Message-Passing Inference on a Factor Graph for Collaborative Filtering

TL;DR

This paper presents a new message-passing algorithm for collaborative filtering that improves performance in cold-start scenarios and is analytically tractable using density evolution techniques.

Contribution

The paper introduces the IMP message-passing algorithm for collaborative filtering, offering better performance with limited data and enabling analytical error bounds.

Findings

IMP outperforms EM and other matrix completion methods with small data.

The approach improves cold-start recommendation quality.

Density evolution analysis is applicable to the MP algorithm.

Abstract

This paper introduces a novel message-passing (MP) framework for the collaborative filtering (CF) problem associated with recommender systems. We model the movie-rating prediction problem popularized by the Netflix Prize, using a probabilistic factor graph model and study the model by deriving generalization error bounds in terms of the training error. Based on the model, we develop a new MP algorithm, termed IMP, for learning the model. To show superiority of the IMP algorithm, we compare it with the closely related expectation-maximization (EM) based algorithm and a number of other matrix completion algorithms. Our simulation results on Netflix data show that, while the methods perform similarly with large amounts of data, the IMP algorithm is superior for small amounts of data. This improves the cold-start problem of the CF systems in practice. Another advantage of the IMP algorithm is that it can be analyzed using the technique of density evolution (DE) that was originally developed for MP decoding of error-correcting codes.