Probabilistic Matrix Factorization Based On Co-occurrence And User Behavior

Among recommendation algorithm,the collaborative filtering algorithm based on matrix factorization is one of the most widely used recommendation techniques.This paper will expand traditional matrix factorization model according to co-occurrence and build co-occurrence relationship matrix based on point mutual relationship in word2Vec,then combine it with user behavior model and probability matrix decomposition,so as to study the impact of the co-occurrence purchase relationship matrix on the recommendation system.First of all,according to the relationship between items' co-occurrence,this paper regard items as point and construct point mutual information matrix,in the same way getting the user point mutual information matrix.Then we combine the historical score and the co-occurrence factor to decompose the two matrices with the rating matrix,after that we get the user-embedding vector and item-embedding vector.Next,we use the topic model to model the user's item set and comment text,and use these two topic vectors to liner fit user's vector.Finally we use probability matrix factorization to model the factors considered in this paper and get the final user and item embedding vector to predict users' purchase result.In this paper,we do experiment and analysis on the public dataset movielens and amazon.And results show that the probabilistic matrix factorization based on co-occurrence and user behavior is more effective compared with the classical recommendation algorithm such as PMF and CTR,and performs better on recommending item co-occurrence purchase set.