Enhancing KGCN-Based Recommendation Algorithms via Attention Mechanism Integration

With the widespread deployment of recommender systems across various application scenarios, improving recommendation accuracy and personalization under data sparsity and cold-start conditions has become a key research challenge. Knowledge graphs, with their structured semantics and rich entity relationship modeling capabilities, provide valuable auxiliary information for recommendation tasks. As a representative graph-based recommendation model, KGCN demonstrates strong performance in entity representation learning. However, its uniform neighbor sampling strategy fails to effectively distinguish the importance of different neighbors, thereby limiting representation accuracy. To address this issue, this paper proposes an attention-based neighbor sampling method built upon DS-KGCN. By constructing user–relation–neighbor triples and introducing attention weight computation, the method enables dynamic modeling and weighted aggregation of key neighbors. Extensive experiments conducted on public datasets such as Book-Crossing and Last. FM show that the proposed method consistently outperforms existing models in terms of AUC, F1, and Recall, validating the effectiveness of the attention-based sampling strategy in enhancing recommendation accuracy and model expressiveness.