Enhancing Recommendation Effect through Knowledge Graph and Similarity Attention Mechanisms

To enhance the performance of recommendation models, numerous researchers have turned to knowledge graphs as auxiliary information. Nevertheless, prevailing recommendation methods primarily focus on utilizing knowledge graphs to encode entities more efficiently, often neglecting the crucial collaborative signals inherent in user-item interactions. This paper proposes the Knowledge-Graph Collaborative Diffusion Network (KCDN) model, which explicitly encodes collaborative signals in the knowledge graph and exploits the similarity attention mechanism to refine the embedding representation. KCDN model comprises of three integral components. The initial segment encompasses a heterogeneous propagation layer dedicated to collaborative signal and knowledge graph propagation. This facet enhances the preliminary representation of users and items. The subsequent element involves the similarity attention mechanism embedding layer, which normalizes the similarity between each ternary head entity and user-interacted item. It subsequently constructs embeddings for each layer through weighted summation, complementing the user and item representations set. The final facet pertains to the prediction layer, wherein the embeddings from distinct propagation layers are aggregated using pooling, summation, and concatenation techniques. This aggregation culminates in the prediction of click-through rates. The experimental results demonstrate the effectiveness of the KCDN paradigm that exhibits better performance and accuracy in several recommendation tasks compared with eight benchmark models in the field of recommendation system.