Personalized Microblog Recommendation

With the advent of Web 2.0, social network has changed people’s habits of using Web pages significantly. Instead of showing users the content of Web pages passively, social network allows users who share interest, backgrounds or real-life connection to interact and collaborate with each other. Therefore, social network has attracted a great number of users and there is a sheer volume of data within a social network. These data are valuable sources for knowledge extraction and decision support.Recommender systems recommend an item(e.g., a product, a web page and a twitter etc.) which a user may be interested in to the user by studying his/her data. They have become very popular in the recent years and are applied in a wide variety of applications including e-commerce websites, enterprise recommendation engine and social networking services. Of all the algorithms of recommender systems, Collaborative filtering is one of the most important ones and it outperforms the traditional Text Filtering in some commercial applications.Currently, Weibo has become one of the most influential networking services throughout the world. Along with its increasing growth of popularity, the large number of information available on Weibo has obstructed people from accessing the information they are interested in. In this paper, we study the Weibo recommendation problem. We start by giving a brief review on existing recommendation algorithms, especially the Collaborative Filtering ones. Then we explore some explicit factors and implicit factors which may influence user’s interest and the preprocessing of data on Weibo. In the previous works, Chen et al. [1] only consider the relationship between user and publisher, and the relationship between weibo(tweet) and user. In our work, we propose to take the relationship between publisher and weibo into consideration and adopt Tensor Factorization to model relationships between user, weibo and publisher. Finally, we present our experimental results which show that our method significantly outperforms the baseline method.