Scalable Recommendation with Poisson Factorization

TL;DR

This paper introduces a scalable Bayesian Poisson matrix factorization model for recommendation systems that efficiently handles sparse, large-scale user-item data by focusing on observed entries, improving predictive accuracy.

Contribution

The paper presents a novel Poisson factorization model with a variational inference algorithm that scales to massive datasets, outperforming existing matrix factorization methods.

Findings

Outperforms state-of-the-art methods on real-world datasets

Efficiently handles large, sparse user-item matrices

Provides scalable inference for massive data sets

Abstract

We develop a Bayesian Poisson matrix factorization model for forming recommendations from sparse user behavior data. These data are large user/item matrices where each user has provided feedback on only a small subset of items, either explicitly (e.g., through star ratings) or implicitly (e.g., through views or purchases). In contrast to traditional matrix factorization approaches, Poisson factorization implicitly models each user's limited attention to consume items. Moreover, because of the mathematical form of the Poisson likelihood, the model needs only to explicitly consider the observed entries in the matrix, leading to both scalable computation and good predictive performance. We develop a variational inference algorithm for approximate posterior inference that scales up to massive data sets. This is an efficient algorithm that iterates over the observed entries and adjusts an approximate posterior over the user/item representations. We apply our method to large real-world user data containing users rating movies, users listening to songs, and users reading scientific papers. In all these settings, Bayesian Poisson factorization outperforms state-of-the-art matrix factorization methods.