Semantic Relation Transfer for Non-overlapped Cross-domain Recommendations

Abstract

Although cross-domain recommender systems (CDRSs) are promising approaches to solving the cold-start problem, most CDRSs require overlapped users, which significantly limits their applications. To remove the overlap limitation, researchers introduced domain adversarial learning and embedding attribution alignment to develop non-overlapped CDRSs. Existing non-overlapped CDRSs, however, have several drawbacks. They ignore the semantic relations between source and target items, leading to noisy knowledge transfer. Moreover, they learn knowledge from both domain-shared and domain-specific preferences and are hence easily misled by the source-domain-specific preferences. To overcome these drawbacks, we propose a novel semantic relation-based knowledge transfer framework (SRTrans). We semantically cluster the source and the target items and calculate their similarities to extract relational knowledge between domains. To transfer the relational knowledge, we develop a new two-tier graph transfer network. Last, we introduce a task-oriented knowledge distillation supervision and combine it with a prediction loss to alleviate the negative impact of the source-domain-specific preferences. Our experimental results on real-world datasets demonstrate that SRTrans significantly outperforms state-of-the-art models.