Graph Retrieval-Augmented LLM for Conversational Recommendation Systems
Zhangchi Qiu, Linhao Luo, Zicheng Zhao, Shirui Pan, Alan Wee-Chung Liew
TL;DR
This work tackles knowledge sparsity in Conversational Recommender Systems by leveraging a training-free Graph Retrieval-Augmented framework, G-CRS, that combines graph-based retrieval with in-context learning (ICL) for LLM-driven recommendations. The approach comprises a three-stage pipeline: Offline Indexing builds a Conversation-Entity Interaction Graph with a frequency matrix ${\mathbb P}$; Online Graph Retrieval uses a graph reasoner to expand seed entities ${\tilde{\mathcal E}}_t$ and Personalized PageRank (PPR) to produce candidate items ${\mathcal I}_k$ and similar conversations ${\mathcal C}_n$ via ${\mathbf r} = \alpha {\mathbf p} + (1-\alpha) {\mathcal A}' {\mathbf r}$ and ${\mathcal A}'={\mathcal A}{\mathcal D}^{-1}$; Retrieval-Augmented Reasoning and Recommendation then formats a structured prompt ${P_t}$ to the LLM to generate ${\mathcal I}_{t+1} \subseteq {\mathcal I}_k$ with a reasoning trace. Experiments on ReDial and INSPIRED show G-CRS achieves superior HR and MRR metrics without task-specific training, demonstrating the effectiveness of graph-structured retrieval in grounding LLM-based CRSs. The work highlights the potential of training-free, graph-aware RAG combined with ICL for scalable, contextually grounded recommendations in dialogue systems.
Abstract
Conversational Recommender Systems (CRSs) have emerged as a transformative paradigm for offering personalized recommendations through natural language dialogue. However, they face challenges with knowledge sparsity, as users often provide brief, incomplete preference statements. While recent methods have integrated external knowledge sources to mitigate this, they still struggle with semantic understanding and complex preference reasoning. Recent Large Language Models (LLMs) demonstrate promising capabilities in natural language understanding and reasoning, showing significant potential for CRSs. Nevertheless, due to the lack of domain knowledge, existing LLM-based CRSs either produce hallucinated recommendations or demand expensive domain-specific training, which largely limits their applicability. In this work, we present G-CRS (Graph Retrieval-Augmented Large Language Model for Conversational Recommender Systems), a novel training-free framework that combines graph retrieval-augmented generation and in-context learning to enhance LLMs' recommendation capabilities. Specifically, G-CRS employs a two-stage retrieve-and-recommend architecture, where a GNN-based graph reasoner first identifies candidate items, followed by Personalized PageRank exploration to jointly discover potential items and similar user interactions. These retrieved contexts are then transformed into structured prompts for LLM reasoning, enabling contextually grounded recommendations without task-specific training. Extensive experiments on two public datasets show that G-CRS achieves superior recommendation performance compared to existing methods without requiring task-specific training.