination for Graph Collaborative Filtering (LaGCF), which contains two components: locality preservation and adaptive combination, where locality preservation explicitly integrates local features with high-order features at each propagation layer for obtaining better performance faster, while adaptive combination adds adaptive weight and identity matrix in aggregation to enhance the representing power of deep-level GCN. Finally, extensive experiments are conducted on four publicly available datasets, and the results demonstrate the eﬀectiveness and superior performance of our model from both analytical and empirical perspectives.
			Zhiqiang Guo (School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China); Chaoyang Wang (School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China); Zhi Li (School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China); Jianjun Li (School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China); Guohui Li ( School of Software Engineering, Huazhong University of Science and Technology, Wuhan, China);
		447	Hyperbolic Personalized Tag Recommendation [Abstract] Abstract Personalized Tag Recommendation(PTR) aims to automatically generate a list of tags for users to annotate web resources, the socalled items, according to users’ tagging preferences. The main challenge of PTR is to learn representations of involved entities (i.e., users, items, and tags) from interaction data without loss of structural properties in original data. To this end, various PTR models have been developed to conduct representation learning by embedding historical tagging information into low-dimensional Euclidean space. Although such methods are eﬀective to some extent, their ability to model hierarchy, which lies in the core of tagging information structures, is restricted by Euclidean space’s polynomial expansion property. Since hyperbolic space has recently shown its competitive capability to learn hierarchical data with lower distortion than Euclidean space, we propose a novel PTR model that operates on hyperbolic space, namely HPTR. HPTR learns the representations of entities by modeling their interactive relationships in hyperbolic space and utilizes hyperbolic distance to measure semantic relevance between entities. Specially, we adopt tangent space optimization to update model parameters. Extensive experiments on real-world datasets have shown the superiority of HPTR over state-of-the-art baselines.
			Weibin Zhao (State Key Lab for Novel Software Technology, Nanjing University, Nanjing, China); Aoran Zhang (Tongda college, Nanjing University of Posts and Telecommunications, Yangzhou, China); Lin Shang (State Key Lab for Novel Software Technology, Nanjing University, Nanjing, China); Yonghong Yu (Tongda college, Nanjing University of Posts and Telecommunications, Yangzhou, China); Li Zhang (Department of Computer Science, Royal Holloway, University of London, Surrey, UK); Can Wang (School of Information and Communication Technology, Griffith University, Brisbane, Australia); Jiajun Chen (State Key Lab for Novel Software Technology, Nanjing University, Nanjing, China); Hongzhi Yin (School of Information Technology and Electrical Engineering, The University of Queensland, Brisbane, Australia);
		14	Collaborative Filtering for Recommendation in Geometric Algebra [Abstract] Abstract At present, recommender system plays an important role in many practical applications. Many recommendation models are based on representation learning, in which users and items are embedded into a low-dimensional vector space, and then historical interactions are used to train the models. We ﬁnd that almost all of these methods model users and items in real-valued embedding space, which neglect the potential value of other non-real spaces. In this paper, we propose a Geometric Algebra-based Collaborative Filtering (GACF) model for recommendation. Speciﬁcally, GACF ﬁrstly uses multivectors to represent users and items. Then GACF uses geometric product and inner product to model the historical interaction between users and items. By using geometric product, the model prediction can obtain inter-dependencies between components of multivectors, which enable complex interactions between users and items to be captured. Extensive experiments on two real datasets demonstrate the eﬀectiveness of GACF.
			Longcan Wu (Northeastern University, Shenyang, China); Daling Wang (Northeastern University, Shenyang, China); Shi Feng (Northeastern University, Shenyang, China); Kaisong Song (Northeastern University, Shenyang, China); Yifei Zhang (Northeastern University, Shenyang, China); Ge Yu (Northeastern University, Shenyang, China);
		397	SAER: Sentiment-opinion Alignment Explainable Recommendation [Abstract] Abstract Explainable recommendation systems not only provide users with recommended results but also explain why they are recommended. Most existing explainable recommendation methods leverage sentiment analysis to help users understand reasons for recommendation results. They either convert particular preferences into sentiment scores or simply introduce the rating as the overall sentiment into the model. However, the simple rating information cannot provide users with more detailed reasons for recommendations in the explanation. To encode more sentiment information, some methods introduce user opinions into the explanations. As the opinion-based explainable recommendation system does not utilize supervision from sentiment, the generated explanations are generally limited to templates. To solve these issues, we propose a model called Sentiment-opinion Alignment Explainable Recommendation (SAER), which combines sentiment and opinion to ensure that the opinion in the explanation is consistent with the user’s sentiment to the product. Moreover, SAER provides informative explanations with diverse opinions for recommended items. Experiments on real datasets demonstrate that the proposed SAER model outperforms state-of-theart explainable recommendation methods.
			Xiaoning Zong (School of Software, Shandong University, Jinan, China); Yong Liu (Alibaba-NTU Singapore Joint Research Institute, Nanyang Technological University, Singapore); Yonghui Xu (Joint SDU-NTU Centre for Artificial Intelligence Research (C-FAIR), Shandong University, Jinan, China); Yixin Zhang (School of Software, Shandong University, Jinan, China); Zhiqi Shen (School of Computer Science and Engineering, Nanyang Technological University, Singapore); Yonghua Yang (Alibaba Group, Hangzhou, China); Lizhen Cui (School of Software, Shandong University, Jinan, China);
		437	Toward Auto-learning Hyperparameters for Deep Learning based Recommender Systems [Abstract] Abstract Deep learning (DL)-based recommendation system (RS) has drawn extensive attention during the past years. Its performance heavily relies on hy- perparameter tuning. However, the most common approach of hyperparameters tuning is still Grid Search—a tedious task that consumes immerse computation- al resources and human efforts. To aid this issue, this paper proposes a general hyperparameter optimization framework for existing DL-based RSs based on differential evolution (DE), named DE-Opt. Its main idea is to incorporate DE into a DL-based RS model’s training process to auto-learn its hyperparameters λ(regularization coefficient) and η(learning rate) simultaneously at layer- granularity. Empirical studies on three benchmark datasets verify that: 1) DE- Opt is compatible with and can automate the training of the most recent DL- based RSs by making their λ and η adaptively learned, and 2) DE-Opt signifi- cantly outperforms the state-of-the-art hyperparameter searching competitors in terms of both higher learning performance and lower runtime.
			Bo Sun (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences, Chongqing 400714, China); Di Wu (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences, Chongqing 400714, China); Mingsheng Shang (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences, Chongqing 400714, China); Yi He (Department of Computer Science, Old Dominion University, Norfolk, Virginia 23462, USA);
7	Research Session #7 (Spatio-temporal Data), (April 13 (Wednesday), 2022, 10.15-12.30, Hall 1)	200	JS-STDGN: A Spatial-Temporal Dynamic Graph Network using JS-Graph for Traffic Prediction [Abstract] Abstract Traffic prediction is a fundamental operation in real-time traffic analysis. A precise prediction of traffic condition can benefit both road users and traffic management agencies. However, since road traffic is decided by multiple static and dynamic factors, traffic prediction is still a challenging task. As the core indicator of traffic condition, many works focus on traffic speed prediction using time-series forecasting approaches. Although current methods take into account the static road topology while modelling, they fail to consider (1) the semantic closeness between road components and (2) congestion caused by upstream/downstream traffic propagation. In this paper, we introduce a Spatial-Temporal Dynamic Graph Network using JS-Graph, which considers both static road features and dynamic traffic flows when forecasting. Specifically, we first propose a data-driven ’JS-Graph’ method that describes the semantic similarity between road nodes. It models the complex spatial correlations that cannot be captured by the traditional spatial adjacency graph. Secondly, we design a dynamic graph attention network that considers the traffic dynamics that happened in previous time slices when predicting the current one to capture the congestion propagation phenomena. Extensive experiments conducted on real-world datasets show that our proposed method is significantly better than baselines.
			Pengfei Li (Department of Computer Science and Technology, Soochow University, China); Junhua Fang (Department of Computer Science and Technology, Soochow University, China); Pingfu Chao (Department of Computer Science and Technology, Soochow University, China); Pengpeng Zhao (Department of Computer Science and Technology, Soochow University, China); An Liu (Department of Computer Science and Technology, Soochow University, China); Lei Zhao (Department of Computer Science and Technology, Soochow University, China);
		311	When Multitask Learning Make a Difference: Spatio-temporal Joint Prediction for Cellular Trajectories [Abstract] Abstract Spatio-temporal joint prediction aims to simultaneously pre- dict the next location and the corresponding switch time for a cellular trajectory. It requires to consider not only the mutual influence of spatio- temporal predicting tasks, but also the signals related to the intentions of the travel. Although multitask learning can support the joint prediction by considering both spatio-temporal signals, existing approaches neglect the effects of travel intentions and fail to model the long-term dependen- cies in trajectory, resulting in sub-optimal results accordingly. To solve these issues, we propose an intention-aware multitask learning method for spatio-temporal joint prediction, such that predicting travel inten- tion is learned as an auxiliary task. Specifically, due to the implicity of travel intention, we design an effective loss function to learn meaningful intention representation, which can capture trajectory’s future moving goal, so as to provide long-term information for spatio-temporal joint prediction. Furthermore, we carefully design a gating mechanism to fuse sequential and intentional information with different weights to reflect their importance in capturing current movement status. Besides, self- attention network is adopted to model the long-term dependencies of far sampling points in a dense trajectory. Finally, extensive experiments on two trajectory datasets demonstrate the superiority of our method.
			Yuan Xu (Soochow University, Suzhou, China); Jiajie Xu (Soochow University, Suzhou, China); Junhua Fang (Soochow University, Suzhou, China); An Liu (Soochow University, Suzhou, China); Lei Zhao (Soochow University, Suzhou, China);
		7	Efficient Retrieval of Top-k Weighted Spatial Triangles [Abstract] Abstract Due to the proliferation of location-based services and IoT devices, a lot of spatial points are being generated. Spatial data analysis is well known to be an important task. As spatial data analysis tools, graphs consisting of spatial points, where each point has edges to its nearby points and the weight of each edge is the distance between the corresponding points, have been receiving much attention. We focus on triangles (one of the simplest sub-graph patterns) in such graphs and address the problem of retrieving the top-k weighted spatial triangles. This problem has important real-life applications, e.g., group search, ur- ban planning, and co-location pattern mining. However, this problem is computationally challenging, because the number of triangles in a graph is generally huge and enumerating all of them is not feasible. To solve this challenge, we propose an efficient algorithm that returns the exact result. Our experimental results on real datasets show the efficiency of our algorithm.
			Ryosuke Taniguchi (Osaka University, Osaka, Japan); Daichi Amagata (Osaka University, Osaka, Japan); Takahiro Hara (Osaka University, Osaka, Japan);
		47	DIOT: Detecting Implicit Obstacles from Trajectories [Abstract] Abstract In this paper, we study a new data mining problem of obstacle detection. Intuitively, given two kinds of trajectories, i.e., reference and query trajectories, the obstacle is a region such that most query trajectories bypass this region, whereas the reference trajectories go through as usual. We introduce a density-based deﬁnition for the obstacle within a new normalized Dynamic Time Warping distance and the density functions tailored for the sub-trajectories to estimate the density variations. With this deﬁnition, we introduce a novel framework DIOT to detect implicit obstacles. Experimental results show that DIOT can capture the nature of obstacles yet detect the obstacles eﬃciently and eﬀectively.
			Yifan Lei (School of Computing, National University of Singapore, Singapore); Qiang Huang (School of Computing, National University of Singapore, Singapore); Mohan Kankanhalli (School of Computing, National University of Singapore, Singapore); Anthony Tung (School of Computing, National University of Singapore, Singapore);
		267	Exploring Sub-skeleton Trajectories for Interpretable Recognition of Sign Language [Abstract] Abstract Recent advances in tracking sensors and pose estimation software enable smart systems to use trajectories of skeleton joint loca- tions for supervised learning. We study the problem of accurately recog- nizing sign language words, which is key to narrowing the communication gap between hard and non-hard of hearing people. Our method explores a geometric feature space that we call ‘sub-skeleton’ aspects of movement. We assess similarity of feature space trajectories using natural, speed invariant distance measures, which enables clear and insightful nearest neighbor classification. The simplicity and scalability of our basic method allows for immediate application in different data domains with little to no parameter tuning. We demonstrate the effectiveness of our basic method, and a boosted variation, with experiments on data from different application domains and tracking technologies. Surprisingly, our simple methods improve sign recognition over recent, state-of-the-art approaches.
			Joachim Gudmundsson (University of Sydney, Australia); Martin P. Seybold (University of Sydney, Australia); John Pfeifer (University of Sydney, Australia);
		494	Significant Engagement Community Search on Temporal Networks [Abstract] Abstract Community search, retrieving the cohesive subgraph which contains the user-specified query vertex, has been widely touched over the past decades. The existing studies on community search mainly fo- cus on static networks. However, real-world networks, such as scientific cooperation networks and communication networks, usually are tempo- ral networks whose each edge is associated with timestamps. Therefore, the previous methods do not work when handling temporal networks. Inspired by this, we study the problem of identifying the significant en- gagement community to which the user-specified query belongs. Specif- ically, given an integer k and a query vertex u, then we search for the subgraph H which satisfies (i) u ∈ H; (ii) the de-temporal graph of H is a connected k-core; (iii) In H that u has the maximum engagement level. To address our problem, we first develop a top-down greedy peel- ing algorithm named TDGP, which iteratively removes the vertices with the maximum temporal degree. To further boost the efficiency, we then design a bottom-up local search algorithm named BULS with several powerful pruning strategies. Lastly, we empirically show the superiority of our proposed solutions on six real-world temporal graphs.
			Yifei Zhang (National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab, School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, 430074, China); Longlong Lin (National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab, School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, 430074, China); Pingpeng Yuan (National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab, School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, 430074, China); Hai Jin (National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab, School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, 430074, China);
		510	Influence Computation for Indoor Spatial Objects [Abstract] Abstract Studies have shown that people spend more than 85% of their time in indoor spaces. Providing varies location-based services(LBS) for indoor space is of great demand and has drew attentions from both industry and academic. The influence computation for spatial objects is one of the important applications in LBS and has broad applications in indoor facility location and indoor marketing. The influence query has been studied extensively in outdoor spaces. However, due to the fundamental difference between indoor and outdoor space, the outdoor techniques can not be applied for indoor space. In this paper, we propose the first indoor influence computation algorithm IRV to efficiently process indoor influence query. The proposed algorithm is based on the state-of-art indoor index structure VIP-Tree and several pruning rules are also presented to reduce the computation cost. The experiment results on both real and synthetic data sets show our proposed method outperforms the baseline algorithm.
			Yue Li (East China Normal University, Shanghai, China 51194501052@stu.ecnu.edu.cn); Guojie Ma (East China Normal University, Shanghai, China gjma@fem.ecnu.edu.cn); Shiyu Yang (Guangzhou University, Guangzhou, China syyang@gzhu.edu.cn; Liping Wang (East China Normal University, Shanghai, China lipingwang@sei.ecnu.edu.cn); Jiujing Zhang (Guangzhou University, Guangzhou, China jiujingzhang@e.gzhu.edu.cn);
		516	NavFPNet: A Unified Framework for Network Localization in GPS-free Navigation Scenarios [Abstract] Abstract Localization is crucial to mobile navigation, which mainly depends on satellite navigation systems such as GPS. Unfortunately, sometimes GPS localization might be missing due to hardware failure or the urban canyon effect. Network localization is a promising way to improve localization experience for users under these conditions. How- ever, there has not been any report on the large-scale applications of network localization methods to GPS-free outdoor scenarios. In this pa- per, we describe the challenges and build a novel navigation fingerprint based localization system, which fuses road network information and Wi-Fi signals scanned outdoors. With large amount of navigation tra- jectories as supervised information, we train a ranking model to obtain a reasonable metric for localization. The localization system has been deployed in Tencent Map, improving navigation accuracy in GPS-free outdoor scenarios for 1.6 million users.
			Jiazhi Ni (Localization Technology Department, Tencent Inc, Beijing, China); Xin Zhang (Localization Technology Department, Tencent Inc, Beijing, China); Beihong Jin (State Key Laboratory of Computer Sciences, Institute of Software, Chinese Academy of Sciences, Beijing, China); Fusang Zhang (State Key Laboratory of Computer Sciences, Institute of Software, Chinese Academy of Sciences, Beijing, China); Xin Li (Localization Technology Department, Tencent Inc, Beijing, China); Qiang Huang (Localization Technology Department, Tencent Inc, Beijing, China); Pengsen Wang (Localization Technology Department, Tencent Inc, Beijing, China); Xiang Li (Localization Technology Department, Tencent Inc, Beijing, China); Ning Xiao (Localization Technology Department, Tencent Inc, Beijing, China); Youchen Wang (Localization Technology Department, Tencent Inc, Beijing, China); Chang Liu (Localization Technology Department, Tencent Inc, Beijing, China);
8	Research Session #8 (Recommendation-II), (April 13 (Wednesday), 2022, 10.15-12.30, Hall 2)	64	M^3-IB: A Memory-augment Multi-modal Information Bottleneck Model for Next-item Recommendation [Abstract] Abstract Modeling of users and items is essential for accurate recom- mendations. Traditional methods focused only on users’ behavior data for recommendation. Several recent methods attempted to use multi- modal data (e.g. items’ attributes and visual features) to better model users and items. However, these methods fail to model users’ dynamic and personalized preferences on different modalities. In addition, be- sides useful information for recommendation, the multi-modal data also contains a great deal of irrelevant and redundant information that may mislead the learning of recommendation models. To solve these problems, we propose a Memory-augment Multi-Modal Information Bottleneck method, named M 3 -IB, for next item recom- mendation. First, we design a memory network framework to main- tain modality-specific knowledge and capture users’ dynamic modality- specific preferences. Second, we propose to model and fuse users’ per- sonalized preferences on different modalities with a multi-modal proba- bilistic graph. Then, to filter out irrelevant and redundant information in multi-modal data, we extend the information bottleneck theory from single-modal to multi-modal scenario and design a multi-modal infor- mation bottleneck (M2IB) model. Finally, we provide a variational ap- proximation and a flexible implementation of the M2IB model for next item recommendation. Experiential results on five real-world data sets demonstrate the promise of the proposed method.
			Yingpeng Du (School of Software and Microelectronics, Peking University, Beijing, PR China); Hongzhi Liu (School of Software and Microelectronics, Peking University, Beijing, PR China); Zhonghai Wu (National Engineering Research Center of Software Engineering, and Key Lab of High Confidence Software Technologies (MOE), Peking University, Beijing, PR China)
		75	Fully Utilizing Neighbors for Session-based Recommendation with Graph Neural Networks [Abstract] Abstract Session-based recommendation (SBR) has attracted many researchers due to its highly practical value in many online services. Recently, graph neural networks (GNN) are widely applied to SBR due to their superiority on learning better item and session embeddings. However, existing GNN-based SBR models mainly leverage direct neighbors, lacking eﬃcient utilization of multi-hop neighbors information. To address this issue, we propose a multi-head graph attention diﬀusion layer to utilize multi-hop neighbors information. Furthermore, we spot the information loss of local contextual aggregation in existing GNN-based models. To handle this problem, we propose positional graph attention aggregation layer to exploit direct neighbors information. Combining these two designs, we propose a novel model named FUN-GNN (Fully Utilizing Neighbors with Graph Neural Networks) for session-based recommendation. Experiments on three real datasets demonstrate its superiority over existing state-of-the-art baselines in terms of Precision and Mean Reciprocal Rank metrics.
			Xingyu Zhang (School of Computer Science and Shanghai Key Laboratory of Data Science, Fudan University, Shanghai, China); Chaofeng Sha (School of Computer Science and Shanghai Key Laboratory of Data Science, Fudan University, Shanghai, China);
		258	Enhancing Graph Convolution Network for Novel Recommendation [Abstract] Abstract Graph convolution network based recommendation methods have achieved great success. However, existing graph based methods tend to recommend popular items yet neglect tail ones, which are actually the focus of novel recommendation since they can provide more surprises for users and more proﬁts for enterprises. Furthermore, current novelty oriented methods treat all users equally without considering their personal preference on popular or tail items. In this paper, we enhance graph convolution network with novelty-boosted masking mechanism and personalized negative sampling strategy for novel recommendation. Firstly, we alleviate the popularity bias in graph based methods by obliging the learning process to pay more attention to tail items which are assigned to a larger masking probability. Secondly, we empower the novel recommendation methods with users’ personal preference by selecting true negative popular samples. Extensive experimental results on three datasets demonstrate that our method outperforms both graph based and novelty oriented baselines by a large margin in terms of the overall F-measure.
