Enhancing Large Language Models with Multi-Type Collaborative Semantics for Recommendation

Abstract

Recommender systems play an essential role in helping users discover relevant content on digital platforms. Traditional collaborative‑filtering approaches focus on user–item interactions and often ignore additional semantic signals such as tags or categories. Recently, large language models (LLMs) have demonstrated strong ability to interpret textual information, but current LLM‑based recommenders either concentrate solely on text or rely on hand‑crafted prompts. This paper presents a unified framework that integrates heterogeneous graph neural networks with LLMs to capture multi‑type relations among users, items, tags and categories. The study employs a heterogeneous graph encoder (HGT) to learn collaborative embeddings, align these embeddings to the LLM token space, and construct structured prompts that incorporate both graph and textual information. Experiments on MovieLens‑1M and Amazon Beauty datasets show that the proposed model improves area under the curve (AUC) and normalized discounted cumulative gain (NDCG) over both traditional and recent LLM‑based baselines. The results suggest that combining graph structure with language understanding leads to more accurate and interpretable recommendations.