LLM-FSGNN: LLM-Guided Feature-Structure Augmentation Graph Neural Network for Cold-Start Recommendation

The cold-start recommendation has been challenging due to the limited historical interactions for new users and new items. Recently, methods based on meta-learning and graph neural networks are effective on this problem. However, these methods mainly focus on the missing user-item interactions in cold-start scenarios, overlooking the missing of user/item feature information, which significantly limits the quality and effectiveness of node embeddings. To address this problem, we propose an innovative Large Language Model-Guided Feature-Structure Augmented Graph Neural Network (LLM-FSGNN). The proposed framework integrates external semantic knowledge with internal graph structures via LLMs. It generates more comprehensive and robust user and item node representations. Specifically, we leverage the semantic reasoning capabilities of LLMs to extract textual descriptions of users and items, constructing a semantic view to enhance ambiguous or missing attributes and enrich node feature representations. Simultaneously, by analyzing item content to better understand user preferences, LLMs can accurately predict users’ potential intentions toward items and uncover latent interaction relationships, thus strengthening structural features. In addition, we introduce a graph structure enhancement module to complement structural relations from a graph-based perspective, mitigating hallucination issues in LLMs. Experimental results on multiple public benchmark datasets demonstrate consistent improvements in cold-start scenarios. For example, LLM-FSGNN achieves 4.8% lower MAE on MovieLens100K, 3.6% lower MAE on MovieLens1M, and 4.2% lower MAE on Amazon compared with the current state-of-the-art methods.