A Personalized Weibo Message Recommendation Algorithm Based On Topical Social Influence

In recent years, an increasing number of people become used to get in time news through microblogging services such as weibo. Users send out hundreds of millions of messages on weibo, which brings along the severe challenge of information overload. Therefore it is of great significance to provide precise weibo recommendations to users by learning how to help the user find the weibo they may be interested in among tons of information.However, nearly all the current algorithms for weibo recommendation ignore the influence of user's interest distribution and topical social influence from the authors of the corresponding messages. In fact people will follow different users for different topics, and they receive different degrees of influence from different topics. If we conduct a deep dig into the user's interest topics as well as social influence, and make effective use of these information to recommend weibo messages, we will be able to elevate the accuracy rate of recommendation, alleviate the cold-start problem and improve the user experience. So far there is no integration method to take both topic interest distribution and users' topical social influence into consideration.In this paper a personalized weibo message recommendation algorithm is proposed based on topical social influence. It consists of three parts:first, using topics obtained by LDA to calculate each user's interest topic distribution; second, establishing the user social influence estimation model based on topic factor graph by mining the intrinsic relationships among users'interest topic distribution, following relations and user's repost relations. The user social influence estimation model can give quantitative estimates of users social influence within different topics; last, integrating the social influence into the traditional latent factor model for recommendation. By predicting user's latent rating for each weibo message and ranking accordingly (a higher rating means a higher probability of repost), we create a personalized recommendation list for each user. Experiments have been done using Sina weibo, from which the results show the recommendation method based on topic social influence significantly outperforms the traditional weibo message recommendation method.