HGAT-Rec: a hierarchical graph attention framework for cross-domain e-commerce recommendation with heterogeneity-aware contrastive alignment

Abstract Cross-domain recommendation over heterogeneous e-commerce networks faces three unresolved technical failures: existing GNN recommenders apply a single shared attention vector across all neighbor pairs, which can introduce a type-conflation bias governed by the inter-type variance of attention coefficients; meta-path importance weights learned in a source domain become systematically miscalibrated under the relational distribution shift of cross-domain transfer; and contrastive learning methods apply type-agnostic perturbations that assign equal dropout intensity to semantically distinct edge types, degrading representation quality in heterogeneous graph settings. We propose HGAT-Rec, which incorporates a heterogeneity-aware contrastive learning (HCL) objective that grounds view construction and sample selection in the typed relational structure of a cross-domain heterogeneous graph: type-stratified edge dropout preserves high-signal interaction channels proportionally to their attention weight, type-conditioned positive samples are drawn from meta-path neighborhoods sharing genuine relational content, and a cross-domain InfoNCE term enforces relative ordering constraints across the full overlapping user population. To support HCL, HGAT-Rec further provides (i) a type-triple-indexed graph attention mechanism parameterized by the joint combination of source node type, target node type, and edge type; (ii) a domain-conditioned meta-path aggregation layer with separate per-domain path importance distributions sharing a domain-invariant query vector; and (iii) a TransR-based knowledge gated fusion module decoupling item semantics from interaction sparsity. Theoretical grounding is provided by two formal propositions characterizing type-conflation bias and contrastive robustness under sparsity. Experiments on the Amazon Product Co-purchasing benchmark show that HGAT-Rec outperforms eleven baselines, achieving 11.20% and 8.70% improvements in NDCG@10 and HR@10 over the strongest competitor, a 31.80% gain for cold-start users, and a sparsity degradation rate of 35.2% under 80% interaction removal versus 49.0% for the next-best baseline.