Attentive Knowledge Graph Embedding for Personalized Recommendation

Knowledge graphs (KGs) have proven to be effective for highquality recommendation. However, existing methods mainly investigate separate paths connecting user-item pairs from KGs, thus failing to fully capture the rich semantics and underlying topology of KGs. We, therefore, propose a novel attentive knowledge graph embedding (AKGE) framwork to exploit the complex subgraphs of KGs linking user-item pairs to help better infer user preference. Specifically, AKGE first employs a distance-aware sampling strategy to automatically extract high-order subgraphs, which represent user-item relations with rich semantics. The subgraphs are then encoded by the proposed attentive graph neural network to help learn accurate user preference over items. Extensive validation shows that AKGE consistently outperforms state-of-the-arts. It additionally provides potential explanations for recommendation results.