Probabilistic Matrix Factorization Recommendation Based On Social Information And Item Exposure

In order to solve the problem of data overload,personalized recommendation service came into being.Collaborative filtering is the most widely used algorithm,which decomposes the user-item matrix(grading,clicks)into low-rank user potential preference matrix and item latent feature matrix using additional auxiliary information such as user social relations,contextual information and item content.In the recommendation of implicit feedback data,the unobserved user-item pair has ambiguity:one is that the user doesn't see the item,the partial data generally contains a lot of noise;the other is that the user has seen the item but doesn't like it,this part of data truly reflect user's potential preferences.In order to make better use of existing social information to improve implicit feedback recommendation performance,this paper proposes a probabilistic matrix factorization model based on users' social network and item exposure.On one hand,the model constrains users'potential preference factor by decomposing the user-user social matrix to alleviate data sparseness to some extent;On the other hand,the model takes item exposure as observation's condition,and models item exposure with item popularity and social information to better solve the unobserved ambiguity in data.Finally,we use expectation maximization algorithm to learn user latent preference matrix and item latent feature matrix.This paper conducts multiple levels of experimentation and analysis on the Lastfm public dataset.The results show that the introduction of user influence and social relationship into the item exposure and probabilistic matrix factorization model can avoid overfitting in the case of sparse data and improve the performance on metrics of recall,MAP and NDCG compared to state-of-art algorithms.