Zero-Inflated Poisson Factorization for Recommendation Systems

A recommendation system confronts two opposing problems. In order to be practical, recommendations systems need to be quick, efficient, and scale well in both computational complexity and memory cost [7, 23]. This is frequently achieved by reducing the dimensionality of the data [7]. For example, suppose we have N users, M items, and a review matrix R ∈ RN×M consisting of positive integer ratings. Existing recommendation systems will try to find sparse or low rank approximations to R, capturing an intuition that high dimensional rating observations are governed by relatively low dimensional preferences [23, 25, 16, 9].