HKGCL: Hierarchical graph contrastive learning for multi-domain recommendation over knowledge graph.

• A hierarchical contrastive approach over knowledge graphs for MDRs is proposed. • A domain-shared contrastive learning task across multiple domains is devised. • Two domain-specific contrastive learning tasks in separate domains are designed. • The experiments have exhibited the superiority of our model over existing methods. Multi-domain recommendation (MDR) aims to improve the recommendation performance in all target domains simultaneously by leveraging rich data from relevant domains. However, conventional approaches either only rely on sparse interactions in each separate scenario, or simply combine multiple shared recommendation solutions, resulting in MDR systems still suffering from severely low recommendation accuracy. To tackle the above issues, we propose a novel recommendation model called Hierarchical Graph Contrastive Learning for Multi-Domain Recommendation over Knowledge Graph (HKGCL) . Different from previous MDR models, our HKGCL can treat each relevant domain as a hierarchy in the interaction network. Based on the LightGCN aggregation strategy, a hierarchical message passing mechanism over the knowledge graph is designed to aggregate hierarchical knowledge representations for users and items. Then, a hierarchical node dropping scheme is proposed as our data augmentation technique on different hierarchical views to obtain more self-supervised semantic signals. Subsequently, three graph contrastive learning tasks in multiple and separate domains are proposed to explore domain-shared and domain-specific preference features for target users, respectively. Additionally, a novel Multi-domain Bayesian Personalized Ranking (MBPR) approach is proposed to assist in training our multi-task learning framework. Extensive experiments conducted on two real-world datasets demonstrate the consistent superiority of our proposed HKGCL over various state-of-the-art baselines, and also verify that the HKGCL can achieve strong performance on sparse interaction scenarios. [ABSTRACT FROM AUTHOR]