Binary Matrix Factorization applied to Netflix dataset analysis

Abstract

In this paper we aim at assessing the potential of Binary Matrix Factorization (BMF) in the implementation of recommendation systems, by analyzing a Netflix dataset. In particular, we study the explanatory power and the prediction capability of a particular BMF algorithm based on a post non-linear mixture model, namely the Post NonLinear Penalty Function (PNL-PF) algorithm. Unlike the majority of BMF methods, PNL-PF is capable of efficiently handling the difficult case of correlated rank-1 binary terms. We show that BMF represents an interesting alternative to classical matrix factorization methods in terms of explanatory power and prediction capability.