LLM-GRec: Large Language Model Enhanced Graph Recommendation via Granular-Ball Computing

Abstract Large Language Models (LLMs) have been introduced into recommender systems to enhance semantic understanding. However, LLM-generated user interests often consist of discrete phrases containing noise, redundancy, and structural instability. Moreover, the semantic gap between generated interests and collaborative recommendation objectives limits effective user--item matching. To address these challenges, we propose LLM-GRec, a robust graph-based recommendation framework enhanced by granular-ball computing. First, historical interactions and collaborative candidate items guide LLMs to generate recommendation-oriented user interests. Second, granular-ball computing organizes fine-grained interests into stable, multi-granularity interest anchors. Third, a unified graph integrates users, items, knowledge entities, and interest anchors for graph representation learning. An anchor alignment constraint encourages semantic consistency between user and item representations in the shared interest space. Furthermore, a quality-aware interest masking and reconstruction mechanism reduces the influence of noisy interests during graph propagation and recommendation learning. Experiments on three real-world datasets demonstrate that LLM-GRec outperforms existing recommendation methods on most evaluation metrics, with particularly notable improvements in sparse scenarios. The proposed framework effectively integrates semantic interest modeling, structured representation learning, and robust graph recommendation. Source code is available at https://github.com/shenjiale534/LLM-GRec.