Study On Personalized Recommendation Algorithm Based Probabilistic Matrix Factorization

Using the known score to predict the unknown score and gain recommendation is the basic task of the personalized recommendation system,in which probability matrix factorization(PMF)model is a typical representative,but there are still some existing bottleneck problems,such as high dimensional data sparsity.To further study PMF model,the main contributions are summarized as follows:(1)Based on the research of the status about collaborative filtering,PMF and trust mechanism at home and abroad,the basic principles of PMF technology and the trust mechanism based on social network are deeply studied and analyzed.According to the demands of items and the data analysis,the HBase storage system and the Spark cluster which are used to deal with the big data are built successfully.(2)Owing to that PMF model is sensitive to the initial value,an improved algorithm based on alternating least square is proposed integrating the user-item bias information.The item latent factor is taken as the initial value of the alternating least squares to improve the recommendation accuracy.The experimental results show the proposed algorithm in thesis significantly improves recommendation accuracy.(3)Based on the user-item bias information,an improved algorithm of social network is proposed using the trust propagation mechanism,which combines the trust with the scores to evaluate the preference relations among users,and then we use gradient descent to obtain the optimal solution.The experimental results verify the validity of the improved algorithm.The improved PMF model proposed in thesis is applied to the actual project,and the prototype system is designed and implemented to meet the expected requirements of items in the personalized recommendation.