For the first time in China, Tsinghua and other teams won the wsdm2022 only best paper award, and Hong Kong Chinese won the "time test Award"

author | West

edit | Chen Caixian

2 month 21 solstice 25 Japan , The first 15 The international Internet search and Data Mining Conference （WSDM 2022） Hold online , The research team from the computer department of Tsinghua University won the only best paper award at the conference ！

This is also since the founding of the conference , The scientific research team from China won the award for the first time .

WSDM（ The pronunciation is 「Wisdom」） By the international computer society （ACM） Its information retrieval （SIGIR）、 data mining （SIGKDD）、 database （SIGMOD） And network information processing （SIGWEB） And so on , Enjoy a high academic reputation in the field of data mining .

Besides , Besides the best Thesis Award ,WSDM The conference also announced 「 Time test Award 」 Award winning work of —— The Chinese University of Hong Kong team “Recommender systems with social regularization”（WSDM 2011）.

 

 

1

WSDM Best Paper Award

According to the official website of the conference , Tsinghua University won this year WSDM The only work for which the best Thesis Award is awarded is “Learning Discrete Representations via Constrained Clustering for Effective and Efficient Dense Retrieval”（ Discrete representation learning based on constrained clustering improves the performance of dense vector retrieval ）.

Thesis link ：https://arxiv.org/pdf/2110.05789.pdf

The author of the paper is ： Zhan Jingtao , Mao Jiaxin , Liu Yiqun , Guo Jiafeng , Zhang min , Ma Shaoping . The first author is zhanjingtao, a doctoral student in the computer department of Tsinghua University , The corresponding author is Professor liuyiqun, Department of computer science, Tsinghua University , The relevant achievements were made by Tsinghua University 、 Renmin University of China 、 The Institute of computing, Chinese Academy of Sciences and other units jointly completed .

Figure note ： Professor Liu Yiqun

With the extensive application of deep learning and pre training language model , Dense vector retrieval has become one of the most important and frequent data operations in Internet search , However, the existing dense vector retrieval model greatly increases the storage overhead and time complexity compared with the traditional index retrieval model , This creates an important bottleneck for performance improvement .

For the above problems , In this paper, we propose a new clustering algorithm based on constrained clustering （Constrained Clustering） A retrieval model for improving dense vector retrieval process RepCONC.

Figure note ： The training flow chart of retrieval model proposed in this paper

The model is based on constrained clustering method to jointly optimize the process of text encoder and vector quantization end-to-end ,RepCONC Constrained dense vectors are uniformly distributed to different quantization centers , Thus, the discernibility of dense vector representation is greatly improved , Improved retrieval performance .

This paper proves the importance of this constraint in theory , The approximate solution of the constrained clustering process is derived by using the optimal transmission theory to improve the efficiency of the algorithm .RepCONC Vector inversion file system that can be used in the industry （IVF） Up operation , Even if you leave GPU Use only CPU It can also achieve better index compression and retrieval results , Compared with the traditional dense vector retrieval method in compression ratio 、 Retrieval performance 、 Time efficiency and other aspects have been significantly improved .

Figure note ： Schematic diagram of constrained clustering process

In addition to the only best paper selected every year , The Congress also elected 3 Best paper nominations （Best Paper Award Runner-Ups）：

• Doubly Robust Off-Policy Evaluation for Ranking Policies under the Cascade Behavior Model（ Tokyo institute of technology ）
• Evaluating Mixed-initiative Conversational Search Systems via User Simulation（ Ticino University ）
• The Datasets Dilemma: How Much Do We Really Know About Recommendation Datasets?（ Nanyang University of technology ）

 

 

2

Time test Award

get WSDM 2022「 Time test Award 」 My job is from the Chinese University of Hong Kong “Recommender Systems with Social Regularization”.

The reason for awarding awards given by the conference is ：

Recommendation system has become an enduring research topic in academia and industry . The committee selected the paper , Because of its importance and influence in the field . This paper explores the relationship between trust and recommendation , Realize that users don't necessarily have similar tastes to the people they trust , But it also affirms the importance of trust to recommendation . The author establishes the most appropriate social connections for several different recommendation tasks , Thus helping to establish the value of incorporating social signals into the recommendation system . therefore , This paper has not only produced a strong influence （ stay  WSDM The time test award has been cited the most times among all the nominations ）, It also foresees the importance of trust and transparency in the recommendation system , It has become an important topic in recent days .

The paper was published in 2011 year WSDM 2011 receive .

Address of thesis ：https://dennyzhou.github.io/papers/RSR.pdf

In this work , The research team of the computer department of the Chinese University of Hong Kong made a pioneering study of the few people who were interested in 、 It's hot now 「 Social recommendation 」 problem . at present , Social recommendation has become a necessary skill for various Internet products , Microblogging 、 Tiktok 、 TaoBao 、 WeChat 「 Have a look 」 And so on .

They are based on the user's social friend information （ Mining data from platforms such as Douban ）, Two social recommendation algorithms are proposed , The objective function is decomposed by the constraint matrix of social regularization term , To help improve the prediction accuracy of the recommendation system . Experimental results show that , Their method is very general , It is suitable for solving many types of trust aware recommendation problems .

More Than This , The paper will also think in reverse , To realize ： The existence of social relationships may reduce the quality of recommendations . Starting from a single trust may also produce recommendations with low accuracy , For example, friends who are good at studying sneakers may not be good at movie recommendation . therefore , The authors used the similarity function very early , An algorithm for identifying target friends based on different recommendation tasks is designed , To model social systems more realistically .

In the paper , They put forward ： They believe , With the rapid development of online social networking sites , Social based research will become more and more popular . The fact proved that , Such is the case .

Reference link ：

1.https://www.wsdm-conference.org/2022/

2.https://mp.weixin.qq.com/s/FLtWupAxoqAthXYQOa3YrQ?v_p=89&WBAPIAnalysisOriUICodes=10000001&launchid=10000365--x&wm=3333_2001&aid=01A3NNUgUONWBBii_bsE-e7BIuQxBjrg6ihZ9c4RkfEruJoPI.&from=10C2093010