The paper introduces KGAT (Knowledge Graph Attention Network), a novel method for recommendation systems that leverages knowledge graphs to model high-order relationships between users, items, and attributes. Traditional methods like factorization machines (FM) and neural factorization machines (NFM) treat each interaction as an independent instance, neglecting the relationships among instances or items. KGAT addresses this by integrating a knowledge graph into the recommendation model, allowing it to capture high-order connectivities in a collaborative knowledge graph (CKG). Key contributions of KGAT include: 1. **Recursive Embedding Propagation**: This layer updates node embeddings based on the embeddings of its neighbors, recursively propagating information to refine node representations. 2. **Attention Mechanism**: This mechanism learns the importance of each neighbor during the propagation process, enhancing the model's ability to capture relevant high-order relations. Empirical results on three public benchmarks (Amazon-book, Last-FM, and Yelp2018) show that KGAT outperforms state-of-the-art methods like Neural FM and RippleNet. The study also demonstrates the interpretability benefits of the attention mechanism and the effectiveness of embedding propagation for high-order relation modeling. The paper highlights the importance of explicitly modeling high-order relations in CKG to improve recommendation quality and provides a comprehensive evaluation of KGAT's performance and components. Future work could explore integrating additional structural information, such as social networks and item contexts, to further enhance recommendation systems.The paper introduces KGAT (Knowledge Graph Attention Network), a novel method for recommendation systems that leverages knowledge graphs to model high-order relationships between users, items, and attributes. Traditional methods like factorization machines (FM) and neural factorization machines (NFM) treat each interaction as an independent instance, neglecting the relationships among instances or items. KGAT addresses this by integrating a knowledge graph into the recommendation model, allowing it to capture high-order connectivities in a collaborative knowledge graph (CKG). Key contributions of KGAT include: 1. **Recursive Embedding Propagation**: This layer updates node embeddings based on the embeddings of its neighbors, recursively propagating information to refine node representations. 2. **Attention Mechanism**: This mechanism learns the importance of each neighbor during the propagation process, enhancing the model's ability to capture relevant high-order relations. Empirical results on three public benchmarks (Amazon-book, Last-FM, and Yelp2018) show that KGAT outperforms state-of-the-art methods like Neural FM and RippleNet. The study also demonstrates the interpretability benefits of the attention mechanism and the effectiveness of embedding propagation for high-order relation modeling. The paper highlights the importance of explicitly modeling high-order relations in CKG to improve recommendation quality and provides a comprehensive evaluation of KGAT's performance and components. Future work could explore integrating additional structural information, such as social networks and item contexts, to further enhance recommendation systems.