Abstract: Compared with the traditional collaborative recommendation algorithm, the Probabilistic Matrix Factorization (PMF) model performed well on large and sparse datasets. However, only the user’s rating information on the term was used in the model, the user and item char-acteristics were not fully considered. Therefore, there is still a lot of room for improvement in terms of recommendation accuracy. PMF Model Combining User Feature and Item Feature (UFIF-PMF) was proposed by combing user attribute features, user preference features and item tag features into PMF model. Firstly, the user attribute similarity according to the attribute information was calculated, and then the user preference similarity was calculated by use of the item label information and the user rating information, thus the user similarity matrix was constructed by weighting. Secondly, the item similarity matrix based on the item label information was constructed. Thirdly, the user similarity matrix and the item similarity matrix were incorporated into PMF model. Finally, the model training and comparison experiments on Movielens dataset were performed. The experimental results show that, compared with PMF, Matrix factorization recom-mendation algorithm based on user’s preference (USPMF) and PMF recommendation algorithm combing item similarity (ISPMF), the Root Mean Square Error (RMSE) of UFIF-PMF decreased by 6.27%, 3.65% and 3.49%, and Mean Absolute Error (MAE) decreased by 8.46%, 4.8% and 4.67% respectively. Furthermore, it effectively alleviates the problems of cold start and data sparsity, and has strong scalability.