			Xuan Ma (School of Computer Science, Wuhan University, Hubei, China); Tieyun Qian (School of Computer Science, Wuhan University, Hubei, China); Yile Liang (School of Computer Science, Wuhan University, Hubei, China); Ke Sun (School of Computer Science, Wuhan University, Hubei, China); Hang Yun (School of Computer Science, Wuhan University, Hubei, China); Mi Zhang (School of Computer Science, Wuhan University, Hubei, China)
		301	PMAR: Multi-Aspect Recommendation Based on Psychological Gap [Abstract] Abstract Review-based recommendations mainly explore reviews that provide actual attributes of items for recommendation. In fact, besides user reviews, merchants have their descriptions of the items. The inconsistency between the descriptions and the actual attributes of items will bring users psychological gap caused by the Expectation Effect. Compared with the recommendation without merchant’s description, users may feel more unsatisfied with the items (below expectation) or be more impulsive to produce unreasonable consuming (above expectation), both of which may lead to inaccurate recommendation results. In addition, as users attach distinct degrees of importance to different aspects of the item, the personalized psychological gap also needs to be considered. In this work, we are motivated to propose a novel Multi-Aspect recommendation based on Psychological Gap (PMAR) by modelling both user’s overall and personalized psychological gaps. Specifically, we first design a gap logit unit for learning the user’s overall psychological gap towards items derived from textual review and merchant’s description. We then integrate a user-item co-attention mechanism to calculate the user’s personalized psychological gap. Finally, we adopt Latent Factor Model to accomplish the recommendation task. The experimental results demonstrate that our model significantly outperforms the related approaches w.r.t. rating prediction accuracy on Amazon datasets.
			Liye Shi (School of Computer Science and Technology, East China Normal University, Shanghai, China); Wen Wu (School of Computer Science and Technology, Shanghai Key Laboratory of Mental Health and Psychological Crisis Intervention, School of Psychology and Cognitive Science, East China Normal University, Shanghai, China); Yu Ji (School of Computer Science and Technology, East China Normal University, Shanghai, China); Luping Feng (School of Data Science and Engineering, East China Normal University, Shanghai, China); Liang He (School of Computer Science and Technology, East China Normal University, Shanghai, China);
		133	Graph Neural Networks with Dynamic and Static Representations for Social Recommendation [Abstract] Abstract Recommender systems based on graph neural networks receive increasing research interest due to their excellent ability to learn a variety of side information including social networks. However, previous works usually focus on modeling users, not much attention is paid to items. Moreover, the possible changes in the attraction of items over time, which is like the dynamic interest of users are rarely considered, and neither do the correlations among items. To overcome these limitations, this paper proposes graph neural networks with dynamic and static representations for social recommendation (GNN-DSR), which considers both dynamic and static representations of users and items and incorporates their relational inﬂuence. GNN-DSR models the short-term dynamic and long-term static interactional representations of the user’s interest and the item’s attraction, respectively. The attention mechanism is used to aggregate the social inﬂuence of users on the target user and the correlative items’ inﬂuence on a given item. The ﬁnal latent factors of user and item are combined to make a prediction. Experiments on three real-world recommender system datasets validate the eﬀectiveness of GNN-DSR.
			Junfa Lin (School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China); Siyuan Chen (School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China); Jiahai Wang (School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China);
		228	GISDCN: A Graph-based Interpolation Sequential Recommender with Deformable Convolutional Network [Abstract] Abstract Sequential recommendation systems aim to predict users’ next actions based on the preferences learned from their historical behaviors. There are still fundamental challenges for sequential recommender. First, with the popularization of online services, recommender needs to serve both the warm- and cold-start users. However, most existing models depending on user-item interactions lose merits due to the difficulty of learning sequential dependencies with limited interactions. Second, users’ behaviors in their historical sequences are often implicit and complex due to the objective variability of reality and the subjective randomness of users’ intentions. It is difficult to capture the dynamic transition patterns from these user-item interactions. In this work, we propose a graph-based interpolation enhanced sequential recommender with deformable convolutional network (GISDCN). For cold-start users, we re-construct item sequences into a graph to infer users’ possible preferences. To capture the complex sequential dependencies, we employ the deformable convolutional network to generate more robust and flexible filters. Finally, we conduct comprehensive experiments and verify the effectiveness of our model. The experimental results demonstrate that GISDCN outperforms most of the state-of-the-art models at cold-start conditions.
			Yalei Zang (Nanjing University of Aeronautics and Astronautics, Nanjing, China); Yi Liu (Nanjing University of Aeronautics and Astronautics, Nanjing, China); Weitong Chen (The University of Queensland, Brisbane, Australia); Bohan Li (Nanjing University of Aeronautics and Astronautics, Nanjing, China); Aoran Li (Nanjing University of Aeronautics and Astronautics, Nanjing, China); Lin Yue (The University of Queensland, Brisbane, Australia); Weihua Ma (Nanjing University of Aeronautics and Astronautics, Nanjing, China);
		359	Core Interests Focused Self-Attention for Sequential Recommendation [Abstract] Abstract Recent studies identify that sequential recommender systems (SRSs) are improved by self-attention mechanism due to its ability to capture the correlation between interactions. However, two major limitations remain unaddressed by existing works. First, user behaviors in long sequences contain many implicit and noisy preference signals that cannot suﬃciently reﬂect users’ actual preferences. Therefore, modeling all the interactive behaviors will worsen the representation of their actual interests. Second, most models only consider the interaction histories as ordered sequences, while ignoring the time interval between interactions, which leads to a loss of eﬀective information. To tackle these issues, we herein propose CIFARec (Core Interests Focused self-Attention based sequential recommendation), which can explicitly extract those interestrelevant interactions from users’ implicit feedback information and focus on users’ core interests adaptively. Meanwhile, our model takes into account time intervals to retain valid information. Extensive experiments on ﬁve benchmark datasets show that our model outperforms various state-of-the-art sequential models consistently.
			Zhengyang Ai (Institute of Information Engineering and School of Cyber Security, Chinese Academy of Sciences, Beijing, China); Shupeng Wang (Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China); Siyu Jia (Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China); Shu Guo (National Computer Network Emergency Response Technical Team/Coordination Center of China, Beijing, China);
9	Research Session #9 (Applications of ML-II), (April 13 (Wednesday), 2022, 10.15-12.30, Hall 3)	51	CaSS: A Channel-aware Self-supervised Representation Learning Framework for Multivariate Time Series Classification [Abstract] Abstract Self-supervised representation learning of Multivariate Time Series (MTS) is a challenging task and attracts increasing research interests in recent years. Many previous works focus on the pretext task of self-supervised learning and usually neglect the complex problem of MTS encoding, leading to unpromising results. In this paper, we tackle this challenge from two aspects: encoder and pretext task, and propose a unified channel-aware self-supervised learning framework CaSS. Specifically, we first design a new Transformer-based encoder Channel-aware Transformer (CaT) to capture the complex relationships between different time channels of MTS. Second, we combine two novel pretext tasks Next Trend Prediction (NTP) and Contextual Similarity (CS) for the self-supervised representation learning with our proposed encoder. Extensive experiments are conducted on several commonly used benchmark datasets. The experimental results show that our framework achieves new state-of-the-art comparing with previous self-supervised MTS representation learning methods (up to +7.70% improvement on LSST dataset) and can be well applied to the downstream MTS classification. 1
			Yijiang Chen (School of Computer Science, Fudan University, Shanghai, China); Xiangdong Zhou (School of Computer Science, Fudan University, Shanghai, China); Zhen Xing (School of Computer Science, Fudan University, Shanghai, China); Zhidan Liu (School of Computer Science, Fudan University, Shanghai, China); Minyang Xu (School of Computer Science, Fudan University, Shanghai, China and Arcplus Group PLC, Shanghai, China);
		312	Transportation-mode Aware Travel Time Estimation via Meta-learning [Abstract] Abstract Transportation-mode aware travel time estimation (TA-TTE) aims to estimate the travel time of a path in a speciﬁc transportation mode (e.g., walking, driving). Diﬀerent from traditional travel time estimation, TA-TTE requires to consider the heterogeneity of transportation modes due to diﬀerent moving characteristics in diﬀerent modes. As a result, when applying classical travel time estimation models, suﬃcient data is needed for each mode to capture mode-dependent characteristics separately. While in reality, it is hard to obtain enough data in some modes, resulting in a severe data sparsity problem. A practical method to solve this problem is to leverage the mode-independent knowledge (e.g., time for waiting for traﬃc lights) learned from other modes. To this end, we propose a meta-optimized method called MetaMG, which learns well-generalized initial parameters to support eﬀective knowledge transfer across diﬀerent modes. Particularly, to avoid negative transfer, we integrate a spatial-temporal memory in meta-learning to cluster trajectories according to spatial-temporal distribution similarity for enhanced knowledge transfer. Besides, a multi-granularity trajectory representation is adopted in our base model to explore more useful features in diﬀerent spatial granularities while improving the robustness. Finally, comprehensive experiments on real-world datasets demonstrate the superior performance of our proposed method over existing approaches.
			Yu Fan (School of Computer Science and Technology, Soochow University, Suzhou, China); Jiajie Xu (School of Computer Science and Technology, Soochow University, Suzhou, China); Rui Zhou (Swinburne University of Technology, Melbourne, Australia); Chengfei Liu (Swinburne University of Technology, Melbourne, Australia);
		380	Efficient Consensus Motif Discovery of All Lengths in Multiple Time Series [Abstract] Abstract Time series motif discovery is an important primitive for the time series data mining. With the explosion of new sensing technology, there is a continuously increasing amount of time series data in every aspect of our lives, from seismology, entomology, human activity moni- toring, medicine and so on. Considering the rich information included in time series, motif discovery has become an essential part of many data mining tasks. In recent years, the problem of consensus motif discovery in multiple time series begins to appear in our vision. For this task, the existing approaches can only search the consensus motif of a fixed length. However, variable-length motif mining is more common in real applica- tions. To address this problem, the brute force version of the existing fixed-length approach is prohibitively expensive. In this paper, we pro- pose an efficient, scalable and exact algorithm VACOMI to search the consensus motif of all lengths in a given motif length range. We evaluate the performance of VACOMI on four real datasets. The results show that VACOMI can reduce up to 96% of the running time compared with the state-of-the-art approach.
			Mingming Zhang (School of Software, Fudan University, Shanghai, China); Peng Wang (School of Computer Science, Fudan University, Shanghai, China); Wei Wang (School of Computer Science, Fudan University, Shanghai, China);
		459	LiteWSC: a Lightweight Framework for Web-Scale Spectral Clustering [Abstract] Abstract Spectral clustering is an effective clustering method for its excellent performance in partitioning non-linearly distributed data. Nevertheless, it suffers from scalability due to its high computational complexity in constructing the Laplacian graph and computing the corresponding eigendecomposition. In the past decades, many efforts have been made to face this problem. However, the computational bottleneck is still a problem in processing extensive data, especially in web-scale scenarios. We present LiteWSC, a simple yet efficient lightweight spectral clustering framework, to cluster web-scale data with limited resource requirements. Our framework has minimal space overhead and does not require explicit overall embeddings computation. We also analyze the theoretical guarantee and performance boundary of LiteWSC. LiteWSC is highly flexible with O(sp) memory requirement, where s and p are the number of samples and the number of prototypes, which are adaptive to the available resource. Therefore, LiteWSC can partition web-scale data (e.g., n = 8, 000k) in an resource-limited host (e.g., memory is restricted to 1 GB). Experiments on real-world, large-scale and web-scale datasets demonstrate both the efficiency and effectiveness of LiteWSC over state-of-the-art methods.
			Geping Yang (Guangdong University of Technology, Faculty of Computer, China); Sucheng Deng (University of Macau, State Key Laboratory of Internet of Things for Smart City and Department of Computer and Information Science, Macau, China); Yiyang Yang (Guangdong University of Technology, Faculty of Computer, China); Zhiguo Gong (University of Macau, State Key Laboratory of Internet of Things for Smart City and Department of Computer and Information Science, Macau, China); Xiang Chen (Sun Yat-Sen University, School of Electronics and Information Technology, China); Zhifeng Hao (Shantou University, College of Engineering, China)
		40	MetisRL: A Reinforcement Learning Approach for Dynamic Routing in Data Center Networks [Abstract] Abstract As the cornerstone of cloud services, data centers most com- monly adopt multi-rooted tree-like topologies, which provide multiple equal-cost paths but suffer from enormous bandwidth loss due to colli- sions and congestions. In this paper, we present MetisRL, a dynamic flow scheduling system combining the centralized software-defined net- working (SDN) controller and reinforcement learning to balance the net- work traffic. From the global view of the entire network, the SDN con- troller in MetisRL gathers historical traffic matrices and monitors the link utilization. Then the reinforcement learning component in the SDN controller uses such information to perform dynamic prediction and out- put flow scheduling decisions accordingly. We evaluate MetisRL with benchmark tests and compare it with classical flow scheduling schemes, which shows that MetisRL performs well in balancing the flow in multiple paths and can efficiently avoid traffic congestions.
			Yuanning Gao (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China); Xiaofeng Gao (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China); Guihai Chen (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China)
		128	Learning Evolving Concepts with Online Class Posterior Probability [Abstract] Abstract In this paper, we propose a new instance-aware concept drift detection and learning algorithm, called CPP, by dynamically capturing the evolving concepts via online Class Posterior Probability. Instead of comparing distributions of two adaptive windows of data or using prediction performance to infer concept drifts, we model and trace the concept drift by investigating the change of class posterior probability via ensemble of classiﬁer chains directly. Building upon the intuitive concept drift modeling, CPP shows several attractive beneﬁts: (a) It is capable of detecting and learning both gradual and abrupt concept drifts eﬀectively, and thus supports accurate predictions. (b) CPP allows distinguishing real concept drifts from noisy instances or virtual concept drifts. (c) The time-changing concepts are captured and learned at the instance level, and this suggests that CPP lends itself to fast concept drift detection and learning. Empirical results show that our method allows eﬀective and eﬃcient concept drift detection and has good prediction performance compared to many baselines.
			Junming Shao (Yangtze Delta Region Institute (Huzhou) and Data Mining Lab, University of Electronic Science and Technology of China, China); Kai Wang (Yangtze Delta Region Institute (Huzhou) and Data Mining Lab, University of Electronic Science and Technology of China, China); Jianyun Lu (Data Mining Lab, University of Electronic Science and Technology of China, China and School of Artificial Intelligence and Big Data, Chongqing College of Electronic Engineering, China); Zhili Qin (Yangtze Delta Region Institute (Huzhou) and Data Mining Lab, University of Electronic Science and Technology of China, China); Qiming Wangyang (Yangtze Delta Region Institute (Huzhou) and Data Mining Lab, University of Electronic Science and Technology of China, China); Qinli Yang (Yangtze Delta Region Institute (Huzhou) and Data Mining Lab, University of Electronic Science and Technology of China, China);
		290	A Trace Ratio Maximization Method for Parameter Free Multiple Kernel Clustering [Abstract] Abstract Multiple Kernel Clustering (MKC) is helpful to leverage complementary information from various contexts and alleviate the diﬃculty of kernel determination. However, the key weighting strategies for optimal kernel learning around individual kernels are not derived from their optimization problems but embedded in a plug-and-play manner and lead to sub-optimal objective function value. More seriously, the hyperparameters, introduced by the additive balance of these two coupled sub-tasks, are hard to determine in unsupervised learning scenarios and lead to inconsistent and less satisfying results. To avoid the problems mentioned above, we present a novel parameter-free MKC method with the trace ratio criterion (TRMKC in short), which minimizes the approximation errors between consensus and base kernels using the corr-entropy induced metric and maximizes the mean similarities based on the consensus kernel. The trade-oﬀ between these two coupled sub-procedures can be automatically balanced, and the performance could be mutually reinforced. To solve the trace ratio criterion and the corr-entropy induced non-quadratic function optimization problem, we present an alternative strategy with monotonic convergence proof, which reformulates it into a series of sub-problems with trace diﬀerence and quadratic programming by utilizing the half-quadratic optimization technique. Extensive MKC experimental results well demonstrate the eﬀectiveness of TRMKC.
			Yan Chen (School of Computer Science, Sichuan University, Chengdu, China); Lei Wang (School of Computer and Information Technology, Shanxi Univiersity, Taiyuan, China); Liang Du (School of Computer and Information Technology, Shanxi Univiersity, Taiyuan, China and Institute of Big Data Science and Industry, Shanxi Univiersity, Taiyuan, China); Lei Duan (School of Computer Science, Sichuan University, Chengdu, China);
10	Research Session #10 (Algorithms-I), (April 13 (Wednesday), 2022, 14.30-16.15, Hall 1)	181	Improving Information Cascade Modeling by Social Topology and Dual Role User Dependency [Abstract] Abstract In the last decade, information diﬀusion (also known as information cascade) on social networks has been massively investigated due to its application values in many ﬁelds. In recent years, many sequential models including those models based on recurrent neural networks have been broadly employed to predict information cascade. However, the user dependencies in a cascade sequence captured by sequential models are generally unidirectional and inconsistent with diﬀusion trees. For example, the true trigger of a successor may be a non-immediate predecessor rather than the immediate predecessor in the sequence. To capture user dependencies more suﬃciently which are crucial to precise cascade modeling, we propose a non-sequential information cascade model named as TAN-DRUD (Topology-aware Attention Networks with Dual Role User Dependency). TAN-DRUD obtains satisfactory performance on information cascade modeling through capturing the dual role user dependencies of information sender and receiver, which is inspired by the classic communication theory. Furthermore, TAN-DRUD incorporates social topology into two-level attention networks for enhanced information diffusion prediction. Our extensive experiments on three cascade datasets demonstrate that our model is not only superior to the state-of-the-art cascade models, but also capable of exploiting topology information and inferring diﬀusion trees.
			Baichuan Liu (School of Data Science, Fudan University, Shanghai 200433, China); Deqing Yang (School of Data Science, Fudan University, Shanghai 200433, China); Yuchen Shi (School of Data Science, Fudan University, Shanghai 200433, China); Yueyi Wang (School of Data Science, Fudan University, Shanghai 200433, China);
		277	Mining Negative Sequential Rules from Negative Sequential Patterns [Abstract] Abstract As an important tool for behavior informatics, negative sequential rules (NSRs, e.g., < ab >⇒ ¬ < cd >) are sometimes much more informative than positive sequential rules (PSRs, e.g., < ab >⇒ < cd >), as they can provide valuable decision-making information from both negative and positive sides. Very limited NSR mining algorithms are available now and most of them discover NSRs only from positive sequential patterns (PSPs, e.g., < abcd >) rather than from negative sequential patterns (NSPs, e.g., < a¬bc¬d >), which may result in a loss of important information. However, discovering NSRs (e.g., < a¬b >⇒< c¬d >) from NSPs is much more diﬃcult than mining NSRs from PSPs because NSPs do not satisfy the downward closure property. In addition, it is very diﬃcult to ﬁnd which kind of NSRs should be generated. This paper proposes a novel algorithm named nspRule to address all these diﬃculties. The experiment results on real-life and synthetic datasets show that nspRule can mine NSRs correctly and eﬃciently w.r.t. several aspects including rule number, runtime, memory usage and scalability.
			Chuanhou Sun (Department of Computer Science and Technology, Qilu University of Technology(Shandong Academy of Sciences), Jinan, China); Xiaoqi Jiang (Department of Computer Science and Technology, Qilu University of Technology(Shandong Academy of Sciences), Jinan, China); Xiangjun Dong (Department of Computer Science and Technology, Qilu University of Technology(Shandong Academy of Sciences), Jinan, China); Tiantian Xu (Department of Computer Science and Technology, Qilu University of Technology(Shandong Academy of Sciences), Jinan, China); Long Zhao (Department of Computer Science and Technology, Qilu University of Technology(Shandong Academy of Sciences), Jinan, China); Zhao Li (Department of Computer Science and Technology, Qilu University of Technology(Shandong Academy of Sciences), Jinan, China); Yuhai Zhao (Northeastern University, Shenyang, China);
		287	CrossIndex: Memory-Friendly And Session-Aware Index for Supporting Crossfilter in Interactive Data Exploration [Abstract] Abstract Crossfilter, a typical application for interactive data exploration (IDE), is widely used in data analysis, BI, and other fields. However, with the scale-up of the dataset, the real-time response of crossfilter can be hardly fulfilled. In this paper, we propose a memory-friendly and session-aware index called CrossIndex, which can support crossfilterstyle queries with low latency. We first analyze a large number of query workloads generated by previous work and find that queries in the data exploration workload are inter-dependent, which means these queries have overlapped predicates. Based on this observation, this paper defines the inter-dependent queries as a session and builds a hierarchical index that can be used to accelerate crossfilter-style query processing by utilizing the overlapped property of the session to reduce unnecessary search space. Extensive experiments show that CrossIndex outperforms almost all other approaches and meanwhile keeps a low building cost.
