Non-overlapping Cross-domain Recommendation via Adapted Tensor Factorization

As a promising strategy dealing with data sparsity issue, cross-domain recommender systems transfer valuable knowledge from auxiliary domains to assist the recommendation task in a target domain. Most existing research focuses on the scenario that the auxiliary domains share the same users or items with the target domain. However, such auxiliary data is only obtainable in quite few circumstances in the real-world. In this paper, we study the general scenario that the auxiliary domains have no overlapped users or items with the target domain, which can be applied widely owing to the easily acquired auxiliary data. We consider a triadic user-item-domain interaction pattern to synthetically model user, item and domain factors in a common space by an adapted tensor factorization model, which can further be effectively solved with closed-form updating rules. The experimental results have shown the effectiveness of our method on data sets publicly available from three domains