Insight at the right spot: Provide decisive subgraph information to Graph LLM with reinforcement learning

Abstract

Large language models (LLMs) cannot see or understand graphs. The current Graph LLM method transform graph structures into a format LLMs understands, utilizing LLM as a predictor to perform graph-learning task. However, these approaches have underperformed in graph-learning tasks. The issues arise because these methods typically rely on a fixed neighbor hop count for the target node set by expert experience, limiting the LLM’s access to only a certain range of neighbor information. Due to the black-box nature of LLM, it is challenging to determine which specific pieces of neighborhood information can effectively assist LLMs in making accurate inferences, which prevents LLMs from generating correct inferences. This study proposes to assist LLM in gaining insight at the right spot by providing decisive subgraph information to Graph LLM with reinforcement learning (Spider). A reinforcement subgraph detection module was designed to search for essential neighborhoods that influence LLM’s predictions. A decisive node-guided network was then applied to guide the reinforcement subgraph network, allowing LLMs to rely more on crucial nodes within the essential neighborhood for predictions. Essential neighborhood and decisive node information are provided to LLM in text form without the requirement of retraining. Experiments on five graph learning datasets demonstrate the effectiveness of the proposed model against all baselines, including GNN and LLM methods.