Research On Personalized Recommendation Algorithm Based-on Probabilistic Graphical Model And Parallel Implementation

With the rapid development of the Internet,many e-commerce platforms have gradually emerged and improved the quality of life of the public.However,with the explosive growth of data scale,Internet platforms that are using recommendation systems(such as Alibaba,Paper Weekly,etc.)are facing serious problems.The problem of information overload makes it impossible to make effective personalized recommendations based on user characteristics.Because the model expression ability of traditional recommendation algorithms is not strong enough,and the probability graph model has both the advantages of probability theory and graph theory,the probability graph model can be used to provide a more interpretable suggestion for the dependence between the variables in the recommendation problem.However,many recommendation algorithms based on probabilistic graph models only use the naive zero-mean spherical Gaussian prior distribution when solving parameters,and cannot achieve iterative parameter solving,so the recommendation effect for existing products(in-matrix)is poor;Many recommendation algorithms cannot solve the cold start problem,that is,new products that have just entered the system(out-of-matrix)will be difficult to recommend to any user because they do not have any purchase history;in addition,some recommendation algorithms incorporate socialization Information is used to alleviate the problem of excessively high sparseness of the score matrix,but the trust exposure factor among users is not considered;finally,the time complexity of the recommendation algorithm is often high,and the program operation efficiency is low.In order to solve the above-mentioned problems,this article mainly focuses on the exposure matrix factorization(Exposure Matrix Factorization,Expo MF)research,and proposes two improved algorithms.The main tasks are as follows:(1)A Variational Autoencoder-based Hybrid Recommendation(VAHR)is proposed.On the basis of the probability graph model of the Expo MF algorithm,Gibbs sampling is used to infer the parameters,so that the complete conditional distribution of a parameter obtained in the previous iteration is used as the prior distribution of the next iteration.The prior distribution and The conjugate relationship between the likelihood functions directly derives the analytical solution of the posterior distribution,thereby realizing iterative parameter inference;using the maximum posterior probability-expectation maximization algorithm to realize the iterative parameter estimation,and analyzing the two solutions The similarities and differences of the methods;the hidden features of the user's exposure vector are extracted and reconstructed by the variational autoencoder,so as to predict the exposure probability of each product to the user;use the parameters inferred above to train to obtain a variation that can extract the hidden features of the product Self-encoder to solve the recommendation problem of new products.Experiments show that compared with other algorithms of the same type,VAHR can effectively improve the recall rate for both in-matrix and out-of-matrix problems.(2)An Exposure-Based Social Recommendation(EBSR)algorithm is proposed.Based on the VAHR algorithm,the two types of social information,social tags and social trust relationships,are integrated to extract hidden features of users and products to alleviate the matrix sparsity problem;considering that the negative feedback value in the social trust relationship matrix may be affected by users Therefore,the user-user trust exposure implicit variable is further introduced on the basis of the user-commodity score exposure hidden variable;it is introduced in the generation process of the two observation variables of the user-product score and the user-user trust relationship.Correct the bias term to achieve more interpretable modeling;finally use Gibbs sampling and the maximum posterior probability-expectation maximization algorithm to solve the parameters,and analyze the similarities and differences between the two solution methods;taking into account the Gibbs in the EBSR algorithm The parallelism of Sampling uses the parallel framework Fork/Join provided by Java to optimize the efficiency of the algorithm.Experiments on Last.fm and the Zhihu data set obtained by crawling show that EBSR can effectively improve the score prediction accuracy compared with the same type of algorithm,and can achieve a higher speedup through parallel implementation.