Abstract-There is a common intuition that the user behavioral pattern and social information of a group may influence its attraction to users. In this paper, we employ user behavioral and social information to predict the user quit rate of social groups and validate the link between social behavioral pattern and group quit rate on a large scale real datasetTencent QQ groups. We routinely model this task as a regression problem, and generate 97 features from user behavioral and social information. Then we use an improved Scalable Orthogonal Regression (iSOR) method to predict the quit rate of QQ group. Our study shows that the quit rate of a group can be predicted with high accuracy, furthermore, the iSOR method selected several import features from the total 97 social behavioral features.Index Terms-Quit rate, social group, social information, user behavior.
I. INTRODUCTIONUnderstanding the social and user behavioral information of groups has been an attractive topic in social network analysis. Most of the existing works have focused on the social community detection (e.g.[1]- [3]), group profiling (e.g.[4], [5]) and group stability analysis (e.g. [6]-[10]) using social behavioral information. But little attention is paid to the link between social behavioral pattern and user quit rate of a group, which has great significance to both social science and the industry. For social science, it helps to understand the human collective behaviors. For industry, the social network operators can serve the customers more intelligently. In this paper, we extract 97 features for quit rate prediction and we explore the relation between social behavioral information and quit rate of a group.There exist various forms of groups in many online social network services [11]. In Tencent QQ groups, for example, user can find groups of her interest by searching keywords or group ID. She can join a group when permitted by the group manager or just quit a group she joined without any restriction. All these groups are created and maintained by QQ users. Users in a group can send messages that all group members can see. The continuous records of the user behavioral and social information in a group provide us the opportunity to study the relation between social behavioral information and quit rate of a group.In this paper, we extract lots of features from social behavioral information and predict the quit rate using an improved linear regression model, which could select significant orthogonal features automatically [12]. Our experiment result shows that the quit rate of a group can be predicted precisely using social behavioral pattern without knowing the conversation content, which validates the link between social behavioral information and group attraction. We also show that the behavioral pattern of group conversation is important to attract users staying in a group.The main contributions of this paper are: 1) Extract useful social features and user behavioral features for quit rate prediction. We totally extract 97 features from use...