Multiple Knowledge Transfer for Cross-Domain Recommendation

Abstract

Collaborative filtering based recommendation systems rely on underlying similarities among users and items across multiple dataset and hence requires sufficiently large amount of ratings data to achieve accurate and reliable results. However, newly established businesses do not have sufficient ratings data and hence this requirement is rarely met. In this research, we propose Multiple Latent Clusters (MultLC) transfer to exploit the correlations among multiple datasets that do not necessarily have an identical dimension of information. In particular, we transfer different aspects of knowledge across different data sources where while transferring each aspect from a source to the target, we only soft-transfer common latent clusters while preserving unique (domain-specific) latent clusters of the target. By soft-transfer, we mean that we minimize the difference among the shared clusters (while not making them identical). Comprehensive experiments on real-world datasets demonstrate the effectiveness of our proposed MultLC over other widely utilized cross-domain recommendation algorithms. The performance improvements demonstrate the benefits of transferring knowledge from multiple sources while preserving the unique information of the target-domain for cross-domain recommendations.