Recently, at the 27th Web Conference (The Web Conference 2018, formerly known as The International World Wide Web Conference - WWW), Chairman of the Conference Pierre-Antoine Champin announced that one of the three awards of the conference - - the unique Best Poster Award was presented to "An Improved Sampler for Bayesian Personalized Ranking by Leveraging View Data", a paper jointly completed by the team of Professor Jin Depeng and Associate Professor Li Yong of the Department of Electronic Engineering, Tsinghua University in cooperation with National University of Singapore. The first author of the paper is Ding Jingtao, a doctoral student of the Department of Electronic Engineering, Tsinghua University (supervisor: Professor Jin Depeng), andYuGuanghui, an undergraduate of the Department of Electronic Engineering also participated in it.
The opportunities and challenges of user behavior prediction in the era of big data coexist: the massive data have put forward new requirements for algorithm efficiency, and the complex user behavior contained in multi-source data requires more accurate and comprehensive modeling. The paper studies the sampler optimization of Bayesian personalized ranking algorithm (BPR for short) in user behavior prediction and solves the two major problems of ubiquitous negative sample sampling inefficiency and sparse user implicit feedback behavior. For the first problem, the paper puts forward the idea of reducing the negative sample candidate set for different users. It is found that when the set is reduced to 0.1%-1% of the full sample set, similar prediction performance can still be achieved, thus realizing the significant improvement of the sampler efficiency; for the second problem, the paper combines the rich characteristics of user behavior in multi-source data and effectively uses the auxiliary feedback data of the users to more accurately model user preferences. The work cooperates with the unicorn company in the domestic vertical e-commerce field of mothers and infants - - Beibei Group. The result of the online environment test on beibei.com shows that when the behavior prediction algorithm is applied to the product recommendations of millions of users, the conversion rate and exposure value and other indicators have been effectively improved by 12%-20%.
Meanwhile, the research group published a long article: "DeepMove: Predicting Human Mobility with Attentional Recurrent Networks" (first author: doctoral student Feng Jie, supervisor: Associate Professor Li Yong). The work introduces the attention mechanism into the human mobility behavior prediction field for the first time. While significant improving the accuracy of individual mobility prediction based on deep learning methods, it can give a regular explanation behind the success of the prediction, providing a promising research idea for realizing interpretable human behavior prediction.
The Web Conference/WWW was initiated by the International World Wide Web Conference Committee (IW3C2). As an authoritative Internet technology academic conference with great international reputation, it has been successfully held for 26 sessions, and is a top academic conference in the cross field of Internet systems and applications, and is also listed as a cross/comprehensive/emerging category A academic conference in theList of International Academic Conferences and Periodicals Recommended by China Computer Federation. Since 1994, experts and scholars from all over the world have gathered at the WWW conference, including leading scholars and industry elites from world-famous universities, research institutions, multinational corporations, and the International Organization for Standardization to jointly discuss cutting-edge Internet-related technologies and continue to promote the development and innovation of Internet technologies. The WWW conference in 1998 included Google's classic search engine papers.
