Effective Graph-based Recommender Systems Using Large Language Models

Effective Graph-based Recommender Systems Using Large Language Models

Speaker: Yuze Chen
Host: Dr Ruihong Qiu

 

Abstract

Large language models (LLMs) provide strong semantic understanding and reasoning capabilities for recommender systems, but often face efficiency bottlenecks. In contrast, graph collaborative filtering (Graph CF) supports efficient large-scale recommendation, but remains highly dependent on the quality of user–item interaction graphs. Existing LLM-enhanced Graph CF methods mainly leverage LLMs to enhance node representations, while the underlying interaction graph structure is largely left unchanged. In real-world scenarios, interaction data can be sparse, may not always reflect users' genuine preferences, and can vary in importance as user intents and contexts change. Therefore, this research further explores leveraging LLM reasoning to directly improve the interaction graph structure, while maintaining the efficiency of Graph CF. This research examines LLM-enhanced interaction modelling in Graph CF through three research questions (RQs) concerning interaction sparsity, reliability, and adaptability.

First, we introduce CLEAR, namely Mining Consensus with LLMs for Edge Augmentation in Graph-based Recommendation. CLEAR leverages LLM semantic reasoning to enhance Graph CF at the interaction level. Candidate interactions are evaluated from complementary collaborative and semantic perspectives, and only those supported by both models are added as consensus edges. By iteratively introducing a small number of high-quality edges, CLEAR alleviates interaction sparsity while retaining efficient recommendation. (RQ1 Interaction Sparsity)

Second, we investigate how LLMs can improve interaction reliability. Observed interactions may be affected by accidental clicks, exposure bias, temporary interests, and other factors, and therefore may not accurately reflect users' genuine preferences. We use LLM assessments to supervise a lightweight model that predicts reliability weights across the graph, allowing questionable interactions to be downweighted without requiring LLM inference for every edge. (RQ2 Interaction Reliability)

Finally, we investigate how LLMs can enable Graph CF to use historical interactions adaptively. As user intents and contexts change, historical interactions may differ in relevance to the current recommendation. An offline LLM assesses which historical interests are most relevant to the current need, and this guidance is distilled into a lightweight predictor that dynamically weights interactions. Graph CF can therefore emphasise context-relevant history while reducing irrelevant information. (RQ3 Interaction Adaptability)

Experiments on Amazon Toys, Amazon Beauty, and Yelp show that the completed RQ1 work, CLEAR, consistently outperforms competitive methods, supporting the effectiveness of LLM-guided interaction enhancement.

Biography

Yuze Chen is a PhD candidate at the School of Electrical Engineering and Computer Science, The University of Queensland, supervised by Professor Xue Li, Associate Professor Rocky Chen, and Dr Priyanka Singh. He received his Master of Data Science from The University of Queensland. His research interests lie in recommender systems, graph-based collaborative filtering, graph learning, and the application of large language models to recommendation.