Generating Global Factual and Counterfactual Explainer for Molecule Under Domain Constraints

Abstract

Graph neural networks (GNNs) are powerful tools for handling graph-structured data but often lack transparency. This paper aims to generate interpretable global explanations for GNN predictions, focusing on real-world scenarios like chemical molecules. We develop an approach that produces both factual and counterfactual explanations while incorporating domain constraints, ensuring validity and interpretability for domain experts. Our contributions include creating global explanations, integrating domain constraints, and improving random walk in global explanations using fragment-based editing. We demonstrate the effectiveness of our approach on AIDS and Mutagenicity datasets, providing a comprehensive understanding of GNNs and aiding domain experts in evaluating generated explanations.