Relationship Recommendation In Social Networks Based On Factorization Machines

In recent years, with the rapid development of social networks, a number of SNS services have attracted hundreds of millions of users’ usage, the vast amounts of information produced by the users has brought the problem of information overload, while it is difficult for users to find the potential friends, the interest topic or information among the huge number of users and information. The relationship recommendation in social networks aims at capturing users’interests and behavior tendencies from users’profile, behavior, social graph and keywords information, and accordingly serving them with potentially interesting users, driving users’ activities and enhancing their experience. In this paper, we mainly study the relationship recommendation problems in social networks based on Factorization Machines model. The major work of this paper is presented as follows:We summarize and review the concept and structure of recommender system, and also the commonly used algorithms and evaluation indicators. The classic collaborative filtering algorithms, recommendation algorithms in social networks and its applicable scenes are summarized and compared theoretically.We deeply describe the development of the factorization model in the recommender system and its core idea, and then the Factorization Machines model is introduced, we describe the expression of Factorization Machines, loss function and parameter solution and how it can be applied in recommender systems.Based on the Factorization Machines model, we extract some key features which have great influence on user behaviors from the Tencent Microblog dataset, and divide the sessions of users’behavior history, capturing the temporal information of the users’behaviors, which significantly improve the prediction accuracy. At the same time, since the rapid development of social networks which leads a large number of cold-start users in test dataset, we handle the cold-start problem by extracting some statistic features that we name as users’bias from the user’s social network. At last, we combine the users’bias with the Factorization Machines model and reach a prefect predication.Finally, we compare our solution with some traditional relationship recommendation algorithms, and the result has illustrated the effectiveness of our method. We also discuss the impact of the value of factorization dimension on the prediction accuracy.