Research on Improved RBM Recommendation Algorithm Based on Gibbs Sampling

Abstract

Restricted Boltzmann Machine (RBM) is an important tool for personalized recommendation prediction, but it ignores the power-law distribution of the Restricted Boltzmann Machine data set, the RBM algorithm can not focus on the tail data sampling of the recommended data set. Therefore, firstly, the recommended data are obtained and the data characteristics are analyzed, then the random Gibbs Sampling initial value of RBM is changed to random selection in the early iteration and the last sampling value in the later iteration, the fixed Gibbs sampling steps were replaced by single-step sampling (CD-1) and multi-step sampling (CD-5),which is Periodic Gibbs Sampling (PGS). The experiment shows that the improved Gibbs sampling initial value and the changed Gibbs sampling steps can effectively improve the sampling performance, the improved RBM algorithm is also more accurate than the original RBM algorithm, the cyclic time Restricted Boltzmann Machine (RTRBM) algorithm and the Probability Matrix Factorization (PMF) algorithm. It shows that the improved RBM algorithm is suitable for the power-law distribution of recommendation data sets, and effectively improves the accuracy of recommendation.