Cross-domain Recommendation with Consistent Knowledge Transfer by Subspace Alignment

Recommender systems have drawn great attention from both academic area and practical websites. One challenging and common problem in many recommendation methods is data sparsity, due to the limited number of observed user interaction with the products/services. Cross-domain recommender systems are developed to tackle this problem through transferring knowledge from a source domain with relatively abundant data to the target domain with scarce data. Existing cross-domain recommendation methods assume that similar user groups have similar tastes on similar item groups but ignore the divergence between the source and target domains, resulting in decrease in accuracy. In this paper, we propose a cross-domain recommendation method transferring consistent group-level knowledge through aligning the source subspace with the target one. Through subspace alignment, the discrepancy caused by the domain-shift is reduced and the knowledge shared local top-n recommendation via refined item-user bi-clustering two domains is ensured to be consistent. Experiments are conducted on five real-world datasets in three categories: movies, books and music. The results for nine cross-domain recommendation tasks show that our proposed method has improved the accuracy compared with five benchmarks.