Unearthing Undiscovered Interests: Knowledge Enhanced Representation Aggregation for Long-Tail Recommendation.

Graph neural networks have achieved remarkable performance in the field of recommender systems. However, existing graph-based recommendation approaches predominantly focus on suggesting popular items, disregarding the significance of long-tail recommendation and consequently falling short of meeting users’ personalized needs. To this end, we propose a novel approach called Knowledge-enhanced Representation Aggregation for Long-tail Recommendation (KRALR). Firstly, KRALR employs a user long-tail interests representation aggregation procedure to merge historical interaction information with rich semantic data extracted from knowledge graph (KG). By utilizing random walks on the KG and incorporating item popularity constraints, KRALR effectively captures the long-tail interests specific to the target user. Furthermore, KRALR introduces a long-tail item representation aggregation procedure by constructing a co-occurrence graph and integrating it with the KG. This integration enhances the quality of the representation for long-tail items, thereby enabling KRALR to provide more accurate recommendations. Finally, KRALR predicts rating scores for items that users have not interacted with and recommends the top N un-interacted items with the highest rating scores. Experimental results on the real-world dataset demonstrate that KRALR can improve recommendation accuracy and diversity simultaneously, and provide a wider array of satisfactory long-tail items for target users. Code is available at https://github.com/ZZP-RS/KRALR .