GraphRAG based semantic representation for hybrid recommendation models

Abstract

Recommender systems have undergone significant evolution, transitioning from traditional context-based techniques to leveraging deep learning and graph-based methods. Existing models often rely heavily on ID-based data, overlooking the wealth of textual and contextual information embedded in user reviews. Therefore, in recent times, much research has been centred around incorporating textual information from explicit feedback through the use of LLMs. For example, RLMRec is a model-agnostic framework that integrates semantic representation learning with large language models combined with state-of-the art ID based collaborative filtering technologies. This paper explores the abilities of Microsoft’s GraphRAG to enhance the quality of textual information and semantic representations used in hybrid recommender models which use both semantic and collaborative signals. We leverage on GraphRAG’s Knowledge Graph and RAG based information extraction to generate richer semantic embeddings from user reviews and item metadata. Our experiments demonstrate that incorporating GraphRAG-derived embeddings improves recommendation quality, particularly in top-ranked positions. This research provides valuable insights into the strengths and limitations of GraphRAG as a semantic representation generation tool and suggests promising avenues toward more contextually aware hybrid recommender systems.