Enhancing High-order Interaction Awareness in LLM-based Recommender Model
The model ELMRec improves LLM-based recommendation by enhancing whole-word embeddings to capture complex, high-order user-item interactions without needing graph pre-training. It also adds a training-free reranking step to fix the LLM's tendency to prioritize older interactions in sequential recommendations. This results in state-of-the-art performance across various datasets.