Exploring Ensemble Strategies for Graph Counterfactual Explanations

Recent advancements in graph neural networks (GNNs) have significantly enhanced the performance of AI systems in tasks such as community detection, user friendship prediction, and drug discovery. However, the opaque nature of these models undermines user trust, especially in sensitive domains like health and finance. Graph Counterfactual Explanation (GCE) methods aim to mitigate this issue by providing insights into model predictions and suggesting user actions for alternative outcomes. Yet, GCEs produced by different methods often vary in quality, diversity, and alignment with the original model’s predictions. This work introduces an ensemble-based approach designed to address these inconsistencies by leveraging multiple GCE methods. Our approach comprises two main strategies: Selection, employing multi-criteria optimization to choose the optimal base explanation for each case, and Aggregation, combining multiple explanations to form a more robust overall explanation. We propose three selection strategies and six aggregation strategies. Our experimental evaluation demonstrates that these ensemble methods, particularly Ideal-Point Multi-Criteria Selection, consistently outperform individual GCE methods across diverse datasets in terms of quality, thereby significantly improving the interpretability of GNNs.