A Recommender System Based on a Double Feature Allocation Model

TL;DR

This paper introduces a Bayesian double feature allocation model for collaborative filtering recommender systems, utilizing Indian buffet processes and a novel semi-consensus Monte Carlo method to improve prediction accuracy and scalability.

Contribution

It presents a new Bayesian model for collaborative filtering that captures user-item features and introduces a semi-consensus Monte Carlo method for scalable inference.

Findings

Effective prediction of user preferences for unseen items.

Demonstrated scalability with large user and item datasets.

Improved inference accuracy over traditional methods.

Abstract

A collaborative filtering recommender system predicts user preferences by discovering common features among users and items. We implement such inference using a Bayesian double feature allocation model, that is, a model for random pairs of subsets. We use an Indian buffet process (IBP) to link users and items to features. Here a feature is a subset of users and a matching subset of items. By training feature-specific rating effects, we predict ratings. We use MovieLens Data to demonstrate posterior inference in the model and prediction of user preferences for unseen items compared to items they have previously rated. Part of the implementation is a novel semi-consensus Monte Carlo method to accomodate large numbers of users and items, as is typical for related applications. The proposed approach implements parallel posterior sampling in multiple shards of users while sharing item-related global parameters across shards.