Semi Nonnegative Matrix Factorization Recommendation Algorithm Based On Bayesian Probability Model

Abstract

Despite model-based Collaborative Filtering (CF) methods are popular in recommendation systems, the latent factorization is extensively concerned. However, due to lack of nonnegative constraints and intrinsic probability semantics, traditional latent factorization methods have problems of low internal semantic interpretability and prediction accuracy with unsatisfactory running time per iteration times. In this paper, we propose a semi nonnegative matrix factorization algorithm based on Bayesian probability model. Using semi supervised learning method, the original rating matrix is initialized into positive and negative mixed which guarantees intrinsic interpretable semantics. Then, in order to add probability interpretation and improve the accuracy of prediction, we utilize the Gauss-Wishart distribution as a prior distribution and use Markov Chain Monte Carlo (MCMC) to achieve Bayesian random approximation inference. The whole experiment mainly considers the number of latent factors D and differences between best results and final results under a certain number of iterations. Facing the sparsity and imbalance in large data sets, the proposed algorithm guarantees the internal interpretability and has high prediction accuracy in the real world data set under the different number of latent factors and iterations with a small convergence time.