			Tianyu Xia (School of Computer Science, Fudan University and Shanghai Key Laboratory of Data Science, Shanghai, China); Hanbing Zhang (School of Computer Science, Fudan University and Shanghai Key Laboratory of Data Science, Shanghai, China); Yinan Jing (School of Computer Science, Fudan University and Shanghai Key Laboratory of Data Science, Shanghai, China); Zhenying He (School of Computer Science, Fudan University and Shanghai Key Laboratory of Data Science, Shanghai, China); Kai Zhang (School of Computer Science, Fudan University and Shanghai Key Laboratory of Data Science, Shanghai, China); X. Sean Wang (School of Computer Science, Fudan University and Shanghai Key Laboratory of Data Science and Shanghai Institute of Intelligent Electronics and Systems, Shanghai, China);
		489	Hierarchical Bitmap Indexing for Range Queries on Multidimensional Arrays [Abstract] Abstract Bitmap indices are widely used in commercial databases for processing complex queries, due to their eﬃcient use of bit-wise operations. Bitmap indices apply natively to relational and linear datasets, with distinct separation of the columns or attributes, but do not perform well on multidimensional array scientiﬁc data. We propose a new method for multidimensional array indexing that considers the spatial component of multidimensional arrays. The hierarchical indexing method is based on sparse n-dimensional trees for dimension partitioning, and bitmap indexing with adaptive binning for attribute partitioning. This indexing performs well on range queries involving both dimension and attribute constraints, as it prunes the search space early. Moreover, the indexing is easily extensible to membership queries. The indexing method was implemented on top of a state of the art bitmap indexing library Fastbit, using tables partitioned along any subset of the data dimensions. We show that the hierarchical bitmap index outperforms conventional bitmap indexing, where an auxiliary attribute is required for each dimension. Furthermore, the adaptive binning signiﬁcantly reduces the amount of bins and therefore memory requirements. 1
			Luboš Krčál (Dept. of Computer Science, Czech Technical University in Prague, Czech Republic); Shen-Shyang Ho (Dept. of Computer Science, Rowan University, Glassboro, NJ, USA); Jan Holub (Dept. of Computer Science, Czech Technical University in Prague, Czech Republic);
		261	A Novel Null-invariant Temporal Measure to Discover Partial Periodic Patterns in Non-uniform Temporal Databases [Abstract] Abstract “Rare item problem” is a fundamental problem in pattern mining. It represents the inability of a pattern mining model to discover the knowledge about frequent and rare items in a database. In the literature, researchers advocated the usage of null-invariant measures as they disclose genuine correlations without being influenced by the object co-absence in the database. Since the existing null-invariant measures consider only an item’s f requency and disregard its temporal occurrence information, they are inadequate to address the rare item problem faced by the partial periodic pattern model. This paper proposes a novel null-invariant measure, called relative periodic-support, to find the patterns containing both frequent and rare items in non-uniform temporal databases. We also introduce an efficient pattern-growth algorithm to find all desired patterns in a database. Experimental results demonstrate that our algorithm is efficient.
			R. Uday Kiran (The University of Aizu, Fukushima, Japan); Vipul Chhabra (IIIT-Hyderabad, Hyderabad, Telangana, India); Saideep Chennupati (IIIT-Hyderabad, Hyderabad, Telangana, India); P. Krishna Reddy (IIIT-Hyderabad, Hyderabad, Telangana, India); Minh-Son Dao (NICT, Tokyo, Japan); Koji Zettsu (NICT, Tokyo, Japan);
		418	A Two-phase Approach for Recognizing Tables with Complex Structures [Abstract] Abstract Tables contain rich multi-dimensional information which can be an important source for many data analytics applications. However, table structure information is often unavailable in digitized documents such as PDF or image files, making it hard to perform automatic analysis over high-quality table data. Table structure recognition from digitized files is a non-trivial task, as table layouts often vary greatly in different files. Moreover, the existence of spanning cells further complicates the table structure and brings big challenges in table structure recognition. In this paper, we model the problem as a cell relation extraction task and propose T2, a novel two-phase approach that effectively recognizes table structures from digitized documents. T2 introduces a general concept termed prime relation, which captures the direct relations of cells with high confidence. It further constructs an alignment graph and employs message passing network to discover complex table structures. We validate our approach via extensive experiments over three benchmark datasets. The results demonstrate T2 is highly robust for recognizing complex table structures.
			Huichao Li (Beijing Institute of Technology, Beijing, China); Lingze Zeng (Beijing Institute of Technology, Beijing, China); Weiyu Zhang (Beijing Institute of Technology, Beijing, China); Jianing Zhang (Beijing Institute of Technology, Beijing, China); Ju Fan (Renmin University of China, Beijing, China); Meihui Zhang (Beijing Institute of Technology, Beijing, China);
		467	A Dynamic Heterogeneous Graph Perception Network with Time-Based Mini-Batch for Information Diffusion Prediction [Abstract] Abstract Information diffusion prediction is an important task to un- derstand how information spreads among users. Most previous studies either only focused on the use of diffusion sequences, or only used so- cial networks between users to make prediction, but such modeling is not sufficient to model the diffusion process. In this paper, we propose a novel Dynamic Heterogeneous Graph Perception Network with Time- Based Mini-Batch (DHGPNTM) that can combine dynamic diffusion graph and social graph for information diffusion prediction. First, we propose a Graph Perception Network (GPN) to learn user embedding in dynamic heterogeneous graphs, and combine temporal information with user embedding to capture users’ dynamic preferences. Then we use a multi-head attention to generate users’ context-dependence embedding, and design a fusion gate to selectively integrate users’ dynamic prefer- ences and context-dependence embedding. The extensive experiments on real datasets demonstrate the effectiveness and efficiency of our model.
			Wei Fan (Heilongjiang University, Harbin, China); Meng Liu (Heilongjiang University, Harbin, China); Yong Liu (Heilongjiang University, Harbin, China);
11	Research Session #11 (Systems), (April 13 (Wednesday), 2022, 14.30-16.15, Hall 2)	310	HEM: A Hardware-aware Event Matching Algorithm for Content-based Pub/Sub Systems [Abstract] Abstract Content-based publish/subscribe (CPS) systems are widely used in many fields to achieve selective data distribution. Event matching is a key component in the CPS system. Many efficient algorithms have been proposed to improve matching performance. However, most of the existing work seldom considers the hardware characteristics, resulting in performance degradation due to a large number of repetitive operations, such as comparison, addition and assignment. In this paper, we propose a Hardware-aware Event Matching algorithm called HEM. The basic idea behind HEM is that we perform as many bit OR operations as possible during the matching process, which is most efficient for the hardware. In addition, we build a performance analysis model that quantifies the trade-off between memory consumption and performance improvement. We conducted extensive experiments to evaluate the performance of HEM. On average, HEM reduces matching time by up to 86.8% compared with the counterparts.
			Wanghua Shi (Shanghai Jiao Tong University, Shanghai, China); Shiyou Qian (Shanghai Jiao Tong University, Shanghai, China);
		396	RotorcRaft: Scalable Follower-Driven Raft on RDMA [Abstract] Abstract State machine replication plays a fundamental role in meeting both the scalability and the fault-tolerance requirement in cloud services. However, the single-point leader is easy to become a bottleneck of scalability because it needs to handle all read and write requests and independently replicate logs in order for all followers. Moreover, machine resources are shared through cloud services where the scale-up of the leader is very expensive. In this paper, we propose a variant of Raft protocol using RDMA named RotorcRaft to significantly offload burden from the leader to followers to relieve the single-point bottleneck. First, RotorcRaft assigns a follower-driven log replication mechanism that exploits hybrid RDMA primitives to relieve part of the burden of leader to followers in log replication. Then, RotorcRaft proposes a quorum follower read that enables followers to handle read requests without the involvement of the leader. Experimental results demonstrate that RotorcRaft has excellent scalability and up to 1.4x higher throughput with 84% latency compared against the state-of-the-art work.
			Xuecheng Qi (East China Normal University, Shanghai, China); Huiqi Hu (East China Normal University, Shanghai, China); Xing Wei (East China Normal University, Shanghai, China); Aoying Zhou (East China Normal University, Shanghai, China);
		545	Efficient Matrix Computation for SGD-based Algorithms on Apache Spark [Abstract] Abstract With the increasing of matrix size in large-scale data analysis, a series of Spark-based distributed matrix computation systems have emerged. Typically, these systems split a matrix into matrix blocks and save these matrix blocks into a RDD. To implement matrix operations, these systems manipulate the matrices by applying coarse-grained RDD operations. That is, these systems load the entire RDD to get a part of matrix blocks. Hence, it may cause the redundant IO when running SGDbased algorithms, since SGD only samples a min-batch data. Moreover, these systems typically employ a hash scheme to partition matrix blocks, which is oblivious to the sampling semantics. In this work, we propose a sampling-aware data loading which uses fine-grained RDD operation to reduce the partitions without sampled data, so as to decrease the redundant IO. Moreover, we exploit a semantic-based partition scheme, which gathers sampled blocks into the same partitions, to further reduce the number of accessed partitions. We modify SystemDS to implement Emacs, efficient matrix computation for SGD-based algorithms on Apache Spark. Our experimental results show that Emacs outperforms existing Spark-based matrix computation systems by 37%.
			Baokun Han (East China Normal University, Shanghai, China); Zihao Chen (East China Normal University, Shanghai, China); Chen Xu (East China Normal University, Shanghai, China); Aoying Zhou (East China Normal University, Shanghai, China);
		204	Parallel Pivoted Subgraph Filtering with Partial Coding Trees on GPU [Abstract] Abstract The pivoted subgraph isomorphism problem is a special subgraph isomorphism problem that focuses on the pivoted nodes rather than the entire subgraphs. The key challenge in adapting existing techniques to the pivoted problem is eliminating their redundant intermediate results. In this paper, we propose a GPU-based pivoted subgraph isomorphism filtering technique, where information of each node is encoded into a series of codes. When performing a pivot subgraph search, the candidate nodes satisfying the coding requirements are collected parallelly on GPU while others are filtered away. Then the final result can be effectively retrieved by a verification process on the filtered nodes. As demonstrated by the experimental results, our method dramatically reduces the processing time of the pivoted subgraph isomorphism problem. Compared to the state-of-the-art GPU-friendly subgraph matching method GpSM which also focuses on filtering effect, the algorithm’s execution time is halved, confirming that our approach can effectively process pivoted subgraph isomorphism queries.
			Yang Wang (Northeastern University, China); Yu Gu (Northeastern University, China); Chuanwen Li (Northeastern University, China);
		317	Txchain: Scaling Sharded Decentralized Ledger via Chained Transaction Sequences [Abstract] Abstract The blockchain has become the most prevalent distributed ledger (DL). Sharding has emerged as a major solution to the scalability bottleneck of DLs. From the underlying data structure of existing sharding schemes, although the Directed Acyclic Graph (DAG)-based topology improves the scalability of DLs compared to chained blocks, the security and reliability of consensus mechanisms in DAG-based DLs have not been verified. Moreover, these schemes suffer from high communication overhead when scaling out. To address these issues, we propose a sharded DL named TxChain, which adopts a novel data structure manipulated by the unit of transaction and constituted by chained transaction sequences of each account. TxChain optimistically processes concurrent transactions and ensures the consistency of all shards via transaction sequence conversion (TSC)-based consensus mechanism. Shards maintain the full replica of Txchain and execute transactions by trustworthy validators, which reduce the frequency of communication with other shards. We theoretically prove the consistency of shards maintained by TSC and demonstrate TxChain’s throughput scales with low latency through extensive experiments.
			Zheng Xu (Fudan University, Shanghai, China); Rui Jiang (Fudan University, Shanghai, China); Peng Zhang (Fudan University, Shanghai, China); Tun Lu (Fudan University, Shanghai, China); Ning Gu (Fudan University, Shanghai, China);
		471	Zebra: An Efficient, RDMA-Enabled Distributed Persistent Memory File System [Abstract] Abstract Distributed file systems (DFSs) play important roles in datacenters. Recent advances in persistent memory (PM) and remote direct memory access (RDMA) technologies provide opportunities in enhancing distributed file systems. However, state-of-the-art distributed PM file systems (DPMFSs) still suffer from a duplication problem and a fixed transmission problem, leading to high network latency and low transmission throughput. To tackle these two problems, we propose Zebra, an efficient RDMA-enabled distributed PM file system — Zebra uses a replication group design for alleviating the heavy replication overhead, and leverages a novel transmission protocol for adaptively transmitting file replications among nodes, eliminating the fixed transmission problem. We implement Zebra and evaluate its performance against state-of-theart distributed file systems on an Intel Optane DC PM platform. The evaluation results show that Zebra outperforms CephFS, GlusterFS, and NFS by 4.38×, 5.61×, and 2.71× on average in throughput, respectively.
			Jingyu Wang (Shanghai Jiao Tong University, Shanghai, China); Shengan Zheng (Shanghai Jiao Tong University, Shanghai, China); Ziyi Lin (Alibaba Group, Hangzhou, China); Yuting Chen (Shanghai Jiao Tong University, Shanghai, China); Linpeng Huang (Shanghai Jiao Tong University, Shanghai, China);
12	Research Session #12 (Applications of ML-III), (April 13 (Wednesday), 2022, 14.30-16.15, Hall 3)	6	Hierarchical Attention Factorization Machine for CTR Prediction [Abstract] Abstract Click-through rate (CTR) prediction is a crucial task in recommender systems and online advertising. The most critical step in this task is to perform feature interaction. Factorization machines are proposed to complete the second-order interaction of features to improve the prediction accuracy, but they are not competent for high-order feature interactions. In recent years, many state-of-the-art models employ shallow neural networks to capture high-order feature interactions to improve prediction accuracy. However, some studies have proven that the addictive feature interactions using feedforward neural networks are inefficient in capturing common high-order feature interactions. To solve this problem, we propose a new way, a hierarchical attention network, to capture high-order feature interactions. Through the hierarchical attention network, we can refine the feature representation of the previous layer at each layer, making it become a new feature representation containing the feature interaction information of the previous layer, and this representation will be refined again through the next layer. We also use shallow neural networks to perform higher-order non-linear interactions on feature interaction terms to further improve prediction accuracy. The experiment results on four real-world datasets demonstrate that our proposed HFM model outperforms state-of-the-art models.
			Lianjie Long (Chongqing University, Chongqing 400044, China); Yunfei Yin (Chongqing University, Chongqing 400044, China); Faliang Huang (Chongqing University, Chongqing 400044, China);
		35	MCRF: Enhancing CTR Prediction Models via Multi-Channel Feature Refinement Framework [Abstract] Abstract Deep learning methods have recently achieved huge success in Click-Through Rate (CTR) prediction tasks mainly due to the ability to model arbitrary-order feature interactions. However, most of the existing methods directly perform feature interaction operations on top of raw representation without considering whether raw feature representation only obtained by embedding layer is sufficiently accurate or not, which leads to suboptimal performance. To address this issue, we design a model-agnostic and lightweight structure named Gated Feature Refinement Layer (GFRL), which dynamically generates flexible and informative feature representation by absorbing both intra-field and contextual information based on specific input instances. We further propose a generalized framework Multi-Channel Feature Refinement Framework (MCRF), which utilizes GFRLs to generate multiple groups of embedding for richer interactions without exploding the parameter space. GFRL and MCRF can be well generalized in many existing CTR prediction methods to improve their performance. Extensive experiments on four real-world datasets show that GFRL and MCRF can achieve statistically significant performance improvements when applied to mainstream CTR prediction algorithms. The integration of MCRF and FM can further push the state-of-the-art performance of CTR prediction tasks. We also identify that refining feature representation is another fundamental but vital direction to boost the performance of CTR prediction models.
			Fangye Wang (Fudan University, Shanghai, China); Hansu Gu (Seattle, United States); Dongsheng Li (Microsoft Research Asia, Shanghai, China); Tun Lu (Fudan University, Shanghai, China); Peng Zhang (Fudan University, Shanghai, China); Ning Gu (Fudan University, Shanghai, China);
		185	Semi-Supervised Graph Learning with Few Labeled Nodes [Abstract] Abstract Graph-based semi-supervised learning, utilizing both a few labeled nodes and massive unlabeled nodes, has aroused extensive attention in the research community. However, for the graph with few labeled nodes, the performance of Graph Convolutional Networks (GCNs) will suffer from a catastrophic decline due to its intrinsic shallow architecture limitation and insufficient supervision signals. To accommodate this issue, we propose a novel Self-Training model (ST-LPGCN) which reinforces the pseudo label generation on the GCNs with Label Propagation algorithm (LPA). By making full use of the advantages of GCNs in aggregating the local node features and LPA in propagating the global label information, our ST-LPGCN improves the generalization performance of GCNs with few labeled nodes. Specifically, we design a pseudo label generator to pick out the nodes assigned with the same pseudo labels by GCN and LPA, and add them to the labeled data for the next selftraining process. To reduce the error propagation of labels, we optimize the transition probability between nodes in LPA under the supervision of the pseudo labels. The extensive experimental results on four real-world datasets validate the superiority of ST-LPGCN for the node classification task with few labeled nodes.
			Cong Zhang (Beijing University of Posts and Telecommunications, China); Ting Bai (Beijing University of Posts and Telecommunications, China); Bin Wu (Beijing University of Posts and Telecommunications, China);
		328	A Deep Reinforcement Learning Based Dynamic Pricing Algorithm in Ride-hailing [Abstract] Abstract Online ride-hailing has become one of the most important transportation ways in the modern city. In the ride-hailing system, the vehicle supply and riding demand is different in different regions, and thus the passengers’ willingness to take a riding service will change dynamically. Traditional pricing strategies cannot make reasonable decisions to set the riding prices with respect to the dynamical supply and demand in different regions, and they cannot make adaptive responses to the real-time unbalanced supply and demand. In addition, the ridehailing platform usually intends to maximize the long-term profit. In this paper, we use deep reinforcement learning to design a multi-region dynamic pricing algorithm to set the differentiate unit price for different regions in order to maximize the long-term profit of the platform. Specifically, we divide the ride-hailing area into several non-overlapping regions, and then propose a model to characterize the passenger’s price acceptance probability. We further model the pricing issue as a Markov decision-making process, and then use deep reinforcement learning to design a multi-region dynamic pricing algorithm (MRDP) to maximize the platform’s long-term profit. We further run extensive experiments based on realistic data to evaluate the effectiveness of the proposed algorithm against some typical benchmark approaches. The experimental results show that MRDP can set the price effectively based on supply and demand to make more profit and can balance the supply and demand to some extent.
			Bing Shi (Wuhan University of Technology, China); Zhi Cao (Wuhan University of Technology, China); Yikai Luo (Wuhan University of Technology, China);
		87	InDISP: An Interpretable Model for Dynamic Illness Severity Prediction [Abstract] Abstract At present, there are a large number of methods for predicting the dynamic illness severity. However, these methods have two limitations: (1) they are not sensitive enough to abrupt changes in illness severity; (2) they are not comprehensive enough to explain the predicted results. To tackle these two challenges, we propose a novel Interpretable model for Dynamic Illness Severity Prediction (InDISP), effectively combining the patient status, structured medical knowledge, and drug usage to predict the trend of Sequential Organ Failure Assessment (SOFA) scores using a temporal convolutional network. The capture of drug usage events makes InDISP get better results with an abrupt change of the SOFA score. In addition, InDISP can explain the influence of features on the prediction results and explore the internal mechanism of their combined influence.
			Xinyu Ma (Southeast University, China); Meng Wang (Southeast University, China); Xing Liu (Central South University, China); Yifan Yang (Tencent, China); Yefeng Zheng (Tencent Jarvis Lab, China); Sen Wang (University of Queensland, Australia);
		249	Generative Adversarial Imitation Learning to Search in Branch-and-Bound Algorithms [Abstract] Abstract Recent studies have shown that reinforcement learning (RL) can provide state-of-the-art performance at learning sophisticated heuristics by exploiting the shared internal structure combinatorial optimization instances in the data. However, existing RL-based methods require too much trial-and-error reliant on sophisticated reward engineering, which is laborious and inefficient for practical applications. This paper proposes a novel framework (RAIL) that combines RL and generative adversarial imitation learning (GAIL) to meet the challenge by searching in branch-and-bound algorithms. RAIL has a policy architecture with dual decoders, corresponding to the sequence decoding of RL and the edge decoding of GAIL, respectively. The two complement each other and restrict each other to improve the learned policy and reward function iteratively.
			Qi Wang (Fudan University, China); Suzanne V. Blackley (Harvard Medical School, Boston, USA); Chunlei Tang (Harvard Medical School, Boston, USA);
		373	Market-aware Dynamic Person-Job Fit with Hierarchical Reinforcement Learning [Abstract] Abstract Person-Job Fit (PJF) is the core of online recruitment. Several recent methods took PJF as a preference-learning problem, and tried to learn two-sided preferences from their historical behaviors. However, they ignored users’ interactive feedbacks (accepted or rejected) received from the other side which may change users’ preferences. In addition, they neglected the status of local market which may affect the final matching results of person-job pairs. To solve these issues, we propose a market-aware dynamic PJF method with hierarchical reinforcement learning (HIRE). We design a two-level hierarchy of reinforcement learning policies. Two low-level policies aim to learn the dynamic preferences of both persons and jobs with consideration of interactive feedbacks, and a high-level policy aims to learn the optimal dynamic matching strategy with consideration of local market state. Extensive experiments on two real-world datasets show the effectiveness of HIRE compared with the state-of-the-art.
			Bin Fu (Peking University, Beijing, China); Hongzhi Liu (Peking University, Beijing, China); Hui Zhao (Peking University, Beijing, China); Yao Zhu (Peking University, Beijing, China); Yang Song (BOSS Zhipin, China); Tao Zhang (BOSS Zhipin, China); Zhonghai Wu (Peking University, Beijing, China);
