Cross-domain recommender system using Generalized Canonical Correlation Analysis

TL;DR

This paper introduces a novel cross-domain recommender system leveraging generalized canonical correlation analysis to improve recommendations, especially for new users, by integrating auxiliary data and content features across multiple domains.

Contribution

It proposes an iterative MAX-VAR GCCA method and a new GCCA-ISSM approach to enhance cold-start recommendations using multi-domain user data and content features.

Findings

Improved accuracy in cross-domain rating predictions.

Enhanced cold-start recommendations with the proposed methods.

Validated effectiveness on Amazon and MovieLens datasets.

Abstract

Recommender systems provide personalized recommendations to the users from a large number of possible options in online stores. Matrix factorization is a well-known and accurate collaborative filtering approach for recommender system, which suffers from cold-start problem for new users and items. Whenever a new user participate with the system there is not enough interactions with the system, therefore there are not enough ratings in the user-item matrix to learn the matrix factorization model. Using auxiliary data such as users demographic, ratings and reviews in relevant domains, is an effective solution to reduce the new user problem. In this paper, we used data of users from other domains and build a common space to represent the latent factors of users from different domains. In this representation we proposed an iterative method which applied MAX-VAR generalized canonical correlation analysis (GCCA) on users latent factors learned from matrix factorization on each domain. Also, to improve the capability of GCCA to learn latent factors for new users, we propose generalized canonical correlation analysis by inverse sum of selection matrices (GCCA-ISSM) approach, which provides better recommendations in cold-start scenarios. The proposed approach is extended using content-based features from topic modeling extracted from users reviews. We demonstrate the accuracy and effectiveness of the proposed approaches on cross-domain ratings predictions using comprehensive experiments on Amazon and MovieLens datasets.