A Multi-Head Attention Network Integrating Knowledge Graph and Collaborative Filtering

In the field of recommendation systems, knowledge graphs (KG) have received widespread attention because they can proficiently address the challenges of sparsity and cold start encountered by conventional recommendation algorithms. The existing KG-based recommendation methods mainly consider how to fully utilize the knowledge information in the knowledge graph, ignoring the collaborative information provided by user-item interaction. This leads to the fact that the embedding vector learned by the existing model fails to effectively express the potential semantics of entities in the vector space. To tackle this challenge, this paper proposes a multi-head attention network integrating knowledge graph and collaborative filtering, which is named as collaborative knowledge-aware multi-head attention network (CKAN-MH). This network introduces a multi-head attention mechanism based on the traditional CKAN model to adaptively focus on subsets of different features, and achieves dynamic weighting of tail entities by adjusting attention weights. The addition of multi-head attention enables the model to better grasp complex relationships and patterns in the data, thereby improving recommendation performance. In addition, we applied the proposed CKAN-MH model on three real-world datasets and conducted experimental evaluations. The results indicate that CKAN-MH is significantly superior to several popular advanced baseline models, demonstrating its effectiveness and superiority.