Interaction-Grounded Semantic Graph Refinement for LLM-Based Recommendation

Abstract

Large language models (LLMs) have recently demonstrated remarkable potential forrecommendation by reframing it as a text-generation task. Recent LLM-based approaches apply GNNsto capture higher-order collaborative patterns from interaction graphs but handle textual data separately,failing to extract higher-order semantic patterns from text. One potential solution is to construct semanticgraphs to better capture such semantic relationships. However, naively selecting top-k connections by profilesimilarity introduces significant challenges: 1) preference gap between textual similarity and actual userbehavior, and 2) semantic distortion where identical attributes carry different meanings for users versus items.To overcome these issues, we propose an Interaction-Grounded Semantic Recommender (IGSRec), whichconstructs an interaction-grounded semantic graph by aligning profile-based connections with observedinteractions. IGSRec employs an LLM-based profile generator, constructs a top-k semantic graph, thenrefines it using a learnable scoring function that identifies relevant semantic neighborhoods conditionedon user-item interactions. Through dual-graph propagation over both the refined semantic and interactiongraphs, IGSRec captures higher-order semantic and collaborative patterns. Experiments on Amazon Reviewbenchmarks demonstrate state-of-the-art performance in both direct and sequential recommendation tasks.