13	Research Session #13 (Security), (April 13 (Wednesday), 2022, 16.30-18.00, Hall 1)	21	ADAPT: Adversarial Domain Adaptation with Purifier Training for Cross-domain Credit Risk Forecasting [Abstract] Abstract Recent research on transfer learning reveals that adversarial domain adaptation effectively narrows the difference between the source and the target domain distributions, and realizes better transfer of the source domain knowledge. However, how to overcome the intra/interdomain imbalance problems in domain adaptation, e.g. observed in crossdomain credit risk forecasting, is under-explored. The intra-domain imbalance problem results from the extremely limited throngs, e.g., defaulters, in both source and target domain. Meanwhile, the disparity in sample size across different domains leads to suboptimal transferability, which is known as the inter-domain imbalance problem. In this paper, we propose an unsupervised purifier training based transfer learning approach named ADAPT (Adversarial Domain Adaptation with Purifier Training) to resolve the intra/inter-domain imbalance problems in domain adaptation. We also extend our ADAPT method to the multi-source domain adaptation via weighted source integration. We investigate the effectiveness of our method on a real-world industrial dataset on cross-domain credit risk forecasting containing 1:33 million users. Experimental results exhibit that the proposed method significantly outperforms the state-of-the-art methods. Visualization of the results further witnesses the interpretability of our method.
			Guanxiong Zeng (Alibaba Group, China.); Jianfeng Chi (Alibaba Group, China.); Rui Ma (Alibaba Group, China.); Jinghua Feng (Alibaba Group, China.); Xiang Ao (University of Chinese Academy of Sciences, Beijing, China.); Hao Yang (Alibaba Group, China.);
		345	Poisoning Attacks on Fair Machine Learning [Abstract] Abstract Both fair machine learning and adversarial learning have been extensively studied. However, attacking fair machine learning models has received less attention. In this paper, we present a framework that seeks to effectively generate poisoning samples to attack both model accuracy and algorithmic fairness. Our attacking framework can target fair machine learning models trained with a variety of group based fairness notions such as demographic parity and equalized odds. We develop three online attacks, adversarial sampling , adversarial labeling, and adversarial feature modification. All three attacks effectively and efficiently produce poisoning samples via sampling, labeling, or modifying a fraction of training data in order to reduce the test accuracy. Our framework enables attackers to flexibly adjust the attack’s focus on prediction accuracy or fairness and accurately quantify the impact of each candidate point to both accuracy loss and fairness violation, thus producing effective poisoning samples. Experiments on two real datasets demonstrate the effectiveness and efficiency of our framework.
			Minh-Hao Van (University of Arkansas, Fayetteville, USA); Wei Du (University of Arkansas, Fayetteville, USA); Xintao Wu (University of Arkansas, Fayetteville, USA); Aidong Lu (University of North Carolina at Charlotte, USA);
		284	Bi-Level Selection via Meta Gradient for Graph-based Fraud Detection [Abstract] Abstract Graph Neural Networks (GNNs) have achieved remarkable successes by utilizing rich interactions in network data. When applied to fraud detection tasks, the scarcity and concealment of fraudsters bring two challenges: class imbalance and label noise. In addition to overfitting problem, they will compromise model performance through the message-passing mechanism of GNNs. For a fraudster in a neighborhood dominated by benign users, its learned representation will be distorted in the aggregation process. Noises will propagate through the topology structure as well. In this paper, we propose a Bi-Level Selection (BLS) algorithm to enhance GNNs under imbalanced and noisy scenarios observed from fraud detection. BLS learns to select instance-level and neighborhood-level valuable nodes via meta gradient of the loss on an unbiased clean validation set. By emphasizing BLS-selected nodes in the model training process, bias towards majority class (benign) and label noises will be suppressed. BLS can be applied on most GNNs with slight modifications. Experimental results on two real-world datasets demonstrate that BLS can significantly improve GNNs performance on graphbased fraud detection.
			Linfeng Dong (Chinese Academy of Sciences (CAS), Beijing, China); Yang Liu (Chinese Academy of Sciences (CAS), Beijing, China); Xiang Ao (Chinese Academy of Sciences (CAS), Beijing, China); Jianfeng Chi (Alibaba Group, Hangzhou, China); Jinghua Feng (Alibaba Group, Hangzhou, China); Hao Yang (Alibaba Group, Hangzhou, China); Qing He (Chinese Academy of Sciences (CAS), Beijing, China);
		344	Contrastive Learning for Insider Threat Detection [Abstract] Abstract Insider threat detection techniques typically employ supervised learning models for detecting malicious insiders by using insider activity audit data. In many situations, the number of detected malicious insiders is extremely limited. To address this issue, we present a contrastive learning-based insider threat detection framework, CLDet, and empirically evaluate its efficacy in detecting malicious sessions that contain malicious activities from insiders. We evaluate our framework along with state-of-the-art baselines on two unbalanced benchmark datasets. Our framework exhibits relatively superior performance on these unbalanced datasets in effectively detecting malicious sessions.
			Vinay M.S. (University of Arkansas, Fayetteville, USA); Shuhan Yuan (University of Arkansas, Fayetteville, USA); Xintao Wu (Utah State University, Logan, USA);
		355	Metadata Privacy Preservation for Blockchain-based Healthcare Systems [Abstract] Abstract Blockchain-based healthcare systems provide a patient-centric and accountable way to manage and share electronic health records. Thanks to its unique features, the blockchain is employed to record the metadata and to carry out access control. Nevertheless, the transparent nature of blockchain also poses a new challenge to these systems. We identify that the metadata stored on the blockchain leaks the relationship between doctors and patients. Based on this relationship, the adversary can launch linkage attacks to infer patients’ information. Hence, it is necessary to protect metadata privacy. However, strong privacy protection may reduce accountability. Striking a balance between accountability and privacy preservation is a major challenge for the blockchain and its applications. In this paper, we first elaborate on the reasons why the metadata could leak the privacy of patients in blockchain-based healthcare systems. After that, we propose privacy-preserving and accountable protocols to deal with this problem for two different healthcare scenarios: the single doctor case and the group consultation case. Finally, the theoretical analysis demonstrates the practicality of our protocols.
			Lixin Liu (Renmin University of China, Beijing, China); Xinyu Li (The University of Hong Kong, Hong Kong, China); Man Ho AU (The University of Hong Kong, Hong Kong, China); Zhuoya Fan (Renmin University of China, Beijing, China); Xiaofeng Meng (Renmin University of China, Beijing, China);
		384	Blockchain-Based Encrypted Image Storage and Search in Cloud Computing [Abstract] Abstract In the traditional cloud computing system, existing ranked image search mechanisms will leak the similarity of search results, even cannot prevent malicious cloud servers from tampering with or forging results. Therefore, with the help of blockchain technology, we propose an encrypted image storage and search system. A threshold encryption method based on the matrix similarity transformation theory is designed to protect the similarity of search results. Besides, a hybrid storage structure based on blockchain and cloud server is designed to ensure the correctness and completeness of search results. Finally, we implement the prototype in Python and Solidity, and conduct performance evaluations on local Ethereum client. The theoretical analysis and performance evaluations using the real-world dataset demonstrate that our proposed system is secure and feasible.
			Yingying Li (Xidian University, Xi’an, China); Jianfeng Ma (Xidian University, Xi’an, China); Yinbin Miao (Xidian University, Xi’an, China); Ximeng Liu (Fuzhou University, Fuzhou, China); Qi Jiang (Xidian University, Xi’an, China);
14	Research Session #14 (Text and Image-III), (April 13 (Wednesday), 2022, 16.30-18.00, Hall 2)	223	Semantic-based Data Augmentation for Math Word Problems [Abstract] Abstract It’s hard for neural MWP solvers to deal with tiny local variances. In MWP task, some local changes conserve the original semantic while the others may totally change the underlying logic. Currently, existing datasets for MWP task contain limited samples which are key for neural models to learn to disambiguate different kinds of local variances in questions and solve the questions correctly. In this paper, we propose a set of novel data augmentation approaches to supplement existing datasets with such data that are augmented with different kinds of local variances, and help to improve the generalization ability of current neural models. New samples are generated by knowledge guided entity replacement, and logic guided problem reorganization. The augmentation approaches are ensured to keep the consistency between the new data and their labels. Experimental results have shown the necessity and the effectiveness of our methods.
			Ailisi Li (Fudan University, Shanghai, China); Yanghua Xiao (Fudan University, Shanghai, China); Jiaqing Liang (Fudan University, Shanghai, China); Yunwen Chen (DataGrand Inc., Shanghai, China);
		455	Aligning Internal Regularity and External Influence of Multi-Granularity for Temporal Knowledge Graph Embedding [Abstract] Abstract Representation learning for the Temporal Knowledge Graphs (TKGs) is an emerging topic in the knowledge reasoning community. Existing methods consider the internal and external influence at either element level or fact level. However, the multi-granularity information is essential for TKG modeling and the connection in between is also underexplored. In this paper, we propose the method that Aligning-internal Regularity and external Influence of Multi-granularity for Temporal knowledge graph Embedding (ARIM-TE). In particular, to prepare considerate source information for alignment, ARIM-TE first models elementlevel information via the addition between internal regularity and the external influence. Based on the element-level information, the merge gate is introduced to model the fact-level information by combining their internal regularity including the local and global influence with external random perturbation. Finally, according to the above obtained multigranular information of rich features, ARIM-TE conducts alignment for them in both structure and semantics. Experimental results show that ARIM-TE outperforms current state-of-the-art KGE models on several TKG link prediction benchmarks.
			Tingyi Zhang (Soochow University, Suzhou, China); Zhixu Li (Fudan University, China); Jiaan Wang (Soochow University, Suzhou, China); Jianfeng Qu (Soochow University, Suzhou, China); Lin Yuan (Soochow University, Suzhou, China); An Liu (Soochow University, Suzhou, China); Lei Zhao (Soochow University, Suzhou, China); Zhigang Chen (iFLYTEK Research, Suzhou, China);
		518	FalCon: A Faithful Contrastive Framework for Response Generation in TableQA Systems [Abstract] Abstract In a practical TableQA system, response generation is a critical module to generate a natural language description of the SQL and the execution result. Due to the complex syntax of SQL and matching issues with table content, this task is prone to produce factual errors. In this paper, we propose FALCON, a FAithfuL CONtrastive generation framework to improve the factual correctness of generated responses. FALCON forces the generation model to identify examples with factual errors in the latent space during training and takes contrastive examples into consideration during inference. We also propose two new automatic metrics to further evaluate faithfulness specialized to this task. Experimental results show FALCON brings a favorable performance improvement on both automatic and human evaluation amongst various baseline methods.
			Shineng Fang (Fudan University, Shanghai, China); Jiangjie Chen (Fudan University, Shanghai, China); Xinyao Shen (Fudan University, Shanghai, China); Yunwen Chen (DataGrand Inc., Shanghai, China); Yanghua Xiao (Fudan University, Shanghai, China);
		80	Tipster: A Topic-Guided Language Model for Topic-Aware Text Segmentation [Abstract] Abstract The accurate segmentation and structural topics of plain documents not only meet people’s reading habit, but also facilitate various downstream tasks. Recently, some works have consistently given positive hints that text segmentation and segment topic labeling could be regarded as a mutual task, and cooperating with word distributions has the potential to model latent topics in a certain document better. To this end, we present a novel model namely Tipster to solve text segmentation and segment topic labeling collaboratively. We first utilize a neural topic model to infer latent topic distributions of sentences considering word distributions. Then, our model divides the document into topically coherent segments based on the topic-guided contextual sentence representations of the pre-trained language model and assign relevant topic labels to each segment. Finally, we conduct extensive experiments which demonstrate that Tipster achieves the state-of-the-art performance in both text segmentation and segment topic labeling tasks.
			Zheng Gong (University of Science and Technology of China, Hefei, China); Shiwei Tong (University of Science and Technology of China, Hefei, China); Han Wu (University of Science and Technology of China, Hefei, China); Qi Liu (University of Science and Technology of China, Hefei, China); Hanqing Tao (University of Science and Technology of China, Hefei, China); Wei Huang (University of Science and Technology of China, Hefei, China); Runlong Yu (University of Science and Technology of China, Hefei, China);
		155	Predicting Rumor Veracity on Social Media with Graph Structured Multi-task Learning [Abstract] Abstract Previous studies have shown that the multi-task learning paradigm with the stance classification could facilitate the successful detection of rumours, but the shared layers in multi-task learning tend to yield a compromise between the general and the task-specific representation of structural information. To address this issue, we propose a novel Multi-Task Learning framework with Shared Multi-channel Interactions (MTL-SMI), which is composed of two shared channels and two task-specific graph channels. The shared channels extract taskinvariant text features and structural features, and the task-specific graph channels, by interacting with the shared channels, extract the taskenhanced structural features. Experiments on two realworld datasets show the superiority of MTL-SMI against strong baselines.
			Yudong Liu (Beijing University of Posts and Telecommunications,Beijing, China); Xiaoyu Yang (Beijing University of Posts and Telecommunications,Beijing, China); Xi Zhang (Beijing University of Posts and Telecommunications,Beijing, China); Zhihao Tang (Beijing University of Posts and Telecommunications,Beijing, China); Zongyi Chen (Beijing University of Posts and Telecommunications,Beijing, China); Liwen Zhang (Beijing University of Posts and Telecommunications,Beijing, China);
		188	KAAS: A Keyword-Aware Attention Abstractive Summarization Model for Scientific Articles [Abstract] Abstract In this work, we focus on abstractive summarization methods for assisting medical researchers in effectively managing information. Particularly, we introduce a COVID-19-related summarization dataset (COVID-SUM) and propose a novel Keyword-aware Attention Abstractive Summarization (KAAS) model. The KAAS model consists of two encoders and one decoder. As for the encoders, one is a standard article encoder built on transformer layers, while the other one is a hierarchical keyword encoder that first encodes the words in a keyword using BiLSTM, and then passes the keyword representations to a transformer layer to connect the keywords in an example. Additionally, a decoder with keyword-focused attention is utilized to further direct the decoding process to generate comprehensive summaries of the scientific articles. We benchmark several summarization methods on the new COVID-SUM dataset and release this dataset in the hope to promote advances to summarization in the COVID-19 medical area 1. Furthermore, we evaluate the KAAS on COVID-SUM, ArXiv, and PubMed datasets. Experimental results demonstrate that KAAS outperforms several state-of-the-art models on these datasets.
			Shuaimin Li (University of Chinese Academy of Sciences, Beijing, China); Jungang Xu (University of Chinese Academy of Sciences, Beijing, China);
		198	PERM: Pre-training Question Embeddings via Relation Map for Improving Knowledge Tracing [Abstract] Abstract Learning informative embedding for educational question (aka. representations) lies at the core of online learning systems. Recent solutions mainly focus on learning question embedding via the questionconcept bipartite graph. However, the student-question-concept global relation is inadequately exploited. Moreover, finer-grained semantic information from student-question and student-concept interactions should also be further revealed. To this end, in this paper, we propose to Pretrain question Embeddings via Relation Map for knowledge tracing, namely PERM. Extensive experiments conducted on two real-world datasets show that PERM has higher expressive power which enables knowledge tracing methods to effectively predict students’ performance.
			Wentao Wang (Northwest Normal University, China); Huifang Ma (Northwest Normal University, China); Yan Zhao (Northwest Normal University, China); Fanyi Yang (Northwest Normal University, China); Liang Chang (Guilin University of Electronic Technology, China);
15	Research Session #15 (Graphs-II), (April 13 (Wednesday), 2022, 16.30-18.00, Hall 3)	98	On Glocal Explainability of Graph Neural Networks [Abstract] Abstract Graph Neural Networks (GNNs) derive outstanding perfor- mance in many graph-based tasks, as the model becomes more and more popular, explanation techniques are desired to tackle its black-box nature. While the mainstream of existing methods studies instance-level explanations, we propose Glocal-Explainer to generate model-level explanations, which consumes local information of substructures in the input graph to pursue global explainability. Specifically, we investigate faithfulness and generality of each explanation candidate. In the literature, fidelity and infidelity are widely considered to measure faithfulness, yet the two metrics may not align with each other, and have not yet been incorporated together in any explanation technique. On the contrary, generality, which measures how many instances share the same explanation structure, is not yet explored due to the computational cost in frequent subgraph mining. We introduce adapted subgraph mining technique to measure generality as well as faithfulness during explanation candidate generation. Furthermore, we formally define the glocal explanation generation problem and map it to the classic weighted set cover problem. A greedy algorithm is employed to find the solution. Experiments on both synthetic and real-world datasets show that our method produces meaningful and trustworthy explanations with decent quantitative evaluation results.
			Ge Lv (The Hong Kong University of Science and Technology, Hong Kong, China); Lei Chen (The Hong Kong University of Science and Technology, Hong Kong, China); Caleb Chen Cao (Huawei Technologies, Hong Kong, China);
		337	Temporal Network Embedding with Motif Structural Features [Abstract] Abstract Temporal network embedding aims to generate a lowdimensional representation for the nodes in the temporal network. However, the existing works rarely pay attention to the effect of mesodynamics. Only a few works consider the structural identity of the motif, while they do not consider the temporal relationship of the motif. In this paper, we mainly focus on a particular temporal motif: the temporal triad. We propose the Temporal Network Embedding with Motif Structural Features (MSTNE), a novel temporal network embedding method that preserves structural features, including structural identity and temporal relationship of the motif during the evolution of the network. The MSTNE samples the neighbor node based on the temporal triads and models the effects of different temporal triads using the Hawkes process. To distinguish the importance of different structural and temporal triads, we introduce the attention mechanism. We evaluate the performance of MSTNE on four real-world data sets. The experimental results demonstrate that MSTNE achieves the best performance compared to several state-of-the-art approaches in different tasks, including node classification, temporal link prediction, and temporal node recommendation.
			Zhi Qiao (Beijing Key Lab of Network Technology, School of Computer Science and Engineering, Beihang University, Beijing, China); Wei Li (Beijing Key Lab of Network Technology, School of Computer Science and Engineering, Beihang University, Beijing, China); Yunchun Li (Beijing Key Lab of Network Technology, School of Computer Science and Engineering, Beihang University, Beijing, China);
		392	What affects the performance of models? Sensitivity Analysis of Knowledge graph embedding [Abstract] Abstract Knowledge graph (KG) embedding aims to embed entities and relations into a low-dimensional vector space, which has been an active research topic for knowledge base completion (KGC). Recent researchers improve existing models in terms of knowledge representation space, scoring function, encoding method, etc., have achieved progressive improvements. However, the theoretical mechanism behind them has always been ignored. There are few works on sensitivity analysis of embedded models, which is extremely challenging.The diversity of KGE models makes it difficult to consider them uniformly and compare them fairly. In this paper, we first study the internal connections and mutual transformation methods of different KGE models from the generic group perspective, and further propose a unified KGE learning framework. Then, we conduct an in-depth sensitivity analysis on the factors that affect the objective of embedding learning. Specifically, in addition to the impact of the embedding algorithm itself, this article also considers the structural features of the dataset and the strategies of the training method. After a comprehensive experiment and analysis, we can conclude that the Head-to-Tail rate of datasets, the definition of model metric function, the number of negative samples and the selection of regularization methods have a greater impact on the final performance.
			Han Yang (Peking University, Beijing, China); Leilei Zhang (Peking University, Beijing, China); Fenglong Su(National University of Defense Technology, Hunan, China); Jinhui Pang (Beijing Institute of Technology, Beijing, China);
		123	CollaborateCas: Popularity Prediction of Information Cascades based on Collaborative Graph Attention Networks [Abstract] Abstract In recent years, with the prosperity of online social me- dia platforms, cascade popularity prediction has attracted much attention from both academia and industry. Due to the recent advance ingraph representation learning technologies, many state-of-the-art prediction methods utilize graph neural network to predict the cascade popularity. However, a significant disadvantage shared by these methods is that they treat each cascade independently, while the collaborations among different cascades are ignored. Therefore, in this paper we propose a novel deep learning model CollaborateCas which utilizes collaborations among different cascades to learn node and cascade embeddings directly and simultaneously. To this end, we first construct a heterogeneous user-message bipartite graph where different cascades are indirectly connected by common participants. To further capture temporal interdependence among users within each cascade, we construct homogeneous cascade graphs where temporal information is modeled as edge features. Experimental results on two real-world datasets show that our approach achieves significantly higher prediction accuracy compared with state-of-the-art approaches.
			Xianren Zhang (CQU-UC Joint Co-op Institute, Chongqing University, Chongqing, China and Key Laboratory of Dependable Service Computing in Cyber Physical Society, Ministry of Education, Chongqing University, Chongqing, China); Jiaxing Shang (College of Computer Science, Chongqing University, Chongqing, China and Key Laboratory of Dependable Service Computing in Cyber Physical Society, Ministry of Education, Chongqing University, Chongqing, China); Xueqi Jia (College of Computer Science, Chongqing University, Chongqing, China and Key Laboratory of Dependable Service Computing in Cyber Physical Society, Ministry of Education, Chongqing University, Chongqing, China); Dajiang Liu (College of Computer Science, Chongqing University, Chongqing, China and Key Laboratory of Dependable Service Computing in Cyber Physical Society, Ministry of Education, Chongqing University, Chongqing, China); Fei Hao (School of Computer Science, Shaanxi Normal University, Xian, China); Zhiqing Zhang (CQU-UC Joint Co-op Institute, Chongqing University, Chongqing, China);
