MAPTRec: a multi-view attention framework for heterogeneous cross-domain recommendation

Cross-domain recommendation offers a promising solution to alleviate data sparsity and cold-start issues by transferring knowledge from a source domain. However, existing methods often fail to integrate multi-view user preferences and are susceptible to transfer noise due to domain gaps. To address these limitations, we propose MAPTRec, a multi-view attention-enhanced preference transfer framework. It explicitly models feature heterogeneity and semantic misalignment across domains. Specifically, MAPTRec introduces a heterogeneous feature-aware graph convolution (HFGC) module to model high-order collaborative signals and semantic coupling in interaction graphs. A cross-view alignment module (CAM) then aligns multi-view features via dual cross-attention and dynamic gating mechanisms. Finally, the self-attentive user encoder (SAU) applies multi-head attention to capture internal dependencies within fused features. Experiments on Amazon and Douban datasets demonstrate that MAPTRec consistently outperforms strong baselines in terms of Recall@K and NDCG@K, particularly under sparse and cold-start settings.