Knowledge Memory Graph convolution network for cross-domain recommendation

Cross-domain recommender systems (CDR) address the data sparsity issue by leveraging information from relevant domains. Nowadays, graph neural networks (GNNs) are widely employed to capture higher-order collaborative relationships, further enhancing the effectiveness of CDR in tackling data sparsity. However, GNN-based CDR methods still face challenges in modeling comprehensive user preferences: (1) existing methods relying solely on GNNs struggle to capture fine-grained user preferences (e.g., attribute level) due to limitations in semantic information decomposition; (2) the learned embeddings are highly abstract, making them difficult to understand and interpret; (3) the optimal performance of existing methods typically bottlenecks at layers 2$\sim$4, which is caused by the inherent over-smoothing problem of GNNs. To address these challenges, we propose a method called Knowledge Memory Graph convolution network for Cross-Domain Recommendation (KMGCDR). Specifically, we incorporate an encyclopedic knowledge graph (KG) to enrich the semantic understanding of user behavior. Then, we extend the GNN-based model with a knowledge-enhanced memory network that can store external KG information, aiming to capture user’s fine-grained preferences in an interpretable manner. This design can also widen the gap between the low-correlation data and reduce the smoothness of the graph representation. To our best knowledge, this is the first attempt to employ a memory network to address the over-smoothing problem in CDR. Extensive experiments on real-world datasets demonstrate the superiority of KMGCDR. Compared to the best-performing SOTA baseline method in six cross-domain scenarios, KMGCDR achieves an average of 10.07% and 5.55% improvements on Amazon and Facebook datasets, respectively.