		356	IncreGNN: Incremental Graph Neural Network Learning by Considering Node and Parameter Importance [Abstract] Abstract Graph Neural Network (GNN) has shown powerful learning and reasoning ability. However, graphs in the real world generally exist dynamically, i.e., the topological structure of graphs is constantly evolving over time. On the one hand, the learning ability of the networks declines since the existing GNNs cannot process the graph streaming data. On the other hand, the cost of retraining GNNs from scratch becomes prohibitively high with the increasing scale of graph streaming data. Therefore, we propose an online incremental learning framework IncreGNN based on GNN in this paper, which solves the problem of high computational cost of retraining GNNs from scratch, and prevents catastrophic forgetting during incremental training. Specifically, we propose a sampling strategy based on node importance to reduce the amount of training data while preserving the historical knowledge. Then, we present a regularization strategy to avoid over-fitting caused by insufficient sampling. The experimental evaluations show the superiority of IncreGNN compared to existing GNNs in link prediction task.
			Di Wei (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China); Yu Gu (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China); Yumeng Song (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China); Zhen Song (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China); Fangfang Li (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China); and Ge Yu (School of Computer Science and Engineering, Northeastern University, Shenyang, Liaoning, China);
		427	Representation Learning in Heterogeneous Information Networks Based on Hyper Adjacency Matrix [Abstract] Abstract Heterogeneous information networks(HINs), which usually contain different kinds of nodes and interactions are very common in real world. The richer semantic information and complex relationships have posed great challenges to current representation learning in HINs. Most existing approaches which use predefined meta-paths suffer from high cost and low coverage. In addition, most of the existing methods cannot capture and learn influential high-order neighbors precisely and effectively. In this paper, we attempt to tackle the problem of meta-paths and influential high-order neighbors by proposing an original method HIN-HAM. HIN-HAM captures influential neighbors of target nodes precisely and effectively by generating the hyper adjacency matrix of the HIN. Then it uses convolutional neural networks with weighted multi-channel mechanism to aggregate different types of neighbors under different relationships. We conduct extensive experiments and comparisons on three real datasets and experimental results show the proposed HIN-HAM outperforms the state-of-the-art methods.
			Bin Yang (School of Software, Fudan University, Shanghai, China); Yitong Wang (School of Software, Fudan University, Shanghai, China);
16	Research Session #16 (Algorithms-II), (April 14 (Thursday), 2022, 10.15-12.30, Hall 1)	194	Discovering Bursting Patterns over Streaming Graphs [Abstract] Abstract A streaming graph is a constantly growing sequence of direct- ed edges, which provides a promising way to detect valuable information in real time. A bursting pattern in a streaming graph represents some interaction behavior which is characterized by a sudden increase in terms of arrival rate followed by a sudden decrease. Mining bursting pattern is essential to early warning of abnormal or notable event. While Bursting pattern discovery enjoys many interesting real-life applications, existing research on frequent pattern mining fails to consider the bursting features in graphs, and hence, may not suffice to provide a satisfactory solution.In this paper, we are the first to address the continuous bursting pattern discovering problem over the streaming graph. We present an auxiliary data structure called BPD for detecting the burst patterns in real time with a limited memory usage. BPD first converts each subgraph into a sequence, and then map it into corresponding tracks based on hash functions to count its frequency in a fixed time window for finding the bursting pattern. Extensive experiments also confirm that our approach generate high-quality results compared to baseline method.
			Qianzhen Zhang (Science and Technology on Information Systems Engineering Laboratory, National University of Defense Technology, Changsha, China) , Deke Guo (Science and Technology on Information Systems Engineering Laboratory, National University of Defense Technology, Changsha, China) , and Xiang Zhao (Science and Technology on Information Systems Engineering Laboratory, National University of Defense Technology, Changsha, China)
		327	GHStore: A High Performance Global Hash based key-value store [Abstract] Abstract Log-Structured Merge tree(LSM-tree) has become the mainstream data structure of persistent key-value(KV) stores, but it suffers from serious write and read amplification. In update intensive workloads, repeated and useless compaction of outdated data makes the problem more serious. So we design an efficient global segmented hashmap to record the level of the latest KV pairs, and we present GHStore based on it, which is a key-value store that improves overall performance in write, read and range query simultaneously for update intensive workloads. A read operation of GHStore does not need to search from top to bottom, and a write-induced compaction operation ignores outdated records. The experiments show that for update intensive workloads, compared to widely-used key-value stores(e.g. RocksDB, Wisckey and PebblesDB), GHStore decreases read latency by 10% - 50%, range query latency by 15% - 60%, while increases write throughput by 4% -55%.
			Jiaoyang Li (Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China and School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China) , Yinliang Yue (Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China and School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China) , and Weiping Wang (Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China and School of Cyber Security, University of Chinese Academy of Sciences, Beijing, China)
		533	Membership Algorithm for Single-Occurrence Regular Expressions with Shuffle and Counting [Abstract] Abstract Since shuffle introduced into regular expressions makes the membership problem NP-hard, and an efficient membership algorithm facilitates processing many membership-based applications, it is an essential work to devise an efficient membership algorithm for regular expressions which can support shuffle. In this paper, we focus on the membership problem for single-occurrence regular expressions with shuffle and counting (SOREFCs). First, we define single-occurrence finite automata with shuffles and counters (SFA(&, #)s), which can recognize the languages defined by SOREFCs. We prove that the membership problem for SFA(&, #)s is decidable in polynomial time. Then, we devise a membership algorithm for SOREFCs by constructing equivalent SFA(&, #)s. Experimental results demonstrate that our algorithm is efficient in membership checking.
			Xiaofan Wang (State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing 100190, China)
		10	(p, n)-core: core decomposition in signed networks [Abstract] Abstract Finding cohesive subgraphs is a key and fundamental prob- lem in network science. k-core is a widely used cohesive subgraph model having many applications. In this paper, we study core decomposition in the signed networks named (p, n)-core that combines k-core and signed edges. (p, n)-core preserves internal sufficient positive edges and deficient negative edges simultaneously to get high-quality cohesive subgraphs. We prove that finding exact (p, n)-core is NP-hard. Therefore, we propose two algorithms with pruning techniques. Finally, we demonstrate the superiority of our proposed algorithms using seven real-world networks.
			Junghoon Kim (Nanyang Technological University, 639798, Singapore) and Sungsu Lim (Chungnam National University, 34134, South Korea)
		120	TROP: Task Ranking Optimization Problem on Crowdsourcing Service Platform [Abstract] Abstract Crowdsourcing has the potential to solve complex problems, especially tasks that are easy for humans but difficult for computers. Service providers of emerging crowdsourcing platforms hope that crowdsourcing tasks on their service platforms can be executed as much as possible in available time. We consider from the perspective of the crowdsourcing service platform and study how to rank tasks to minimize the maximum timeout of tasks. We first formalize the Task Ranking Optimization Problem (TROP) and study its offline version. In the case of online scenario, we propose an Iterative Hungarian Algorithm for Task Ranking Optimization Problem, considering task deadline and click transfer rate with ranking amplification. Experiments on a real crowdsourcing service platform and the simulations based on real datasets demonstrate the superiority of proposed algorithm.
			Jiale Zhang (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China), Haozhen Lu (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China) , Xiaofeng Gao (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China), Ailun Song (Tencent Inc., Shenzhen, China), and Guihai Chen (Tencent Inc., Shenzhen, China)
		240	HATree: A Hotness-Aware Tree Index with In-Node Hotspot Cache for NVM/DRAM-Based Hybrid Memory Architecture [Abstract] Abstract The emerging of Non-Volatile Memory (NVM) has changed the traditional DRAM-only memory system. Compared to DRAM, NVM has the advantages of non-volatility and large capacity. However, as the read/write speed of NVM is still lower than that of DRAM, building DRAM/NVM-based hybrid memory systems is a feasible way of adding NVM into the current computer architecture. This paper aims to optimize the well-known B + -tree for hybrid memory. We present a new index called HATree (Hotness-Aware Tree) that can identify and maintain hot keys with an in-node hotspot cache. The novel idea of HATree is using the unused space of the parent of leaf nodes (PLNs) as the hotspot data cache. Thus, no extra space is needed, but the in-node hotspot cache can efficiently improve query performance. We present the new node structures and operations of HATree and conduct experiments on synthetic workloads using real Intel Optane DC Persistent Memory. The comparative results with three existing state-of-the-art indices, including FPTree, LBTree, and BaseTree, suggest the efficiency of HATree.
			Gaocong Liu (University of Science and Technology of China, Hefei, China) , Yongping Luo (University of Science and Technology of China, Hefei, China) , Peiquan Jin (University of Science and Technology of China, Hefei, China and Key Laboratory of Electromagnetic Space Information, CAS, Hefei, China)
		349	Utilizing Expert Knowledge and Contextual Information for Sample-limited Causal Graph Construction [Abstract] Abstract This paper focuses on causal discovery, which aims at infer- ring the underlying causal relationships from observational samples. Existing methods of causal discovery rely on a large number of samples. So when the number of samples is limited, they often fail to produce correct causal graphs. To address this problem, we propose a novel framework: Firstly, given an expert-specified causal subgraph, we leverage contextual and statistical information of the variables to expand the subgraph with positive-unlabeled learning. Secondly, to ensure the faithfulness of the causal graph, with the expanded subgraph as the constraint, we resort to a structural equation model to discover the entire causal graph. Experimental results show that our method achieves significant improvement over the baselines, especially when only limited samples are given.
			Xuwu Wang (Shanghai Key Lab. of Data Science, School of Computer Science, Fudan University) , Xueyao Jiang (Shanghai Key Lab. of Data Science, School of Computer Science, Fudan University) , Sihang Jiang (Shanghai Key Lab. of Data Science, School of Computer Science, Fudan University), Zhixu Li (Shanghai Key Lab. of Data Science, School of Computer Science, Fudan University), and Yanghua Xiao (Shanghai Key Lab. of Data Science, School of Computer Science, Fudan University and Fudan-Aishu Cognitive Intelligence Joint Research Center, Shanghai, China)
		463	Towards Unification of Statistical Reasoning, OLAP and Association Rule Mining: Semantics and Pragmatics [Abstract] Abstract Over the last decades, various decision support technologies have gained massive ground in practice and theory. Out of these technologies, statistical reasoning was used widely to elucidate insights from data. Later, we have seen the emergence of online analytical processing (OLAP) and association rule mining, which both come with specific rationales and objectives. Unfortunately, both OLAP and association rule mining have been introduced with their own specific formalizations and terminologies. This made and makes it always hard to reuse results from one domain in another. In particular, it is not always easy to see the potential of statistical results in OLAP and association rule mining application scenarios. This paper aims to bridge the artificial gaps between the three decision support techniques, i.e., statistical reasoning, OLAP, and association rule mining and contribute by elaborating the semantic correspondences between their foundations, i.e., probability theory, relational algebra, and the itemset apparatus. Based on the semantic correspondences, we provide that the unification of these techniques can serve as a foundation for designing next-generation multi-paradigm data mining tools.
			Rahul Sharma (Information Systems Group, Tallinn University of Technology, Tallinn, Estonia) , Minakshi Kaushik (Information Systems Group, Tallinn University of Technology, Tallinn, Estonia) , Sijo Arakkal Peious (Information Systems Group, Tallinn University of Technology, Tallinn, Estonia) , Mahtab Shahin (Information Systems Group, Tallinn University of Technology, Tallinn, Estonia), Amrendra Singh Yadav (Vellore Institute of Technology - VIT Bhopal, India), and Dirk Draheim (Information Systems Group, Tallinn University of Technology, Tallinn, Estonia)
17	Research Session #17 (Recommendation-III), (April 14 (Thursday), 2022, 10.15-12.30, Hall 2)	300	Learning Social Influence from Network Structure for Recommender Systems [Abstract] Abstract The purchase decision of users is influenced by their basic preference of items, as well as the social influence of peers. Most social recommendation methods focus on incorporating the semantic collaborative information of social friends. In this paper, we argue that the semantic strength of their friends is also influenced by the subnetwork structure of friendship groups, which had not been well addressed in social recommendation literature. We propose a deep adversarial social model (SoGAN) that can automatically integrate the subnetwork structure of social groups and their semantic information into a unified recommendation framework. Specifically, we first align users in two different views, i.e., the “social-friend” view and “co-purchase” view. Then a generative adversarial network is used to learn the structure information of social groups to enhance the performance of recommender systems. We utilize the structural similarity between two views to produce true samples in SoGAN, and generate the mimic data based on the similarity between the semantic representations of users in two views. By discriminating the true instances based on structure similarity, we naturally inject the structure information into semantic learning of users. Extensive experiments on three real-world datasets, show the superiority of incorporating the social structure impact in recommender systems.
			Ting Bai (Beijing University of Posts and Telecommunications and Key Laboratory of Trustworthy Distributed Computing and Service (BUPT), Ministry of Education), Yanlong Huang (Beijing University of Posts), and Bin Wu (Beijing University of Posts and Telecommunications and Key Laboratory of Trustworthy Distributed Computing and Service (BUPT), Ministry of Education)
		398	Meta-path Enhanced Lightweight Graph Neural Network for Social Recommendation [Abstract] Abstract Social information is widely used in recommender systems to alleviate data sparsity. Since users play a central role in both user- user social graphs and user-item interaction graphs, many previous social recommender systems model the information diffusion process in both graphs to obtain high-order information. We argue that this approach does not explicitly encode high-order connectivity, resulting in potential collaborative signals between user and item not being captured. Moreover, direct modeling of explicit interactions may introduce noises into the model and we expect users to pay more attention to reliable links. In this work, we propose a new recommendation framework named Meta-path Enhanced Lightweight Graph Neural Network (ME-LGNN), which fuses social graphs and interaction graphs into a unified heterogeneous graph to encode high-order collaborative signals explicitly. We consider using a lightweight GCN to model collaborative signals. To enable users to capture reliable information more efficiently, we design meta-paths to further enhance the embedding learning by calculating meta-path dependency probabilities. Empirically, we conduct extensive experiment on three public datasets to demonstrate the effectiveness of our model.
			Hang Miao (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun, China and College of Computer Science and Technology, Jilin University, Changchun,China), Anchen Li (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun, China and College of Computer Science and Technology, Jilin University, Changchun,China) and Bo Yang (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun, China and College of Computer Science and Technology, Jilin University, Changchun,China)
		402	Multi-view Multi-behavior Contrastive Learning in Recommendation [Abstract] Abstract Multi-behavior recommendation (MBR) aims to jointly con- sider multiple behaviors to improve the target behavior’s performance. We argue that MBR models should: (1) model the coarse-grained commonalities between different behaviors of a user, (2) consider both individual sequence view and global graph view in multi-behavior modeling, and (3) capture the fine-grained differences between multiple behaviors of a user. In this work, we propose a novel Multi-behavior Multi-view Contrastive Learning Recommendation (MMCLR) framework, includ- ing three new CL tasks to solve the above challenges, respectively. The multi-behavior CL aims to make different user single-behavior representations of the same user in each view to be similar. The multi-view CL attempts to bridge the gap between a user’s sequence-view and graphview representations. The behavior distinction CL focuses on modeling fine-grained differences of different behaviors. In experiments, we conduct extensive evaluations and ablation tests to verify the effectiveness of MMCLR and various CL tasks on two real-world datasets, achieving SOTA performance over existing baselines. Our code will be available on https://github.com/wyqing20/MMCLR
			Yiqing Wu (Key Lab of Intelligent Information Processing of Chinese Academy of Sciences (CAS), Institute of Computing Technology, CAS, Beijing 100190 and China University of Chinese Academy of Sciences, Beijing 100049, China and WeChat Search Application Department, Tencent, China) , Ruobing Xie (WeChat Search Application Department, Tencent, China), Yongchun Zhu (Key Lab of Intelligent Information Processing of Chinese Academy of Sciences (CAS) and Institute of Computing Technology, CAS, Beijing 100190) , Xiang Ao(Key Lab of Intelligent Information Processing of Chinese Academy of Sciences (CAS) and Institute of Computing Technology, CAS, Beijing 100190) , Xin Chen(WeChat Search Application Department, Tencent, China) , Xu Zhang (WeChat Search Application Department, Tencent, China) , Fuzhen Zhuang (Institute of Artificial Intelligence, Beihang University, Beijing 100191, China and Xiamen Institute of Data Intelligence, Xiamen, China) , Leyu Lin (WeChat Search Application Department, Tencent, China) , and Qing He (Key Lab of Intelligent Information Processing of Chinese Academy of Sciences (CAS) and Institute of Computing Technology, CAS, Beijing 100190)
		452	Diffusion-based Graph Contrastive Learning for Recommendation with Implicit Feedback [Abstract] Abstract Recent studies on self-supervised learning with graph-based recommendation models have achieved outstanding performance. They usually introduce auxiliary learning tasks that maximize the mutual information between representations of the original graph and its augmented views. However, most of these models adopt random dropout to construct the additional graph view, failing to differentiate the importance of edges. The insufficiency of these methods in capturing structural properties of the user-item interaction graph leads to suboptimal recommendation performance. In this paper, we propose a Graph Diffusion Contrastive Learning (GDCL) framework for recommendation to close this gap. Specifically, we perform graph diffusion on the user-item interaction graph. Then, the diffusion graph is encoded to preserve its heterogeneity by learning a dedicated representation for every type of relations. A symmetric contrastive learning objective is used to contrast local node representations of the diffusion graph with those of the useritem interaction graph for learning better user and item representations. Extensive experiments on real datasets demonstrate that GDCL consistently outperforms state-of-the-art recommendation methods.
			Lingzi Zhang, Yong Liu, Xin Zhou, Chunyan Miao, Guoxin Wang, and Haihong Tang
		354	Deep Graph Mutual Learning for Cross-domain Recommendation [Abstract] Abstract Cross-domain recommender systems have been increasingly important for helping users find satisfying items from different domains. However, existing approaches mostly share/map user features among different domains to transfer the knowledge. In fact, user-item interactions can be formulated as a bipartite graph and knowledge transferring through the graph is a more explicit way. Meanwhile, these approaches mostly focus on capturing users’ common interests, overlooking domain-specific preferences. In this paper, we propose a novel Deep Graph Mutual Learning framework (DGML) for cross-domain recommendation. In particular, we first separately construct domain-shared and domain-specific interaction graphs, and develop a parallel graph neural network to extract user preference in corresponding graph. Then the mutual learning procedure uses extracted preferences to form a more comprehensive user preference. Our extensive experiments on two real-world datasets demonstrate significant improvements over state-of-the-art approaches.
			Yifan Wang (School of Computer Science, Peking University, Beijing, China) , Yongkang Li (School of Computer Science, Peking University, Beijing, China), Shuai Li (School of Computer Science, Peking University, Beijing, China), Weiping Song (School of Computer Science, Peking University, Beijing, China) , Jiangke Fan (Meituan) , Shan Gao (Meituan), Ling Ma (Meituan), Bing Cheng (Meituan), Xunliang Cai (Meituan), Sheng Wang (Paul G. Allen School of Computer Science, University of Washington), Ming Zhang (School of Computer Science, Peking University, Beijing, China)
		453	GELibRec: Third-Party Libraries Recommendation Using Graph Neural Network [Abstract] Abstract Third-party libraries have become an indispensable part of the software. The function provided by well-tested third-party libraries can be reused through their programming interfaces, significantly in- creasing developers’ software quality and productivity of developers. However, the vast number of third-party libraries and the complex dependencies are the main obstacles to efficiently exploiting the available resources. So, advanced methods are needed to explore the dependencies between projects and third-party libraries to make meaningful recommendations. This paper proposes GELibRec, which combines graph embedding and collaborative filtering to recommend libraries to developers. We extract a dataset from the open-source dataset libraries.io, and the experimental results show that GELibRec outperforms these methods concerning various quality metrics.
			Chengming Zou (Hubei Key Laboratory of Transportation Internet of Things, Wuhan University of Technology, Wuhan, China and Peng Cheng National Laboratory, Shenzhen, China),Zhenfeng Fan (School of Computer Science and Technology, Wuhan University of Technology, Wuhan, China)
