RAGE: RAG Enhanced LLM Explainer for Heterogeneous Graphs

Raghvi Baloni, Chong Yihui, Ankur Teredesai

Keywords: Heterogeneous Graphs, Explainable AI (XAI), Retrieval Augmented Generation (RAG), Large Language Models

Abstract:

Generating accurate and interpretable explanations for predictions on heterogeneous graphs remains a significant challenge due to their multi-typed structures and complex relational dependencies. While Large Language Models (LLMs) have demonstrated strong performance in natural language tasks, their ability to provide grounded explanations for heterogeneous graphs is still underexplored. In this work, we introduce RAGE (Retrieval-Augmented Graph Explainer), a novel framework that enhances explanation quality by integrating Retrieval-Augmented Generation (RAG) with structured graph retrieval. RAGE retrieves subgraphs directly relevant to a given query, ensuring that explanations remain closely aligned with the dataset’s inherent structure.
We evaluate RAGE on two heterogeneous graph datasets, DBLP and Goodreads, across multiple LLMs. Through comprehensive experiments, we demonstrate that RAGE achieves comparable or superior predictive performance to metapath-based approach, while improving scalability. Furthermore, our qualitative evaluation highlights that RAGE produces more coherent and contextually accurate explanations, reducing the hallucination risks associated with indirect explanation approaches.
By offering a directly interpretable alternative to metapath-based explanation, RAGE provides a compelling framework for enhancing LLM-based explanation over heterogeneous graphs.

Submission Number: 7