Behavior analysis of RBM for estimating latent factor vectors from rating matrix
Hiroki Shibata, Yasufumi Takama · 2017
This paper analyzes the behavior of Restricted Boltzmann Machine (RBM) when it is applied to estimate latent factor vectors from rating matrix. Recently some RBM models are applied to predict ratings for recommendation systems. However, some papers also pointed out that RBM-based recommender systems could not achieve a clear improvement compared with conventional methods like Singular Value Decomposition (SVD) despite using more advanced and higher-cost statistical model. While the reason of low performance should be revealed by analyzing RBM's behavior, detailed analysis has not yet been investigated. This paper proposes alternative implementation of RBM-based collaborative filtering and analyzes its behavior. Experimental results with artificial datasets shows that it retrieves completely original latent vectors in some cases.