18	Research Session #18 (Applications of ML-IV), (April 14 (Thursday), 2022, 10.15-12.30, Hall 3)	135	Temporal Knowledge Graph Entity Alignment via Representation Learning [Abstract] Abstract Entity alignment aims to construct a complete knowledge graph (KG) by matching the same entities in multi-source KGs. Existing methods mainly focused on the static KG, which assumes that the relationship between entities is permanent. However, almost every KG will evolve over time in practical applications, resulting in the need for entity alignment between such temporal knowledge graphs (TKGs). In this paper, we propose a novel entity alignment framework suitable for TKGs, namely Tem-EA. To incorporate temporal information, we use recurrent neural networks to learn temporal sequence representations. Furthermore, we use graph convolutional network (GCN) and translation-based embedding model to fully learn structural information representation and attribute information representation. Based on these two representations, the entity similarity is calculated separately and combined using linear weighting. To improve the accuracy of entity alignment, we also propose a concept of nearest neighbor matching, which matches the most similar entity pair according to distance matrix. Experiments show that our proposed model has a significant improvement compared to previous methods.
			Xiuting Song (School of Computer and Communication Engineering, Northeastern University (Qinhuangdao), Qinhuangdao 066004, China), Luyi Bai (School of Computer and Communication Engineering, Northeastern University (Qinhuangdao), Qinhuangdao 066004, China), Rongke Liu (School of Computer and Communication Engineering, Northeastern University (Qinhuangdao), Qinhuangdao 066004, China), Han Zhang (School of Computer and Communication Engineering, Northeastern University (Qinhuangdao), Qinhuangdao 066004, China)
		208	Human Mobility Identification by Deep Behavior Relevant Location Representation [Abstract] Abstract This paper focuses on Trajectory User Link (TUL), which aims at identifying user identities through exploiting their mobility pat- terns. Existing TUL approaches are based on location representation, a way to learn location associations by embedding vectors that can indicate the level of semantic similarity between the locations. However, existing methods for location representation don’t consider the semantic diversity of locations, which will lead to a misunderstanding of the semantic information of trajectory when linking anonymous trajectories to candidate users. To solve this problem, in this paper, we propose Deep Behavior Relevant Location representation (DBRLr) to map the polysemous locations into distinct vectors, from the perspective of users’ behavior to reflect the semantic polysemy of locations. To learn this representation, we build a Location Prediction-based Movement Model (LP-based MM), which learns user behavior representation at each visited location from a large history trajectory corpora. LP-based MM considers both Continuity and Cyclicity characteristics of user’s movement. We employ the combination of the intermediate layer representation in LP-based MM as DBRLr. An effective recurrent neural network is used to link anonymous trajectories with candidate users. Experiments are conducted on two real-world datasets, and the result shows that our method performs beyond existing methods.
			Tao Sun (Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China and School of Computing Science and Technology, University of Chinese Academy of Sciences, Beijing, China), Fei Wang (Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China) , Zhao Zhang (Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China), Lin Wu (Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China), and Yongjun Xu (Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China)
		469	Dual Confidence Learning Network for Open-World Time Series Classification [Abstract] Abstract Traditional time series classification methods hold a basic closed-world assumption that the classes of test examples must have been seen by the classifier. However, in the real world, new examples that do not belong to any training class are constantly generated. The existing open-world classification methods cannot be well applied to time series data because they mainly solve problems in the field of computer vision without mining temporal features. In addition, these methods are mostly based on overconfident prediction results, and an overconfident classifier may produce predictive probabilities with high confidence even for incorrect predictions. In this paper, we propose a novel dual confi- dence learning network for open-world time series classification, which leverages temporal deep neural model to capture temporal features, and designs a dual confidence mechanism to identify which known class or unknown class an example belongs to. Specifically, temporal confidence is learned from the likelihood sequence to reflect the true correctness risk and Weibull distribution confidence is learned to reflect the open space risk. Experimental results evaluated on real-world datasets demonstrate that the proposed model can accurately identify examples of unknown classes without sacrificing the classification performance of known class examples.
			Junwei Lv (Key Laboratory of Knowledge Engineering with Big Data, Ministry of Education, Hefei University of Technology, Hefei 230009, China and School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China), Ying He (Key Laboratory of Knowledge Engineering with Big Data, Ministry of Education, Hefei University of Technology, Hefei 230009, China and School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China), Xuegang Hu (Key Laboratory of Knowledge Engineering with Big Data, Ministry of Education, Hefei University of Technology, Hefei 230009, China and School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China), Desheng Cai (School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China), Yuqi Chu (School of Computer Science and Information Engineering, Hefei University of Technology, Hefei 230009, China) and Jun Hu (National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China)
		530	Port Container Throughput Prediction Based on Variational AutoEncoder [Abstract] Abstract The prediction of port container throughput has a significant impact on many of the port’s operations. However, accurate prediction of throughput is a difficult problem due to the complexity of the port environment and the uncertainty of port operations. In this paper, we proposed an approach combining self-attention mechanism and variational autoencoder to forecast the operating time of each container. First, we used self-attention mechanism to capture the features between adjacent containers. Then to reduce the influence of missing data, we designed a variational autoencoder(VAE) module to model the latent variables in the port. Finally, the output layer combined the results of these two parts to obtain the final forecast of the loading and discharging time of containers. The throughput of the entire port can be inferred from the forecasted container operation time. Furthermore, we also proposed dynamic programming algorithms to estimate the distribution of the throughput with the help of variational autoencoder module. Experiment results on port throughput prediction in the real-world datasets show that our approach has superior performance at prediction accuracy. Moreover, experiments conducted at different time intervals demonstrate the effectiveness of our approach on various time scales. And the effectiveness of the dynamic programming algorithms is demonstrated through our case study.
			Jingze Li (School of Computer Science, Fudan University, Shanghai, China, Shanghai Key Laboratory of Data Science, Fudan University, Shanghai, China and Shanghai Institute of Intelligent Electronics & Systems, Fudan University, Shanghai, China) , Shengmin Shi ( School of Computer Science, Fudan University, Shanghai, China, Shanghai Key Laboratory of Data Science, Fudan University, Shanghai, China and Shanghai Institute of Intelligent Electronics & Systems, Fudan University, Shanghai, China) , Tongbing Chen (School of Computer Science, Fudan University, Shanghai, China, Shanghai Key Laboratory of Data Science, Fudan University, Shanghai, China and Shanghai Institute of Intelligent Electronics & Systems, Fudan University, Shanghai, China), Yu Tian (Shanghai International Port (Group) Co., Ltd., Shanghai, China) ,Yihua Ding (Shanghai Harbor e-logistics Software Co., Ltd., Shanghai, China), Yiyong Xiao (NeZha Smart Port&Shipping Technology (Shanghai) Co., Ltd., Shanghai, China), and Weiwei Sun (School of Computer Science, Fudan University, Shanghai, China, Shanghai Key Laboratory of Data Science, Fudan University, Shanghai, China and Shanghai Institute of Intelligent Electronics & Systems, Fudan University, Shanghai, China)
		150	Robust Dynamic Pricing in Online Markets with Reinforcement Learning [Abstract] Abstract In online markets, reinforcement learning (RL) is a promis- ing way for dynamic pricing, due to its ability in maximizing long-term cumulative return. However, directly optimizing RL policies in the on- line markets can be costly since RL requires trial-and-error with the environment, which may lead to drastic revenue loss. In this paper, we propose a robust dynamic pricing algorithm using RL. The main idea is to train the dynamic pricing policy in an adversarial simulation environment built with a generative adversarial framework. In this framework, the generator is trained to: 1) imitate real customers behaviors; 2) generate adversarial behaviors. The algorithm is proved to converge under certain assumptions. The experiment results show that our algorithm can be comparable with the algorithm directly trained in the real environment. Moreover, it outperforms other baseline significantly in different scenarios.
			Bolei Zhang (School of Computer, Nanjing University of Posts and Telecommunications Nanjing, PR China and State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, PR China) and Fu Xiao (School of Computer, Nanjing University of Posts and Telecommunications Nanjing, PR China)
		246	Multi-Memory enhanced Separation Network for Indoor Temperature Prediction [Abstract] Abstract Indoor temperature prediction is vital to predictive control on district heating systems. Due to the data collection in practice, there always exist residential areas with limited historical data. Transferring the knowledge from residential areas with sufficient data is of great help to address the data scarcity problem. However, it is still challenging as the data distribution shifts among residential areas and shifts over time. In this paper, we proposed a Multi-Memory enhanced Separation Network (MMeSN) to predict indoor temperature for residential areas with limited data. MMeSN is a parameter-based multi-source transfer learning method, mainly consisting of two components: Source Knowledge Memorization and Memory-enhenced Aggregation. Specifically, the former component jointly decouples the domain-independent & domain-specific information which separately memorize the specific historical patterns for each source. The latter component memorizes the historical relationships between the target and multiple sources and further aggregates the domain-specific & domain-independent information. We conduct extensive experiments on a real-world dataset, and the results demonstrate the advantages of our approach.
			Zhewen Duan (Xidian University, Xi’an, JD Intelligent Cities Research and JD iCity, JD Technology, Beijing, China) , Xiuwen Yi (Xidian University, Xi’an and JD Intelligent Cities Research) , Peng Li , Dekang Qi (Xidian University, Xi’an and JD Intelligent Cities Research and Southwest Jiaotong University, Chengdu), Yexin Li (Xidian University, Xi’an and JD Intelligent Cities Research), Haoran Xu (Xidian University, Xi’an and JD Intelligent Cities Research), Yanyong Huang (Southwestern University of Finance and Economics, Chengdu), Junbo Zhang (Xidian University, Xi’an and JD Intelligent Cities Research) , and Yu Zheng (Xidian University, Xi’an, JD Intelligent Cities Research and JD iCity, JD Technology, Beijing, China)
		422	TEALED: A Multi-Step Workload Forecasting Approach Using Time-Sensitive EMD and Auto LSTM Encoder-Decoder [Abstract] Abstract Many data-driven methods and machine learning techniques are constantly being applied to the database management system (DBMS), which are based on the judgment of future workloads to achieve a better tuning result. We propose a novel multi-step workload forecasting approach named TEALED which applies time-sensitive empirical mode decomposition and auto long short-term memory based encoder-decoder to predict resource utilization and query arrival rates for DBMSs. We first improve the empirical mode decomposition method by considering time translation and extending short series. Then we utilize the encoder-decoder network to extract features from decomposed workloads and generate workload predictions. Moreover, we combine hyperparameter search technologies to guarantee performance under varying workloads. The experiment results show the effectiveness and robustness of TEALED, and indicate the ability of multi-step workload forecasting.
			Xiuqi Huang (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China), Yunlong Cheng (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China), Xiaofeng Gao (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China), and Guihai Chen (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China)
19	Research Session #19 (Knowledge Bases), (April 14 (Thursday), 2022, 14.30-16.15, Hall 1)	44	Triple-as-Node Knowledge Graph and Its Embeddings [Abstract] Abstract Knowledge Graphs (KGs) aim at semantically representing the world’s knowledge in the form of machine-readable graphs composed of subject-relation-object triples (facts). However, most previous KGs only consider the relationship between individual entities, ignoring connections between facts and entities, which are commonly used to depict useful information about the properties of facts. To this end, we formally introduce FactKG, a new KG form which incorporates fact nodes and extends relations from entity-level to fact-level. This new structure challenges some previous KG techniques. One of the key challenges to FactKG is how to learn compatible representation of entities and facts. In this paper, we mainly focus on the embedding task of FactKG. We contribute a benchmark WD16K with additional fact-relevant relations, and a framework FactE, which can represent facts, entities and relations in the same space via attention. Experiments demonstrate that FactE not only significantly outperforms state-of-the-art models but also brings remarkable benefits for disambiguation of 1-N relations, revealing its potential usefulness.
			Xin Lv (Department of Computer Science and Technology, BNRist and KIRC, Institute for Artificial Intelligence, Tsinghua University, Beijing 100084, China), Jiaxin Shi (Department of Computer Science and Technology, BNRist and KIRC, Institute for Artificial Intelligence, Tsinghua University, Beijing 100084, China), Shulin Cao (Department of Computer Science and Technology, BNRist and KIRC, Institute for Artificial Intelligence, Tsinghua University, Beijing 100084, China), Lei Hou (Department of Computer Science and Technology, BNRist and KIRC, Institute for Artificial Intelligence, Tsinghua University, Beijing 100084, China), and Juanzi Li (Department of Computer Science and Technology, BNRist and KIRC, Institute for Artificial Intelligence, Tsinghua University, Beijing 100084, China)
		170	LeKAN: Extracting Long-tail Relations via Layer-Enhanced Knowledge-Aggregation Networks [Abstract] Abstract Long-tailed relation extraction is a crucial task in the infor- mation extraction field for extracting the long-tailed, imbalanced relation between two annotated entities based on related context. Although many works have been devoted to distinguishing valid instances from noisy data and have achieved promising performance, such studies still have critical defects: works based on nonhierarchical relations ignore the correlations among the relations, and those based on hierarchical relations neglect the hierarchy of the relation structure, which is unbalanced and causes difficulty in extracting data-poor classes. In this paper, a novel layer- enhanced knowledge aggregation network, named LeKAN, is presented to classify the relations between two annotated entities from text, especially long-tailed relations, which are very common in various corpora. Inspired by the election mechanism, we aggregate the ancestors of long-tailed relation classes into new relation representations to prevent the long-tailed relations from being ignored. Specifically, we use GraphSAGE to learn the relational knowledge from an existing knowledge graph via class embedding. Moreover, we aggregate the acquired relational knowledge into the LeKAN by layer-enhanced knowledge-aggregating attention mechanism. Comprehensive experimental results demonstrate that the new method yields considerable improvement over other relation extraction methods on a large-scale benchmark dataset with a long-tailed distribution.
			Xiaokai Liu (National Engineering Research Center for Big Data Technology and System,Services Computing Technology and System Lab, Cluster and Grid Computing Lab and School of Cyber Science and Engineering, Huazhong University of Science and Technology, China), Feng Zhao (National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab and School of Computer Science and Technology, Huazhong University of Science and Technology, China), Xiangyu Gui (National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab and School of Computer Science and Technology, Huazhong University of Science and Technology, China), Hai Jin (National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab and School of Computer Science and Technology, Huazhong University of Science and Technology, China)
		264	TRHyTE: Temporal Knowledge Graph Embedding based on Temporal-Relational Hyperplanes [Abstract] Abstract Temporal Knowledge Graph Embedding (TKGE) aims at encoding evolving facts with high-dimensional vectorial representations. Although a representative hyperplane-based TKGE approach, namely HyTE, has achieved remarkable performance, it still suffers from several problems including (i) ignorance of latent temporal properties and diversity of relations; (ii) neglect of temporal dependency between adjacent hyperplanes; (iii) inefficient static random negative sampling method; (iv) incomplete testing on partial time information. To address these issues, we propose TRHyTE, a novel Temporal-Relational Hyperplane based TKGE model, which defines three typical properties, including interval, open-interval, and instantaneousness temporal, for relations and correspondingly constructs three relational sub-KGs, supporting distinguishing learning for facts. Within each sub-KG, TRHyTE transforms entities into relation space first, and then explicitly projects transformed entities and relations into temporal-relational hyperplanes to learn time-relation-aware embeddings. Moreover, Gate Recurrent Unit is leveraged to simulate TKG evolution so as to capture temporal dependency between adjacent hyperplanes. Additionally, we develop a dynamic negative samples mechanism for robust training. In testing phase, an expand-and-best-merge strategy is crafted to realize a complete testing on all valid time intervals. Extensive experiments on two well-known benchmarks verify the effectiveness of our proposals.
			Lin Yuan (School of Computer Science and Technology, Soochow University, China), Zhixu Li (Shanghai Key Lab of Data Science, School of Computer Science, Fudan University), Jianfeng Qu (School of Computer Science and Technology, Soochow University, China), Tingyi Zhang (School of Computer Science and Technology, Soochow University, China),An Liu (School of Computer Science and Technology, Soochow University, China), Lei Zhao (School of Computer Science and Technology, Soochow University, China), and Zhigang Chen (State Key Laboratory of Cognitive Intelligence, iFLYTEK, China and iFLYTEK Research, Suzhou, China)
		168	ExKGR: Explainable Multi-hop Reasoning for Evolving Knowledge Graph [Abstract] Abstract Knowledge graph reasoning is a popular approach to pre- dict new facts in knowledge graphs (KGs) suffering from inherent in- completeness. Compared with the popular embedding-based approach,multi-hop reasoning approach is more interpretable. Multi-hop reasoning can be modeled as reinforcement learning (RL) in which the RL agent navigates in the KG. Despite high interpretability, the knowledge in real world evolves by the minute, previous approaches are based on static KG. To address the above challenges, we propose an explainable multi-hop reasoning approach (ExKGR) for practical scenario, aiming to reason the emerging entity in evolving KGs and provide evidentiary reasoning paths. Specifically, ExKGR can represent emerging entities by inductive learning of neighbors and the query. Furthermore, we restrict the RL action space of supernodes. Also, we use a dynamic reward instead of a binary reward in prior approaches. The experimental results on four benchmark datasets demonstrate that our approach significantly outperforms prior approaches.
			Cheng Yan (National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab andSchool of Computer Science and Technology, Huazhong University of Science and Technology, China), Feng Zhao (National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab andSchool of Computer Science and Technology, Huazhong University of Science and Technology, China), Hai Jin (National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab andSchool of Computer Science and Technology, Huazhong University of Science and Technology, China)
		243	Improving Core Path Reasoning for the Weakly Supervised Knowledge Base Question Answering [Abstract] Abstract Core Path Reasoning (CPR) is an essential part of the knowledge base question answering (KBQA), which determines whether the answer can be found correctly and indicates the reasonableness of the path. The lack of effective supervision of the core path in weakly supervised KBQA faces great challenges in finding the correct answer through long core paths. Furthermore, even if the correct answer is found, its path might be spurious that is not semantically relevant to the question. In this paper, we focus on solving the CPR problem in weakly supervised KBQA. We introduce a CPR model that aligns questions and paths in a step-by-step reasoning manner from explicit text semantic matching and implicit knowledge bases structure matching. Additionally, we propose a two-stage learning strategy to alleviate the spurious path problem efficiently. We first find relatively correct paths and then use hard Expectation-Maximization to learn the best matching path iteratively. Extensive experiments on two popular KBQA datasets demonstrate the strong competitiveness of our model compared to previous state-of-the-art methods, especially in solving long path and spurious path problem.
			Nan Hu (School of Computer Science and Engineering, Southeast University, China), Sheng Bi (School of Computer Science and Engineering, Southeast University, China), Guilin Qi (School of Computer Science and Engineering, Southeast University, China), Meng Wang (School of Computer Science and Engineering, Southeast University, China), Yuncheng Hua (School of Computer Science and Engineering, Southeast University, China), and Shirong Shen (School of Computer Science and Engineering, Southeast University, China)
		360	Counterfactual-guided and Curiosity-driven Multi-Hop Reasoning over Knowledge Graph [Abstract] Abstract Recently, multi-hop reasoning over incomplete Knowledge Graphs (KGs) to predict missing facts has attracted widespread attention due to its desirable effectiveness and interpretability. It typically adopts the Reinforcement Learning (RL) framework and traverses over the KG to reach the target answer and find evidential paths. However, existing methods often give all reached paths equal hit rewards. Intuitively, not all paths have the same contribution to the proof of the reasoning process. Moreover, the severely sparse rewards obtained after a multi-step traversal are usually insufficient to encourage a sophisticated RL-based model to work well. In order to tackle the above two problems, we propose a novel Counterfactual-guided and Curiosity-driven Knowledge Graph multi-hop Reasoning model (CoCuKGR). CoCuKGR constructs counterfactual relation reasoning tasks to estimate the semantic contribution to the query relation of each path and give each arrival path a different soft reward that can distinguish its validity. In addition, our method leverages the curiosity mechanism to generate curiosity-driven intrinsic rewards, which can not only alleviate the reward sparsity issue but also drive the agent to explore the environment more thoroughly to find more abundant paths. Experimental results show that our proposed model outperforms existing multi-hop reasoning methods significantly.
			Dan Shi (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun, China and College of Computer Science and Technology, Jilin University, Changchun, China), Anchen Li (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun, China and College of Computer Science and Technology, Jilin University, Changchun, China), and Bo Yang (Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun, China and College of Computer Science and Technology, Jilin University, Changchun, China)
