Answers are grounded in this paper and cite their sources.
Counterfactual explanations improve the interpretability of graph neural networks (GNNs) by identifying small structural or feature perturbations that alter model predictions. They are especially useful in applications sensitive to relational changes, including molecular property prediction, fraud detection, and recommendation systems. However, existing methods often rely on local importance scores or heuristic search, limiting their ability to capture global graph structure and generate prediction-relevant graph edits. We propose LLM-CFExplainer, a general framework that uses large language models (LLMs) to assist counterfactual generation for GNNs. The framework extracts neighborhood-level influence information from a trained GNN and converts it, together with the original graph information, into structured prompts. The LLM then performs relational reasoning and proposes counterfactual edits subject to prediction-flipping, domain consistency, and basic topological constraints. A lightweight verification module evaluates the validity of the proposed edits and ensures consistency with the original graph. Experiments on multiple benchmark datasets show that LLM-CFExplainer improves counterfactual validity, coverage of perturbable regions, and coherence of generated explanations. These results demonstrate the potential of LLMs as reasoning tools for enhancing the interpretability of graph learning models. Our code is available at https://github.com/tanhaochen001-ux/LLM-CFExplainer .
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: