UniCRSxGraphRAG: Enhancing recommendations and delivering explainable insights in conversational recommender systems

Abstract

Conversational Recommender Systems (CRSs) are an up-and-coming advancement in the AI race, offering their users personalized recommendations via natural language dialogue in various domains. These systems aim to leverage responsive and context-aware interactions to provide accurate recommendations. Traditional CRSs have three main modules: conversation, recommendation, and semantic fusion. The semantic fusion component, in particular, presents challenges in maintaining consistent meaning across components. In addition, traditional CRS systems also struggle to provide transparent explanations of their recommendations. UniCRSxGraphRAG aims to provide a novel unified approach to address these challenges. This project leverages a large language model (LLM)- based framework—GraphRAG—to dynamically construct knowledge graphs and eliminate the need for explicit semantic alignment strategies. Unlike static KGs (e.g., DBpedia), GraphRAG extracts entities and their relationships via LLMs, enabling stronger natural reasoning traceability. This framework allows us to improve the explainability of recommendations without requiring hand-crafted semantic fusion. Furthermore, for the conversation and recommendation tasks, we build on the success of UniCRS to use its prompt-based learning paradigm to adapt pre-trained LLMs to the downstream tasks. The experimental results are evaluated on a benchmark dataset and demonstrate improved performance in downstream conversation and recommendation tasks, providing transparency in the recommendation process.