		388	Visualizable or Non-visualizable? Exploring the Visualizability of Concepts in Multi-modal Knowledge Graph [Abstract] Abstract An important task in image-based Multi-modal Knowledge Graph construction is grounding concepts to their corresponding images. However, existing research omits the intrinsic properties of different concepts. Specifically, there are some concepts that can not be characterized visually, such as mind, texture, session cookie and so on. In this work, we define concepts like these as non-visualizable concepts (NVC) and the others like dog that have clear and specific visual representations as visualizable concepts (VC). And, we propose a new task of distinguishing VCs from NVCs, which has rarely been tackled by the existing efforts. To address this problem, we propose a multi-modal classification model combining concept-related features from both texts and images. Due to the lack of enough training samples especially for NVC, we select concepts in ImageNet as the instances for VC, and also propose a webly-supervised method to get a small set of instances for NVC. Based on the small training set, we modify the basic two-step positive-unlabeled learning strategy to train the model. Extensive evaluations demonstrate that our model significantly outperforms a variety of baseline approaches.
			Xueyao Jiang (Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University, Shanghai, China), Ailisi Li (Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University, Shanghai, China), Jiaqing Liang (Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University, Shanghai, China), Bang Liu (Mila & DIRO, Universit´e de Montr´eal, Montr´eal, Qu´ebec, Canada), Rui Xie (Meituan, Shanghai, China), Wei Wu (Meituan, Shanghai, China),Zhixu Li (Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University, Shanghai, China), and Yanghua Xiao (Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University, Shanghai, China and Fudan-Aishu Cognitive Intelligence Joint Research Center, Shanghai, China)
20	Research Session #20 (Text and Image-IV), (April 14 (Thursday), 2022, 14.30-16.15, Hall 2)	90	Information Networks based Multi-semantic Data Embedding for Entity Resolution [Abstract] Abstract Entity resolution (ER) is an ongoing topic in data integration and data governance, which attracts considerable attention from multiple research fields. Recently, deep learning techniques have been substantially applied to entity resolution. We focus on entity resolution with graph based multi-semantic data embedding. In ER, data with attributes cannot be well represented by common word embeddings from natural language processing. In this work, data with attributes are modeled as a family of multitype bipartite information networks, each of which captures a specific type of semantics in data. Based on this, multi-semantic embeddings of data are collectively learned through the family of information networks. Particularly, a novel method is introduced to learn similarity based bipartite network embeddings. Generated tailored data embeddings are put into a flexible hierarchical ER framework, which outputs ER classification distributions. Our approach is comprehensively evaluated on a group of datasets, which presents its effectiveness.
			Chenchen Sun (School of Computer Science and Engineering, Tianjin University of Technology, Tianjin, China,Engineering Research Center of Learning-Based Intelligent System (Ministry of Education),Tianjin University of Technology, Tianjin, China and College of Intelligence and Computing, Tianjin University, Tianjin, China), Derong Shen (School of Computer Science and Engineering, Northeastern University, Shenyang, China) and Tiezheng Nie (School of Computer Science and Engineering, Northeastern University, Shenyang, China)
		269	Empowering Transformer with Hybrid Matching Knowledge for Entity Matching [Abstract] Abstract Transformers have achieved great success in many NLP tasks. The self-attention mechanism of Transformer learns powerful representation by conducting token-level pairwise interactions within the input sequence. In this paper, we propose a novel entity matching framework named GTA. GTA enhances Transformer for relational data representation by injecting additional hybrid matching knowledge. The hybrid matching knowledge is obtained via graph contrastive learning on a designed hybrid matching graph, in which the dual-level matching and multiple granularity interactions are modeled. In this way, GTA utilizes the prelearned knowledge of both hybrid matching and language modeling. This effectively empowers Transformer to understand the structural features of relational data when performing entity matching. Extensive experiments on open datasets show that GTA effectively enhances Transformer for relational data representation and outperforms state-of-the-art entity matching frameworks.
			Wenzhou Dou (School of Computer Science and Engineering, Northeastern University,Shenyang, China), Derong Shen (School of Computer Science and Engineering, Northeastern University, Shenyang, China), Tiezheng Nie (School of Computer Science and Engineering, Northeastern University,Shenyang, China), Yue Kou (School of Computer Science and Engineering, Northeastern University,Shenyang, China), Chenchen Sun (School of Computer Science and Engineering, Tianjin University of Technology, Tianjin, China),Hang Cui (University of Illinois at Urbana-Champaign, United States), and Ge Yu (School of Computer Science and Engineering, Northeastern University,Shenyang, China)
		445	Fake Restaurant Review Detection Using Deep Neural Networks with Hybrid Feature Fusion Method [Abstract] Abstract Online review websites all over the world have developed by leaps and bounds in the past decade, followed by more and more fake reviews that are very difficult to identify. This not only affects the decision-making of consumers, but also seriously undermines the fairness of the consumption market. Most existing research ignored the abundant metadata in reviews and its effective combination with the review text. In this paper, we firstly collect reviews from one of the most popular Chinese review platforms Dianping, and release two fake review datasets we construct, which makes up for the relative lack of datasets in this field. Afterward, we present some novel and effective features for fake review detection by analyzing the characteristics of the collected reviews. In addition, we propose a multimodal fake review detection model BAM (BERT + Attention + MLP), which takes the review text and the extracted features as input, and uses neural networks as well as multimodal fusion technology to realize the recognition of fake reviews. The exper- imental results show that BAM has low dependence on the size of the dataset. Compared with baseline models, it can be found that BAM has a better detection performance.
			Yifei Jian (School of Cyber Science and Engineering, Sichuan University, Chengdu, China), Xingshu Chen (School of Cyber Science and Engineering, Sichuan University, Chengdu, China), and Haizhou Wang (School of Cyber Science and Engineering, Sichuan University, Chengdu, China)
		179	Medical image fusion based on pixel-level nonlocal self-similarity prior and optimization [Abstract] Abstract “Self-similarity” is a common characteristic of medical im- ages. That is, small-scale features often appear in multiple locations in the image frequently. Therefore, the global search for similar pixels helps to infer the pixel value of a certain location, which can be used for the extraction of image details. In this paper, a two-stage image decomposition framework (TS-PLNSS) is proposed. It combines the pixel-level nonlocal self-similarity prior and pixel intensity attributes of the image. First, the source image is adaptively decomposed into three scales using different thresholds: texture layer, structure layer, and base layer. Then, different fusion strategies are adopted according to the characteristics of different layers. Among them, an optimization function that integrates structural information is designed to preserve the salient information of the source image to the maximum extent. Finally, the fused medical image is obtained through image reconstruction. The proposed method is compared with seven state-of-the-art multimodal medical image fusion methods to verify its effectiveness. The dataset consists of 110 image pairs, including CT/SPECT, MRI/PET, MRI/SPECT, MRI/CT, and GFP/PC. The subjective and objective evaluation results show that the TS-PLNSS method has better ability to distinguish and retain important information and texture details without distortion.
			Rui Zhu (Key Laboratory of Symbolic Computation and Knowledge Engineering, Ministry of Education, Jilin University, Changchun, 130012, China and College of Computer Science and Technology, Jilin University, Changchun, 130012, China), Xiongfei Li (Key Laboratory of Symbolic Computation and Knowledge Engineering, Ministry of Education, Jilin University, Changchun, 130012, China and College of Computer Science and Technology, Jilin University, Changchun, 130012, China), Yu Wang (Key Laboratory of Symbolic Computation and Knowledge Engineering, Ministry of Education, Jilin University, Changchun, 130012, China and College of Computer Science and Technology, Jilin University, Changchun, 130012, China), and Xiaoli Zhang (Key Laboratory of Symbolic Computation and Knowledge Engineering, Ministry of Education, Jilin University, Changchun, 130012, China and College of Computer Science and Technology, Jilin University, Changchun, 130012, China)
		187	Knowledge-enhanced Interactive Matching Network for Multi-turn Response Selection in Medical Dialogue Systems [Abstract] Abstract Recently, the response selection for retrieval-based dialogue systems has gained enormous attention from both academic and industrial communities. Although the previous methods achieve promising results for intelligent customer service systems and open-domain chatbots, the response selection in medical dialogues suffers from lower performance because of the strong dependency on the domain knowledge. In this paper, we construct two specialized medical knowledge bases and propose a Knowledge-enhanced Interactive Matching Network (KIMN) for multi-turn response selection in medical dialogue systems. Compared with previous response selection approaches, the KIMN adopts pre-trained language model to alleviate the limited training data problem, and incorporates internal and external medical domain knowledge to perform interactive matching between responses and contexts. The experiments on a real-world medical dialogue dataset show that our proposed model consistently outperforms the strong baseline methods by large margins, which shows that our proposed model can retrieve more accurate responses for medical dialogue systems.
			Ying Zhu (School of Computer Science and Engineering, Northeastern University, Shenyang, China), Shi Feng (School of Computer Science and Engineering, Northeastern University, Shenyang, China), Daling Wang (School of Computer Science and Engineering, Northeastern University, Shenyang, China), Yifei Zhang (School of Computer Science and Engineering, Northeastern University, Shenyang, China), Donghong Han (School of Computer Science and Engineering, Northeastern University, Shenyang, China)
		268	A Three-Stage Curriculum Learning Framework with Hierarchical Label Smoothing for Fine-Grained Entity Typing [Abstract] Abstract In this paper, we study the noisy labeling problem on the fine-grained entity typing (FET) task. Most existing methods propose to divide the training data into “clean” and “noisy” sets and use different strategies to deal with them during the training process. However, the “clean” samples used in these methods are not actually clean, some of them also contain noisy labels. To overcome this issue, we propose a three-stage curriculum learning framework with hierarchical label smoothing to train the FET model, which can use relatively clean data to train the model and prevent the model from overfitting to noisy labels. Experiments conducted on three widely used FET datasets show that our method achieves the new state-of-the-art performance. Our code is publicly available at https://github.com/xubodhu/NFETC-CLHLS.
			Bo Xu (School of Computer Science and Technology, Donghua University, Shanghai, China), Zhengqi Zhang (School of Computer Science and Technology, Donghua University, Shanghai, China), Chaofeng Sha (School of Computer Science, Fudan University, Shanghai, China and Shanghai Key Laboratory of Intelligence Processing, Shanghai, China), Ming Du (School of Computer Science and Technology, Donghua University, Shanghai, China), Hui Song (School of Computer Science and Technology, Donghua University, Shanghai, China), and Hongya Wang (School of Computer Science and Technology, Donghua University, Shanghai, China)
		395	TaskSum: Task-Driven Extractive Text Summarization for Long News Documents Based on Reinforcement Learning [Abstract] Abstract A popular and state-of-the-art family of extractive summa- rization is to explore pre-trained language models through reinforce- ment learning (RL). Despite gaining promising results, existing RL-based methods suffer from three drawbacks. First, they often adopt sparse reward signal schemes, which only give rewards to some of the extracted sentences, and result in neglecting salient sentences. Second, they often deem summarization as an independent task and neglect the latent connections existing between summarization and other downstream tasks, that could provide insightful hints to guide the upstream summarization task in return. Third, the length of input sequences in most summarization methods is restricted by the utilized pre-trained language models. To address these problems, we propose a novel RL-based Seq2Seq extractive summarization model, namely TaskSum, which combines extractive text summarization with multiple associated tasks via a dense reward signal scheme. Moreover, we implement a BERT-based hierarchical encoder to effectively encode documents of arbitrary length. Empirical results demonstrate that TaskSum can overcome the above-mentioned drawbacks of existing RL-based summarization methods and achieve significantly better results for long documents.
			Moming Tang (East China Normal University), Dawei Cheng (Tongji University), Cen Chen (East China Normal University), Yuqi Liang (Seek Data Inc), Yifeng Luo (East China Normal University), and Weining Qian (East China Normal University)
21	Research Session #21 (Recommendation-IV), (April 14 (Thursday), 2022, 14.30-16.15, Hall 3)	12	MDKE: Multi-level Disentangled Knowledge-based Embedding for Recommender Systems [Abstract] Abstract Recommender systems aim to dig out the potential interests of users and find out items that might be connected with target users. Accuracy of the recommendation list is crucial for user-oriented applications. Many knowledge-based approaches combine graph neural networks with exploring node structural similarity, while paying little attention to semantically distinguishing potential user interests and item attributes. Therefore, personalized node embeddings fail to be captured after simply aggregating neighborhood information. In this work, we propose Multi-level Disentangled Knowledge-based Embedding (MDKE) for item recommendation. We divide the embedding learning of users and items into disentangled semantic-level and structural-level subspaces. Specifically, semantic-level embeddings correspond to item content vector, which is learned by knowledge graph embedding, and user profile vector that is obtained through a carefully designed preference propagation mechanism. For structural level embeddings, a graph embedding-based module is proposed to capture local structure features on the redefined knowledge interaction graph. The latent vectors of users and items can be obtained by aggregating the aforementioned multi-level embeddings. Experiments on two real-world datasets verify the effec- tiveness of MDKE versus a series of state-of-the-art approaches.
			Haolin Zhou (MoE Key Lab of Artificial Intelligence,Department of Computer Science and Engineering,Shanghai Jiao Tong University, Shanghai, China), Qingmin Liu (MoE Key Lab of Artificial Intelligence,Department of Computer Science and Engineering, Shanghai Jiao Tong University, Shanghai, China), Xiaofeng Gao (MoE Key Lab of Artificial Intelligence,Department of Computer Science and Engineering,Shanghai Jiao Tong University, Shanghai, China), and Guihai Chen (MoE Key Lab of Artificial Intelligence, Department of Computer Science and Engineering,Shanghai Jiao Tong University, Shanghai, China)
		401	Intention Adaptive Graph Neural Network for Category-aware Session-based Recommendation [Abstract] Abstract Session-based recommendation (SBR) is proposed to recom- mend items within short sessions given that user profiles are invisible in various scenarios nowadays, such as e-commerce and short video recommendation. There is a common scenario that user specifies a tar- get category of items as a global filter, however previous SBR settings mainly consider the item sequence and overlook the rich target category information. Therefore, we define a new task called Category-aware Session-Based Recommendation (CSBR), focusing on the above scenario, in which the user-specified category can be efficiently utilized by the recommendation system. To address the challenges of the proposed task, we develop a novel method called Intention Adaptive Graph Neural Network (IAGNN), which takes advantage of relationship between items and their categories to achieve an accurate recommendation result. Specifically, we construct a category-aware graph with both item and category nodes to represent the complex transition information in the session. An intention-adaptive graph neural network on the category-aware graph is utilized to capture user intention by transferring the historical interaction information to the user-specified category domain. Extensive experiments on three real-world datasets are conducted to show our IAGNN outperforms the state-of-the-art baselines in the new task.
			Chuan Cui (Tongji University, Shanghai, China), Qi Shen (Tongji University, Shanghai, China), Shixuan Zhu (Tongji University, Shanghai, China), Yitong Pang (Tongji University, Shanghai, China), Yiming Zhang (Tongji University, Shanghai, China),Hanning Gao (Tongji University, Shanghai, China), and Zhihua Wei (Tongji University, Shanghai, China)
		430	Gated Hypergraph Neural Network for Scene-aware Recommendation [Abstract] Abstract To improve e-commercial recommender systems, researchers have never stopped exploring the interactions between users and items. Unfortunately, most existing methods only explore one or some certain components of the entire interactions. In fact, the entire interaction process is much richer and more complex, including but not limited to “who purchases what items in which merchant under what interaction environments”. Furthermore, many interactions have common features, thus forming a scene, a kind of prior knowledge for predicting user interactions. In this paper, we make the first attempt to study the scene-aware recommendation, which provides better recommendations with the entire interaction modeling and the scene prior knowledge. To this end, we propose a novel gated hypergraph neural network for Scene-aware Recommendation (SREC). Particularly, we first construct a heterogeneous scene hypergraph to model the entire interactions and scene prior knowledge. Then we propose a novel scene-aware gate mechanism-based hypergraph neural network to enrich their representations. Finally, we design a separable score function to predict the matching scores among user, scene, merchant and interaction environments for training and inference procedures. Extensive experiments demonstrate that our SREC can fully leverage the scene prior knowledge and outperforms state-of-the-art methods on real industrial datasets.
			Tianchi Yang (Beijing University of Posts and Telecommunications, China), Luhao Zhang (Meituan, China), Chuan Shi (Beijing University of Posts and Telecommunications, China), Cheng Yang (Beijing University of Posts and Telecommunications, China), Siyong Xu (Beijing University of Posts and Telecommunications, China), Ruiyu Fang (Meituan, China), Maodi Hu (Meituan, China), Huaijun Liu (Meituan, China), Tao Li (Meituan, China), Dong Wang (Meituan, China)
		13	Multi-Behavior Recommendation with Two-Level Graph Attentional Networks [Abstract] Abstract Multi-behavior recommendation learns accurate embeddings of users and items with multiple types of interactions. Although existing multi-behavior recommendation methods have been proven effective, the following two insights are often neglected. First, the semantic strength of different types of behaviors is ignored. Second, these methods only consider the static preferences of users and the static feature of items. These limitations motivate us to propose a novel recommendation model AMR (Attentional Multi-behavior Recommendation) in this paper, which captures hidden relations in user-item interaction network by constructing multi-relation graphs with different behavior types. Specifically, the node-level attention aims to learn the importance of neighbors under specific behavior, while the behavior-level attention is able to learn the semantic strength of different behaviors. In addition, we learn the dynamic feature of target users and target items by modeling the dependency relation between them. The results show that our model achieves great improvement for recommendation accuracy compared with other state-of-the-art recommendation methods.
			Yunhe Wei (College of Computer Science and Engineering, Northwest Normal University, Lanzhou Gansu 730070, China), Huifang Ma (College of Computer Science and Engineering, Northwest Normal University, Lanzhou Gansu 730070, China, Guangxi Key Lab of Multi-source Information Mining and Security, and Guangxi Normal University, Guilin Guangxi, 541004, China Guangxi Key Lab of Trusted Software, Guilin University of Electronic Technology, Guilin Guangxi, 541004, China), Yike Wang (College of Computer Science and Engineering, Northwest Normal University, Lanzhou Gansu 730070, China), Zhixin Li (Guangxi Key Lab of Multi-source Information Mining and Security, Guangxi Normal University, Guilin Guangxi, 541004, China), and Liang Chang (Guangxi Key Lab of Trusted Software, Guilin University of Electronic Technology, Guilin Guangxi, 541004, China)
		157	Toward Paper Recommendation by Jointly Exploiting Diversity and Dynamics in Heterogeneous Information Networks [Abstract] Abstract Current recommendation works mainly rely on the semantic information of meta-paths sampled from the heterogeneous information network (HIN). However, the diversity of meta-path sampling has not been well guaranteed. Moreover, changes in user’s reading preferences and paper’s audiences in the short term are often overshadowed by longterm fixed trends. In this paper, we propose a paper recommendation model, called COMRec, where the diversity and dynamics are jointly exploited in HIN. To enhance the semantic diversity of meta-path, we propose a novel in-out degree sampling method that can comprehensively capture the diverse semantic relationships between different types of entities. To incorporate the dynamic changes into the recommended results, we propose a compensation mechanism based on the Bi-directional Long Short-Term Memory Recurrent Neural Network (Bi-LSTM) to mine the dynamic trend. Extensive experiments results demonstrate that COM-Rec outperforms the representative baselines.
			Jie Wang (School of Computer Science and Technology, Soochoow University, Suzhou, China), Jinya Zhou (School of Computer Science and Technology, Soochoow University, Suzhou, China), Zhen Wu (School of Computer Science and Technology, Soochoow University, Suzhou, China), and Xigang Sun (School of Computer Science and Technology, Soochoow University, Suzhou, China)
		166	Enhancing Session-based Recommendation with Global Context Information and Knowledge Graph [Abstract] Abstract Predicting a user’s next click by utilizing a short anonymous behavior is a challenging problem in the real-life session-based recommendation(SBR). Most existing methods usually learn the users’ preference from current session. However, they seldom consider global context information or knowledge graph and failed to distill high-quality item from similar sessions. In this work, we combine Global Context information with Knowledge Graph, and develop a new framework to enhance session-based recommendation (GCKG). Technically, we model a global knowledge graph, exploiting a knowledge aware attention mechanism for better learning item embeddings. Then, we leverage an attention network and a gated recurrent unit to learn session representations. Furthermore, session representations are augmented simultaneously through construct- ing a similar session referral circle. Comprehensive experiments demonstrate that GCKG significantly outperforms the state-of-the-art methods of existing SBR.
			Xiaohui Zhang (College of Computer Science and Engineering, Northwest Normal University, Lanzhou Gansu 730070, China), Huifang Ma (College of Computer Science and Engineering, Northwest Normal University, Lanzhou Gansu 730070, China and Guangxi Key Lab of Multi-source Information Mining and Security, Guangxi Normal University, Guilin Guangxi 541004, China), Zihao Gao (College of Computer Science and Engineering, Northwest Normal University, Lanzhou Gansu 730070, China), Zhixin Li (Guangxi Key Lab of Multi-source Information Mining and Security, Guangxi Normal University, Guilin Guangxi 541004, China), and Liang Chang (Guangxi Key Laboratory of Trusted Software, Guilin University of Electronic Technology, Guilin Guangxi 541004, China)

