DKGFN: Deep Knowledge Graph Fusion Network for Recommendation

Knowledge graphs (KG) are rich in semantic entities and multi-hop relational structures, which can effectively capture implicit associations among item attributes. In contrast, user-item interaction graphs reflect explicit behavioral patterns between users and items. Integrating these two types of graphs enables the discovery of more complex and fine-grained high-order correlations between user preferences and item semantics, thereby enhancing the expressive power of representation learning and improving the accuracy of personalized recommendations. However, existing methods often lack deep integration between knowledge graphs and user-item interaction graphs. To address this limitation, we propose a deep dynamic fusion framework named DKGFN, which leverages layer-wise contrastive learning and a dual-view attention mechanism to achieve deep fusion of the two graph modalities. Experimental results demonstrate that DKGFN outperforms strong baselines by 8.3% and 11.2% in terms of Recall and NDCG, respectively, validating the effectiveness of our approach.