Predicting User Activity Level in Social Networks Yin Zhu † , Erheng Zhong † , Sinno Jialin Pan ‡ , Xiao Wang ⋆ , Minzhe Zhou ⋆ , Qiang Yang †⋄ † Hong Kong University of Science and Technology, Hong Kong, China ‡ Institute for Infocomm Research, Singapore ⋆ Renren Inc., Beijing, China ⋄ Huawei Noah’s Ark Lab, Hong Kong, China {yinz,ezhong,qyang}@cse.ust.hk, jspan@i2r.a-star.edu.sg {xiao.wang,minzhe.zhou}@renren-inc.com ABSTRACT Social media such as Facebook, Renren and Twitter provide an ide- al ground to study how to predict users’ future activities based on their past social behavior. An important measure of the behavior is activity level, such as users’ level of weekly activeness, or bi- nary classifications in terms of active or inactive. This prediction problem is closely related to Social Customer Relationship Man- agement (Social CRM). Compared to traditional CRM, social CR- M exhibit some special characteristics, in terms of user diversity, social influence, and dynamic nature of social networks. Users’ so- cial diversity property implies that a global predictive model may not be precise for all users. However, the historical data of in- dividual users are too sparse to enable high-quality personalized models. The social influence property suggests that relationships between users can be embedded to further boost the prediction re- sults on individual users. Finally, the dynamical nature of social networks means that users’ behaviors change over time. To address these challenges, we develop a personalized and socially regular- ized time-decay model for accurate user activity level prediction. We conduct experiments on the social media Renren to validate the effectiveness of our proposed model to demonstrate the superior performance when compared with traditional supervised learning methods as well as node classification methods in social networks. Categories and Subject Descriptors H.2.8 [Database Applications]: [Data mining] Keywords Social Network Analysis, User Activity, Prediction 1. INTRODUCTION The number of active users in a social network is a critical mea- sure of its popularity, which can be used as a signal of investment value for investors. In many social network companies’ quarterly Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full cita- tion on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re- publish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. CIKM’13, Oct. 27–Nov. 1, 2013, San Francisco, CA, USA. Copyright 2013 ACM 978-1-4503-2263-8/13/10 ...$15.00. 1 5 10 15 20 25 0 1 2 3 4 5 6 7 Week Weekly Active Days Alice Bob Candy 18 Apr 2012 9 Oct 2012 Figure 1: The weekly active days of three Renren users over 25 weeks. reports, such as that of Facebook (FB) 1 and Renren (RENN) 2 , the number of Monthly Active Users (MAUs) and other user active- ness measures are published. Because these numbers are the strong indicators of popularity and investment value, social network com- panies adopt various strategies to attract new users and maintain old users, e.g., building faster and more stable services, providing better recommendations, developing innovative UI, and support- ing personalized services. An important strategy for increasing the number of active users is to give incentives to users who are inactive or going to be inactive. But if a user is already inactive for a long time (i.e., lost users), it is much harder to activate the user again than when he/she only shows a sign of becoming inactive. This fact motivates us to explore how to predict a user’s future active- ness either in terms of the level of activity or binary classification into active or inactive. By accurately predicting the future activ- ity levels of users, we can track potentially lost users in an early stage and give them incentives to stay active. The prediction mod- el can also shed lights on the explanation of what user behaviors show correlations with their future activity levels. These insights can help improve user-maintenance strategies. Figure 1 shows the weekly online days of three users over 25 weeks from Apr. 18, 2012 to Oct. 9, 2012 in Renren (the detail- s of the dataset are described in the experiment section). Among the three users, Alice and Bob are close friends and exhibit similar patterns of weekly online days. Candy becomes relatively inactive after the 11’th week, and has no more online actions since the 18’th week. If we were able to identify a user like Candy who was still active by some time point but was about to decline his/her online 1 Reports are available at http://investor.fb.com 2 Reports are available at http://ir.renren-inc.com