Explaining Recommendations of Factorization-Based Collaborative Filtering Algorithms

Abstract

Recommender systems try to predict users' preferences on items, given their
feedback. Netflix, a DVD rental company in US, announced the Netflix Prize
competition in 2006. In that competition two powerful matrix factorization
(MF) algorithms were proposed, one using alternating least squares (ALS),
the other using gradient descent Both approaches aim to give a low-rank
representation of the matrix of ratings provirled by users on movies.
Recently, an algorithm was proposed to explain predictions of ALS-based
recommendations. This algorithm can explain why ALS "thinks" that a par-
ticular movie will suit the user's needs, where the explanation consists of
those movies previously rated by the user which are most relevant to the given
recommendation.
We propose a method that can explain predictions of gradient descent-based
MF in the analog way. We validate the proposed method by experimentation.