Friend Recommendation of Microblog in Classification Framework: Using Multiple Social Behavior Features Han Siyao Department of Information Science Beijing Language and Culture University Beijing, China hsy_blcu@163.com Xu Yan Department of Information Science Beijing Language and Culture University Institute of Computing Technology Chinese Academy of Sciences Beijing, China xuy@blcu.edu.cn Abstract—In recent years, microblog has been experiencing an explosive growth, which brings much inconvenience to users to build a healthy social circle in this chaos online world. Friend recommendation can automatically recommend potential friends, filter out the useless information, and facilitate the healthy development of social network. A novel friend recommendation approach is proposed in this paper. First, three kinds of social behavior features, i.e., social rating feature, social content features and social relation features, are extracted to represent the relationship of each user pair in the large-scale microblog data. Based on these features, a binary classifier is trained to determine whether the second user in each pair should be recommended to the first one. In this way, the original recommendation problem is transformed to a binary classification problem so that the sparseness problem of collaborative filtering method can be solved properly. Experiments shows that our approach improves the performance of friend recommendation compared with the traditional collaborative filtering methods. Keywords—Social Network; Friend Recommendatione; User Classification; Collaborative Filtering; Feature Extraction I. INTRODUCTION In recent years, Microblog is becoming one of the most popular social applications with increasing influence. The population of micro-bloggers has grown explosively. According to the statistical report on Internet development in China released by CNNIC at January 2014, microblog users’ coverage reached 55.4 percent in overall Internet users. The rapid development of microblog brings us a serious information overload problem, which brings much inconvenience for users accessing to valid information. Microblog recommendation system can filter out the useless information, and facilitate the continuous healthy development of social network. Generally, microblog recommendation system can be divided into tweet recommendation and friend recommendation. In this paper, we focus on the friend recommendation. Friend recommendation plays an important role in information sharing over microblog platforms. Generally, users would like to follow other users who post interesting messages and who have the similar interests with them. However, assisting users to find new people to follow is not a simple task. Microblog itself provides search services to help people find new users to follow by recommending popular users or the friends of their friends. However, these services do not offer the most relevant users to follow for one user. In our paper, Microblog friend recommendation system is trying to find the most relevant users by analyzing the tremendous amounts of social interaction data to extract useful features. We take advantage of the social behavior data, extract social content feature and social relation features of a user pair<   ,  >that are valuable for friend recommendation. Moreover, we also combine the traditional collaborative filtering algorithm, which provides a rating feature (which we called social rating feature). In microblog platform, explicit user-to-user ratings are not available. To solve this problem, we propose a novel friend ranking scheme using an adaptive CF algorithm. We frame recommendation problem as a classification problem, using the content, relation feature and CF rating feature for sample representation. Then we train a classifier to distinct the recommended users from non-recommended users. The contributions of this paper are as follows: ‚ We propose a sample concept—user pair<  ,  > for friend recommendation. ‚ We propose a novel feature extraction for user pairs, which is valuable and informative for representing the relationship between two users in microblog services. ‚ We transform the recommendation problem into a classification problem, which is innovative and efficient. By combining advanced classification model, it will achieve better performances. It also solve the sparseness problem of traditional recommendation methods. II. RELATED WORK The study of recommendation system was originated in movie recommendation system from MovieLens team in the 1990s. Then it is widely used in commercial fields to