Mixture Of User Behavior Modeling With Probabilistic Matrix Factorization For Recommendation

With the explosive growth of network information, people pay more and more attention to recommendation system. Collaborative filtering is the most successful technology and wildly used for recommendation. And probabilistic matrix factorization, as a key technology in collaborative filtering, is frequently searched in recent years. In this paper, we combine user behavior topic modeling with probabilistic matrix factorization, and introduce an algorithm named "Mixture of User Behavior Modeling with Probabilistic Matrix Factorization" to study the impact of user behavior on recommendation system.First, we divide the user behaviors into two categories, the items one user rates and the text information of a user. Then we apply topic modeling on these two behaviors, use linear regression to gain a user’s composite topic distribution. Then we use the composite topic distribution to restrain a user’s latent factor, and combine the probabilistic matrix factorization to get the final model. At last, we use the final user and item latent matrix to predict the users’ ratings on items.In this paper, we examine our algorithm on two public datasets from lastfm.com and delicious.com and analyze the result. We find that our algorithm is more effective compared with some state-of-art recommendation algorithms, and performs better on cold-start users more obviously.