Research On Recommendation Method Based On Bayesian Local Probabilistic Matrix Factorization

The recommendation system is a product of the development of The Times.It is an efficient and intelligent information filtering platform,which can be recommended to users by users who may be interested in it.Collaborative filtering is the most successful,most widely used technology in recommendation system,and the Probability Matrix Factorization Model is a powerful branch of the collaborative filtering recommendation algorithm,especially in the face of mass data,data sparseness.This paper summarizes the current knowledge of the common recommendation algorithm and the basic content.On the basis of the further study of probability matrix factorization algorithm,spectral clustering method is introduced to users of social information mining.At the same time,we use the Bayesian method to improve the Probabilistic Matrix Factorization Model so that the model no longer requires manual tuning.Another innovation is that we use the Gibbs sampling algorithm to sample the model to further improve the accuracy of the model.The main contents and innovation of this paper are as follows.1.In-depth study of the probability matrix decomposition model,this paper introduces the principle of the algorithm in detail,including the matrix decomposition and introduction of probability distribution of the relevant knowledge and applied to real data set,the advantages of the data sparse problem are verified.2.The spectral clustering algorithm is introduced,before matrix factorization,user similarity by calculation with the method of clustering dividing the original score matrix into matrix,which is BLPMF model is proposed in this paper can effectively reduce the root cause of the problem of "cold start" influence.3.The traditional probabilistic matrix decomposition algorithm is optimized by Bayesian method.System parameters as obey Gaussian-Wishart distribution,to estimate the parameters of the distribution parameters(hyper-parameter)rather than like PMF algorithm,the system parameters as a fixed value estimate,need to adjust the parameters with a careful hand,otherwise easy to appear over fitting phenomenon.4.Using the Gibbs sampling algorithm in MCMC,the hyper-parameter,the feature vector(such as the user and the item)is sampled,and the optimal hyper-parameter is automatically selected.5.Through the Bayesian Local Probability Matrix Factorization algorithm for case study,we study on application examples in APP activity data sets,verify the validity of the BLPMF algorithm,and the possible affect effect of related factors were analyzed,and verified the correctness and validity of the algorithm.