6. Industry Sessions

Industry Session Id	Session Name	Paper ID	Paper Title
	(Date, Time, and Hall)
1	Industry Session # 1 Advancing Recommendation Systems, (April 12 (Tuesday), 2022, 10.15-12.30 , Hall 4)	553	A Joint Framework for Explainable Recommendation with Knowledge Reasoning and Graph Representation [Abstract] Abstract With the development of recommendation systems (RSs), researchers are no longer only satisfied with the recommendation results, but also put forward requirements for the recommendation reasons, which helps improve user experience and discover system defects. Recently, some methods develop knowledge graph reasoning via reinforcement learning for explainable recommendation. Different from traditional RSs, these methods generate corresponding paths reasoned from KG to achieve explicit explainability while providing recommended items. But they suffer from a limitation of the fixed representations that are pretrained on the KG, which leads to a gap between KG representation and explainable recommendation. To tackle this issue, we propose a joint framework for explainable recommendation with knowledge reasoning and graph representation. A sub-graph is constructed from the paths generated through knowledge reasoning and utilized to optimize the KG representations. In this way, knowledge reasoning and graph representation are optimized jointly and form a positive regulation system. Besides, due to more than one candidate in the step of knowledge reasoning, an attention mechanism is also employed to capture the preference. Extensive experiments are conducted on public real-world datasets to show the superior performance of the proposed method. Moreover, the results of the online A/B test on the large-scale Meituan Waimai (MTWM) KG consistently show our method brings benefits to the industry.
			Luhao Zhang (Meituan, China); Ruiyu Fang (Beijing University of Posts and Telecommunications, China); Tianchi Yang (Meituan, China); Maodi Hu (Meituan, China); Tao Li (Meituan, China); Chuan Shi (Beijing University of Posts and Telecommunications, China); Dong Wang (Meituan, China);
		568	XDM: Improving Sequential Deep Matching with Unclicked User Behaviors for Recommender System [Abstract] Abstract Deep learning-based sequential recommender systems have recently attracted increasing attention from both academia and industry. Most of industrial Embedding-Based Retrieval (EBR) systems for recommendation share the similar ideas with sequential recommenders. Among them, how to comprehensively capture sequential user interest is a fundamental problem. However, most existing sequential recommendation models take as input clicked or purchased behavior sequences from user-item interactions. This leads to incomprehensive user representation and sub-optimal model performance, since they ignore the complete user behavior exposure data, i.e., items impressed yet unclicked by users. In this work, we attempt to incorporate and model those unclicked item sequences using a new learning approach in order to explore better sequential recommendation technique. An efficient triplet metric learning algorithm is proposed to appropriately learn the representation of unclicked items. Our method can be simply integrated with existing sequential recommendation models by a confidence fusion network and further gain better user representation. The offline experimental results based on realworld E-commerce data demonstrate the effectiveness and verify the importance of unclicked items in sequential recommendation. Moreover we deploy our new model (named XDM) into EBR of recommender system at Taobao, outperforming the previous deployed generation SDM.
			Fuyu Lv (Alibaba Group, Hangzhou, China); Mengxue Li (Alibaba Group, Hangzhou, China); Tonglei Guo (Alibaba Group, Hangzhou, China); Changlong Yu (The Hong Kong University of Science and Technology, Hong Kong, China); Fei Sun (Alibaba Group, Hangzhou, China); Taiwei Jin (Alibaba Group, Hangzhou, China); Wilfred Ng (The Hong Kong University of Science and Technology, Hong Kong, China);
		584	Mitigating Popularity Bias in Recommendation via Counterfactual Inference [Abstract] Abstract Popularity bias is a common problem in recommender systems. Existing research mainly tracks this problem by re-weighting training samples or leveraging a small fraction of unbiased data. However, the effect of popularity bias in user behavior data may lead to sacrifices in recommendation. In this paper, we exploit data bias from click behavior to derive popularity bias representation, and investigate how to mitigate its negative impact from a causal perspective. Motivated by causal effects, we propose a novel counterfactual inference framework named Mitigating Popularity Bias in Recommendation via Counterfactual Inference (MPCI), which enables us to capture the popularity bias as the direct causal effect of the prediction score, and we eliminate popularity bias by subtracting the direct popularity bias effect from the total causal effect. In this way, MPCI reduces popularity bias by decreasing the influence of popular items on model training. Extensive experiments on two real-world datasets demonstrate the superiority of our methods over some strong baselines and prove the effectiveness of mitigating popularity bias in recommender systems.
			Ming He (Beijing University of Technology, Beijing, China); Changshu Li (Beijing University of Technology, Beijing, China); Xinlei Hu (Beijing University of Technology, Beijing, China); Xin Chen (Beijing University of Technology, Beijing, China); Jiwen Wang (Beijing University of Technology, Beijing, China);
		594	Efficient Dual-process Cognitive Recommender Balancing Accuracy and Diversity [Abstract] Abstract In this paper, we propose a dual-process cognitive recommendation system for sequential recommendations. The framework includes an intuitive representation module (System 1) and an inference module (System 2). System 1 is designed to understand the user’s historical interaction sequences with external knowledge graph. System 2 is built to make recommendations by reinforcement learning to consider long-term returns and diversity. We demonstrate the performance of our method on a wide range of recommendation datasets. Experiments show significant improvement over the state-of-the-art models regarding both relevance and diversity.
			Yixu Gao (Fudan University, Shanghai, China); Kun Shao (Huawei Noah’s Ark Lab, Beijing, China); Zhijian Duan (Peking University, Beijing, China); Zhongyu Wei (Fudan University, Shanghai, China); Dong Li (Huawei Noah’s Ark Lab, Beijing, China); Bin Wang (Huawei Noah’s Ark Lab, Beijing, China); Mengchen Zhao (Huawei Noah’s Ark Lab, Beijing, China); Jianye Hao (Huawei Noah’s Ark Lab, Beijing, China);
		629	Learning and Fusing Multiple User Interest Representations for Sequential Recommendation [Abstract] Abstract Deep learning is known to be effective at automating the generation of representations, which has achieved great success by learning efficient representations from data, especially for user and item representations in sequential recommendation. However, most existing methods usually represent a user’s interest by one independent representation vector, which is inadequate to capture user’s diverse interests. We aim to fully characterize the diversity of user interests in this study to improve the recommendation performance. We propose a Multiple User Interest Representations (MUIR) model that learns and fuses user’s interests from different aspects. To learn different levels of user interests, we specifically leverage two self-attention-based modules that better capture user’s local and global interests respectively. By considering the information of the recent interacted items, we further design a gating module to balance the interests, which is capable of modeling how user interests evolve and interact in recent behavior sequence. Extensive experiments conducted on the real-world datasets demonstrate that our model, MUIR, outperforms existing state of-the-art methods significantly.
			Ming He (Beijing University of Technology, Beijing, China); Tianshuo Han (Beijing University of Technology, Beijing, China); Tianyu Ding (Beijing University of Technology, Beijing, China);
2	Industry Session # 2 Data Management and Search, (April 13 (Wednesday), 2022, 10.15-12.30 , Hall 4)		Invited Talk / Panel: DocLens: Digiizing healthcare documents [Abstract] Abstract India generates on average more than a million hand-written prescriptions everyday, and more than a quarter million lab reports per day. Because of its non-digitized form, these documents are often lost under piles of other documents, or thrown away after their current use. These are thus rarely used for individual-level or population-level analysis. In this project, we aim to solve this problem, by converting hand-written prescriptions and lab reports to structured digital form, such that it can be used for analysis and help multiple entities in the medical ecosystem. Our work aligns with the National Digital Health Mission (NHDM) goal to build privacy-preserving personal health records and help people track their medical history. In this talk, we will primarily focus on information extraction from hand-written prescriptions. We will start with the overall problem at hand, and get into the details of one specific task - medicine name extraction from doctors' handwritten prescriptions. Due to the scarcity of intricately labeled data, we will discuss methods of information extraction from such data using limited supervision. Our work improves existing OCR technology for this use case by infusing knowledge from the medical domain. Gaurav Aggarwal (Google); Sujoy Paul (Google); Narayan Hegde (Google);
		627	Query-Document Topic Mismatch Detection [Abstract] Abstract Query-document topic match is one of the central signals for ranked information retrieval. Returning documents which are too broad or insufficient or off-topic compared to the query leads to a poor user experience. Thus, given a query and a document, it is critical to estimate degree of topical mismatch between the two. Previous work has either focused on very broad topics (like health, education, science, etc.) or predicted extremely-fine-level topic mismatch using query reformulations which suffer from sparsity. Predicting query-document topic mismatch is difficult because it needs semantic understanding of both the query and the document. In this paper, we model the problem as a five-class classification problem using a novel Transformer-based architecture. Our technique takes the query as input along with a detailed document representation including title, URL, snippet, key-phrases and topic distribution, and outputs one of these five grades of topic match: Very Unsatisfactory, Unsatisfactory, Neutral, Satisfactory, and Very Satisfactory. On a large dataset of ∼2.43M query-document pairs, we show that our proposed method can provide an AUC of 0.75.
			Sahil Chelaramani (Microsoft); Ankush Chatterjee (Microsoft); Sonam Damani (Microsoft); Kedhar Nath Narahari (Microsoft); Meghana Joshi (Microsoft); Manish Gupta (Microsoft); Puneet Agrawal (Microsoft);
		561	Beyond QA: ‘Heuristic QA’ Strategies in JIMI [Abstract] Abstract JD Instant Messaging intelligence (JIMI) is an intelligent service robot designed for creating an innovative online shopping experience in E-commerce. We will introduce a framework that combines the ‘intelligent prediction’ and ‘user click’ to facilitate the user input and to format the user inputs as standard queries, which we call as ‘heuristic QA’. It consists of three strategies: intent prediction before user querying, query auto-completion during user querying and next query prediction after user querying. Until now about 1/3 user queries are from heuristic QA in JIMI, which significantly improves the user experience.
			Shuangyong Song (JD AI Research, Beijing, China); Bo Zou (JD AI Research, Beijing, China); Jianghua Lin (JD AI Research, Beijing, China); Xiaoguang Yu (JD AI Research, Beijing, China); Xiaodong He (JD AI Research, Beijing, China);
		571	SQLG+: Efficient 𝑘-hop Query Processing on RDBMS [Abstract] Abstract Graph algorithms (e.g., k-hop queries) are widely used to find the deep association of data in various real-world applications such as business recommendation and fraud detection. However, most of the data are still stored in relational database (i.e., RDBMS) and the performance is rather limited when processing graph queries on RDBMS due to the inherent hardness of complicated table join. In this paper, we propose a fast interactive engine SQLG+, which can be integrated to any RDBMS and enable them to process k-hop graph queries efficiently. Different from naive table-join implementations, SQLG+ caches important nodes with their adjacency lists in memory (i.e., graph cache) and generates a hybrid query plan which combines the ability of graph cache and RDBMS. Also, SQLG+ removes duplicates in the end of each hop (using AdaptiveSet) and expands the frontiers in different ways. Furthermore, dynamic BFS/DFS switch is adopted to achieve the balance between query performance and memory occupation. Extensive experiments show that SQLG+ outperforms the state-of-the-art RDBMS-based implementations by up to several orders of magnitude and is even comparable to the fastest graph databases.
			Li Zeng (Huawei Technologies Co., Ltd, China); Jinhua Zhou (Huawei Technologies Co., Ltd, China); Shijun Qin (Huawei Technologies Co., Ltd, China); Haoran Cai (Huawei Technologies Co., Ltd, China); Rongqian Zhao (Huawei Technologies Co., Ltd, China); Xin Chen (Huawei Technologies Co., Ltd, China);
3	Industry Session # 3 Industrial Machine Learning Applications, (April 14 (Thursday), 2022, 10.15-12.30 , Hall 4)	570	Modeling Long-Range Travelling Times with Big Railway Data [Abstract] Abstract Big Railway Data, such as train movement logs and timetables, have become increasingly available. By analyzing these data, insights about train movement and delay can be extracted, allowing train operators to make smarter train management decisions. In this paper, we study the problem of performing long-range analysis on Big Railway Data, such as estimating the remaining journey time, i.e., the amount of time for a given train to reach the terminal station. We study how existing statistical and machine learning methods, designed for short-range analysis (e.g., estimating the traveling time between two adjacent stations), can be extended to perform long-range analysis. We further design a method, called a-LSTM, based on LSTM (long short-term memory) neural network and attention models. Extensive evaluation on a large amount of train movement data provided by a train service provider in Hong Kong shows that a-LSTM is more effective than other solutions in predicting traveling times.
			Wenya Sun (The University of Hong Kong); Tobias Grubenmann (University of Bonn, Germany); Reynold Cheng (The University of Hong Kong); Ben Kao (The University of Hong Kong); Waiki Ching (The University of Hong Kong);
		579	Multi-scale Time Based Stock Appreciation Ranking Prediction via Price Co-movement Discrimination [Abstract] Abstract The prediction of the stock market trends is an important problem and has attracted tremendous research interest. However, previous methods often consider modeling each stock separately and rarely leverage the information between different stocks to jointly train a model. In this paper, we address the problem of predicting the stock market trends and bring two key insights. First, we show that a better prediction model can be trained by simultaneously considering the features of correlated stocks. Unlike previous methods, our model does not rely on any prior manual input knowledge. Second, we observe that stock trend information on a single time scale is confined and not sufficient because the holding period can be different among investors. We thus design an encoder with multiple time scales to capture features for different time granularity. On top of these, we present a novel stock trend prediction framework called MPS. Extensive experiments are conducted on both the China A-Shares and NASDAQ markets, and results show that MPS outperforms baselines on different holding periods.
			Ruyao Xu (East China Normal University, Shanghai, China); Dawei Cheng (Tongji University, Shanghai, China); Cen Chen (East China Normal University, Shanghai, China); Siqiang Luo (Nanyang Technological University, Singapore); Yifeng Luo (East China Normal University, Shanghai, China); Weining Qian (East China Normal University, Shanghai, China);
		591	RShield: A Refined Shield for Complex Multi-Step Attack Detection based on Temporal Graph Network [Abstract] Abstract Complex multi-step attacks (i.e., CMA) have caused severe damage to core information infrastructures of many organizations. The graph-based methods are well known as the ability for learning complex interaction patterns of systems and users with discrete graph snapshots. However, such methods are challenged by the computer networking model characterized by a natural continuous-time dynamic graph. In this paper, we propose RShield, a temporal graph network-based CMA detection and defense method. It first constructs the continuous-time dynamic graph based on interactions among users and entities from various log records. Then it trains the detection model offline and performs streaming detection for live online network events. A prototype of RShield has been implemented. The experimental evaluation shows that RShield can achieve superior detection performance than the state-of-the-art methods in both transductive and inductive settings.
			Weiyong Yang (Nanjing University, Nanjing, China); Peng Gao (NARI Information & Communication Technology co., ltd, Nanjing, China); Hao Huang (Nanjing University, Nanjing, 210008, China); Xingshen Wei (NARI Information & Communication Technology co., ltd, Nanjing, China); Wei Liu (NARI Information & Communication Technology co., ltd, Nanjing, China); Shishun Zhu (NARI Information & Communication Technology co., ltd, Nanjing, China); Wang Luo (NARI Information & Communication Technology co., ltd, Nanjing, China);
		592	Inter-and-Intra Domain Attention Relational Inference for Rack Temperature Prediction in Data Center [Abstract] Abstract In a data center, predicting the rack temperature then generating alarms when an exception is detected can prevent server failure caused by high rack temperature. Each measuring point records the temperature of the rack over time, and each pair of measuring points may be associated with services or locations. Therefore, the rack temperature prediction problem can be modeled as a graph-based prediction problem. In this case, the prediction of the rack temperature depends not only on its own historical temperature but also on the temperature of racks having the same service or located near each other. Furthermore, the temperature of the rack is actually determined by various factors such as IT workloads and cold aisle temperature. Existing graph-based prediction methods do not consider the influence of these domains during the prediction, but only consider the temperature domain itself. To overcome this challenge, we propose an Inter-and-Intra domain Attention Relational Inference (I2A-RI) model: an unsupervised model that learns the relations between time series variables from different domains and utilizes the inferred interaction structure to achieve accurate dynamical predictions. Two attention modules, the guidance domain attention (GDA) module and the intra-domain attention (IDA) module, are proposed in I2A-RI, which encodes the inter-and-intra domain information to guide the learning procedure. Experiments on the real-world rack temperature dataset show that I2A-RI outperforms other state-ofthe-art models since it takes the advantage of the ability to infer the potential interactions across domains. The benefits of the two proposed attention modules are also verified in the experiments.
			Fang Shen (Alibaba Group, China); Zhan Li (Alibaba Group, China); Bing Pan (Alibaba Group, China); Ziwei Zhang (Tsinghua University, Beijing, China); Jialong Wang (Alibaba Group, China); Wendy Zhao (Alibaba Group, China); Xin Wang (Tsinghua University, Beijing, China); Wenwu Zhu (Tsinghua University, Beijing, China);

7. Demo and PhD Consortium Session

Date, Time, and Hall	Session Name	Paper ID	Paper Title
April 12 (Tuesday), 2022, 13.30-14.45, Hall 1	Demo presentations	598	An Interactive Data Imputation System; [Abstract] Abstract In this demonstration, we propose an interactive data imputation system, called DITS, built upon ten state-of-the-art imputation algorithms and a novel variational generative adversarial imputation network. It consists of three modules, namely source uploader, algorithm evaluation, and interactive imputation. In the source uploader module, DITS allows users to register new imputation and prediction algorithms. Then, DITS is able to make users more aware of various integrated imputation algorithms via algorithm evaluation module, so as to support both lay and power users to operate a customized data imputation for their target dataset via the interactive imputation module. Using a public incomplete meteorologic dataset, we demonstrate that, DITS is capable of assisting users to effectively address real-life missing data issues. Yangyang Wu (Zhejiang University); Xiaoye Miao (Zhejiang University)*; Yuchen Peng (Zhejiang University); Lu Chen (Zhejiang University); Yunjun Gao (Zhejiang University); Jianwei Yin (Zhejiang University).
		599	FoodChain: A Food Delivery Platform Based on Blockchain for Keeping Data Privacy; [Abstract] Abstract Marketplace systems for delivery food retain valuable sensitive data besides charging high fees for each order. On the other hand, Blockchain has being adopted as a distributed database to reduce costs and remove intermediary actors. We have developed FoodChain that is a blockchain based system, decentralizes the food delivery solution and keeps data in public ledgers by ensuring the privacy of sensitive data. Thus, customers, restaurants and couriers can verify and track their sensitive data that are stored in the blockchain. The interactions that occur directly between the involved actors are tracked by smart contracts, which in turn, may contain rules and conditions about the order delivery agreement. Moreover, all sensitive data is encrypted before being sent to the blockchain and can be decrypted using asymmetric encryption. The proposed demo presents the following FoodChain features: smart contracts running on Fantom testnet blockchain, an API to connect with a public database and three web-based decentralized applications (DApp). Rodrigo B Folha (Federal University of Pernambuco)*; Valéria Times (Federal University of Pernambuco); Arthur Carvalho (Miami University); André Araújo (Federal University of Alagoas); Henrique Couto (Federal University of Alagoas); Flaviano Viana (Federal University of Pernambuco).
		601	A scalable lightweight RDF knowledge retrieval system; [Abstract] Abstract Currently, there are numerous knowledge retrieval systems available to researchers, among which the RDF retrieval system is the most common. However, in practice, these systems are often plagued with problems, such as long index constructions and loading times and require large amounts of disk storage space, which are faults that make the system non-conducive to the dynamic incremental updates of data. This paper proposes a scalable lightweight RDF retrieval system, which has the following characteristics: 1) optimization of the index structure to reduce disk occupation and speed up the construction; 2) query optimizations based on query strategy selection; and 3) an interactive visual operation interface. The final evaluation results show that our system uses less disk space and has faster index construction times, and the query performance is competitive with the latest RDF retrieval systems. Yuming Lin (Guilin University of Electronic Techonlogy)*; Chuangxin Fang (Guilin University of Electronic Technology).
		603	CO-AutoML: An Optimizable Automated Machine Learning System; [Abstract] Abstract In recent years, many automated machine learning (AutoML) techniques are proposed for the automatic selection or design machine learning models. They bring great convenience to the use of machine learning techniques, but are difficult for users without programming experiences to use, and lack of effective optimization scheme to respond to users’ dissatisfaction with final results. To overcome these defects, we develop CO-AutoML, a user-friendly and optimizable AutoML system. COAutoML allows users to interact with the system in a customized mode. Besides, it can continuously optimize the search space of the AutoML technique based on reinforce policy and graph neural network (GNN), and thus provide users with more powerful machine learning schemes. Our system empowers ordinary users to easily and more effectively use AutoML techniques, which has a certain application value and practical significance. Our demonstration video: https://youtu.be/nGnmA7noeJA Chunnan Wang (HIT); Hongzhi Wang (Harbin Institute of Technology)*; Xu Bo (HIT); Xintong Song (Harbin Institute of Technology); Xiangyu Shi (Harbin Institute of Technology); Yuhao Bao (Harbin Institute of Technology).
		604	OIIKM: A System for Discovering Implied Knowledge from Spatial Datasets Using Ontology; [Abstract] Abstract Spatial association analysis is an important task in spatial data mining. The co-location pattern, an important expression form of spatial association analysis, has guided users' decisions in many aspects, such as city service, business, etc. However, current methods aim at mining co-location patterns consisting of fine-grained features, they ignore the background knowledge including known relationships of things in a certain domain. Moreover, current methods generate numerous and independent patterns, which causes users to be confused when making the following decisions. Unlike existing works, in this demonstration, we present a system called OIIKM (Ontology-based Interesting Implied Knowledge Miner), which employs ontology to integrate user knowledge during the mining process, to discover more knowledge (represented it as colocation patterns consisting of ontology concepts) implied in the spatial datasets. Besides, to alleviate user confusion, OIIKM provides a visual OntologyTree for the user to select ontology concepts she/he is interested in, by which only a few patterns are displayed to provide more guidance for better decisions. Long Wang (Guilin University Of Electronic Technology)*; Xuguang Bao (Guilin University of Electronic Technology); Liang Chang (Guilin University of Electronic Technology); Tianlong Gu (Guilin University Of Electronic Technology).
		608	IDMBS: An Interactive System to Find Interesting Co-location Patterns Using SVM; [Abstract] Abstract Spatial co-location pattern mining is an important task in spatial data mining. However, traditional mining frameworks cannot help a particular user effectively discover interesting co-location patterns according to his specific interest because traditional mining algorithms decide the prevalence (frequency) of a co-location pattern only by a user-specified real number. Thus, in order to discover the user’s real interesting co-location patterns, in this demonstration, we present IDMBS (Interactive data mining based on support vector machine), an interactive mining system, to discover user-preferred co-location patterns based on SVM. With IDMBS, users only need to go through a few rounds of interactions to efficiently discover the user-preferred patterns. IDMBML contains a filtering algorithm and an SVM model. The patterns selected by the filtering algorithm are annotated by the user, and the SVM model trains these patterns in order to discover more user-preferred co-location patterns. IDMBS can effectively and accurately discover the user-preferred patterns. Yuxiang Zhang (Guilin University of Electronic Technology)*; Xuguang Bao (Guilin University of Electronic Technology); Liang Chang (Guilin University of Electronic Technology); Tianlong Gu (Guilin University Of Electronic Technology).
		609	A Secure Trajectory Similarity Search System; [Abstract] Abstract To achieve the privacy protection on trajectory processing in outsourced environments, in this paper, we propose and demonstrate a Secure Trajectory Similarity Search System (SeTS3) to process the similarity search on encrypted trajectories. We design our system with four layers following the non-colluded dual-cloud model. Based on our previous work, three secure trajectory similarity search algorithms, which support different similarity metrics, are implemented. To assist the secure computations on encrypted trajectories, we develop a cryptographic toolkit in SeTS3 based on Paillier cryptosystem. To achieve a better demonstration, we apply map APIs to visualize the trajectories, and the search performance of the implemented algorithms can be displayed. Yiping Teng (Shenyang Aerospace University)*; Fanyou Zhao (Shenyang Aerospace University); Jiayv Liu (Shenyang Aerospace University); Mengfan Zhang (Shenyang Aerospace University); Jihang Duan (Shenyang Aerospace University); Zhan Shi (Shenyang Aerospace University).
		612	Data-based Insights for the Masses: Scaling Natural Language Querying to Middleware Data; [Abstract] Abstract In this demonstration, we focus on middleware data obtained from devices like the network routers and power meters which may be of interest to a technician fixing a customer complaint or a user trying to self-diagnose their utility usage. The users in our case are often unaware of both the data details and database querying language which is in contrast to typical natural language to structured query (NL2SQL) situations where the business analyst knows their domain data but not the querying techniques. We adapt the rule-based NL2SQL approach to our problem and in particular, focus on queries about users, devices and spatio-temporal properties that are unique to this setting. We demonstrate an Alexa-based system, implemented using open-source Rasa, that can answer router usage queries in a home setting and easily extended for power usage or other utilities. Kausik Lakkaraju (University of South Carolina)*; Vinamra Palaiya (Tantiv4); Sai Teja Paladi (University of South Carolina); Chinmayi Appajigowda (Tantiv4); Biplav Srivastava (AI Institute, University of South Carolina); Lokesh Johri (tantiv4).
		615	Identifying relevant sentences for travel blogs from Wikipedia articles; [Abstract] Abstract Travel blogs are a rich source of information about tourist spots. Millions of travellers use them to decide where to go and what to do. However, they are often devoid of factual information, which leads to a lack of trust and credibility surrounding blogs. On the other hand, Wikipedia is rich with accurate factual descriptions but is often bland and unfit for blogs. We propose a system to identify blog-worthy sentences on Wikipedia to augment travel blogs with factual information to make them more reliable and trustworthy. We curate a dataset of over 1.83 million Wikipedia sentences from 234,305 travel-related articles and assign a blog-worthiness score to each sentence by comparing them against 20 million sentences from ∼600K blogs. Our best BERT based model provides an F1 score of 0.74 to identify if a sentence is blog-worthy or not. Arnav Kapoor (International Institute of Information Technology, Hyderabad)*; Manish Gupta (Microsoft).
	PhD Consortium presentations	630	Neuro-Symbolic XAI for Computational Drug Repurposing; [Abstract] Abstract An emerging challenge in high stakes fields such as healthcare is to build a more transparent artificial intelligence. Among the fields which use AI techniques, is drug development, and more specifically drug repurposing. DR involves finding a new indication for an existing drug. The hypotheses generated by DR techniques must be validated. Therefore, the mechanism of generation must be understood. In this project, we focus on DR using link prediction algorithms. Link prediction consists of generating hypotheses about the relationships between a known molecule and a given target. More specifically, this PhD project aims at creating methods allowing to make transparent prediction in the context of drug repurposing but also to understand how the organization of data in a knowledge graph changes the quality of predictions. Martin Drancé Martin Drancé (University of Bordeaux)
		639	Leveraging Non-negative Matrix Factorization for Document Summarization; [Abstract] Abstract This paper outlines the doctoral research work carried out to develop unsupervised approaches for extractive single document summarization using Non-negative Matrix Factorization (NMF). NMF is a popular topic modeling technique, which divulges inter-relation between three prime semantic units of the document. We exploit the interrelationship among the three semantic units, viz. terms, sentences, and latent topics, in a novel manner to extract informative sentences from the salient topics in the document. The three methods developed during this doctoral study are language-, domain- and collection-independent. Alka Khurana (Department of Computer Science, University of Delhi).

Programme Schedule (Download the pdf)

Important Note: DASFAA-2022 Awards

Program Details

1. Keynote Speakers

Title:

Abstract:

Bio:

Title:

Abstract:

Bio:

Title:

Abstract:

Bio:

Title:

Abstract:

Bio:

Title:

Abstract:

Bio:

2. Panel Session

Overview:

Moderator:

Panelists:

(i) Georgia Koutrika (Athena Research, Greece)

(ii) Jaydeep Sen (IBM Research, India)

(iii) Immanuel Trummer (Cornell University, USA)

(iv) Lei Zou (Peking University, China)

3. Tutorial talks

Title:

Speakers:

Brief outline of the tutorial:

Speakers Bio

Title:

Speakers:

Brief outline of the tutorial:

Speakers Bio

Title:

Speakers:

Brief outline of the tutorial:

Speakers Bio

Title:

Speakers:

Brief outline of the tutorial:

Speakers Bio

Title:

Speakers:

Brief outline of the tutorial:

Speakers Bio

4. Workshops (April 11, 2022, Monday, 09.00- 18.00)

5. Research Sessions