Cross-Domain Recommendation via Preference Propagation GraphNet

Recommendation can be framed as a graph link prediction task naturally. The user-item interaction graph built within a single domain often suffers from high sparsity. Thus, there has been a surge of approaches to alleviate the sparsity issue via cross-domain mutual augmentation. The SOTA cross-domain recommendation algorithms all try to bridge the gap via knowledge transfer in the latent space. We find there are mainly three problems in their formulations: 1) their knowledge transfer is unaware of the cross-domain graph structure. 2) their framework cannot capture high-order information propagation on the graph. 3) their cross-domain transfer formulations are generally more complicated to be optimized than the unified methods. In this paper, we propose the Preference Propagation GraphNet (PPGN) to address the above problems. Specifically, we construct a Cross-Domain Preference Matrix (CDPM) to model the interactions of different domains as a whole. Through the propagation layer of PPGN, we try to capture how user preferences propagate in the graph. Consequently, a joint objective for different domains is defined, and we simplify the cross-domain recommendation into a unified multi-task model. Extensive experiments on two pairs of real-world datasets show PPGN outperforms the SOTA algorithms significantly.