Cross-domain collaborative recommendation without overlapping entities based on domain adaptation

Recommender systems are the systems that take advantages of users’ historical behavior data to model the users’ behavior preferences to recommend things to users. However, recommender systems often suffer from data sparsity issues, due to a lack of adequate preference data, which degrades the overall recommendation performance. Cross-domain recommender systems were later developed to transfer knowledge from the auxiliary domain with rich user behavior data to help improve the recommendation performance of the target domain. Most of the existing cross-domain recommendation methods assume that overlapping entities are shared between domains, and then use them as a bridge for knowledge transfer across domains. However, this assumption does not universally hold. In this scenario, the existing cross-domain recommendation methods rarely consider the distribution inconsistency between domains, but directly transfer the cluster-level knowledge learned from the auxiliary domain to the target domain, which cannot ensure the consistency of knowledge transfer. Therefore, when overlapping entities are not shared between domains, how to effectively transfer knowledge is a key challenge for cross-domain recommender systems. Here, we propose a Cross-Domain Collaborative Recommendation without Overlapping Entities Based on Domain Adaptation, called CCR-DA. We find that CCR-DA can simultaneously achieve the consistency of knowledge transfer and avoid negative transfer in a unified framework. Specifically, we first seamlessly integrate the Maximum Mean Discrepancy (MMD) regularization constraints into the weighted collective matrix tri-factorization to reduce the distribution discrepancy between domains, so as to ensure the consistency of knowledge transfer. Then we further incorporate the graph regularization of user and item graphs from the two domains into the above framework to maintain the inherent geometric structure of each domain, thereby avoiding negative transfer. Experimental results on three categories of cross-domain recommendation tasks constructed from six real-world data sets show that our CCR-DA method outperforms.