Research About Recommendation Algorithm Based On Probabilistic Matrix Factorization

Development of social network brings a new chance to recommender system. Recommending with social links can improve the accuracy of recommendations, and make the users trust why the system give them such suggestions. So more and more e-commerce websites and research workers want to make contributions to recommender systems. In this paper, we do some work on collaborative filtering recommender algorithms, which may face the problems of low accuracy and sparse social relations in real e-commerce websites. We propose new algorithms to solve those problems and do some experiments to prove the improvements of our algorithms, our analytic works in concrete are mainly as follows:1. To raise the accuracy of recommendation algorithm which is based on probabilistic matrix factorization, we propose a new algorithm which is called Topic-based Friends Refining Probabilistic Matrix Factorization(TFR-PMF). TFR-PMF can be used to make recommendations when users’social links are dense, it can solve the problem of considering too many friends’influence when recommending, which may decrease the accuracy of social recommendations. TFR-PMF treats users as sets of tastes(topics), and it tries to mine users’tastes by topic models first, and then it divides users’ social links by those tastes, after that, the final recommendations are only affected by those social relations which has the same tastes with user-item pairs. The experiments show that our algorithm can get a lower RMSE than other 4 PMF-based algorithms, which means our TFR-PMF model can improve the accuracy of social recommendations.2. To solve the problem of accuracy decreasing when social links are sparse, which are faced by TFR-PMF and other social recommendation algorithms, we propose a new algorithm called Topic-based User Clustering Probabilistic Matrix Factorization (TUC-PMF). This algorithm also tries to find users’tastes by topic model, and then it cluster all users by their common tastes. Later, it tries to build social links for users whose social relations are sparse from the results of clustering. So those users’social links become denser after building, it may get a better performance because social recommendations can get a lower RMSE when social links are dense. The experiments prove that our algorithm can improve the accuracy of recommendations for users whose social links are sparse.