User-Oriented Interest Representation on Knowledge Graph for Long-Tail Recommendation

Graph neural networks have demonstrated impressive performance in the field of recommender systems. However, existing graph neural network recommendation approaches are proficient in capturing users’ mainstream interests and recommending popular items but fall short in effectively identifying users’ niche interests, thus failing to meet users’ personalized needs. To this end, this paper proposes the User-oriented Interest Representation on Knowledge Graph (UIR-KG) approach, which leverages the rich semantic information on the knowledge graph to learn users’ long-tail interest representation. UIR-KG maximizes the recommendation of long-tail items while satisfying users’ mainstream interests as much as possible. Firstly, a popular constraint-based long-tail neighbor selector is proposed, which obtains the target user’s long-tail neighbors by conducting high-order random walks on the collaborative knowledge graph and constraining item popularity. Secondly, a knowledge-enhanced hybrid attention aggregator is proposed, which conducts high-order aggregation of users’ long-tail interest representations on the collaborative knowledge graph by comprehensively combining relationship-aware attention and self-attention mechanisms. Finally, UIR-KG predicts the ratings of uninteracted items and provides Top-N recommendation results for the target user. Experimental results on real datasets show that compared with existing relevant approaches, UIR-KG can effectively improve recommendation diversity while maintaining recommendation accuracy, especially in long-tail recommendations. Code is available at https://github.com/ZZP-RS/UIR-KG