C2Explainer: Customizable Mask-based Counterfactual Explanation for Graph Neural Networks
Jiali Ma, Ichigaku Takigawa, Akihiro Yamamoto · 2025
As Graph Neural Networks (GNNs) have become increasingly popular for a wide range of applications involving graph data, there's a growing need to explain GNNs' predictions due to their complex black-box nature.Counterfactual Explanations (CEs) for GNNs are the tasks to generate a counterfactual graph that is similar to the input graph but predicted to have a different class label, providing insights into how an outcome could have differed if certain structures or features of the input graph had been changed.Common methods for graph CEs involve optimizing an edge mask with continuous values between 0 and 1, which faces two major issues: (1) such fractional masks can introduce unintended semantic changes, known as the "introduced evidence" problem; and (2) the mask can only be applied to existing edges, which limits the explanations to edge deletion only.To overcome these limitations, we propose a novel framework for graph CEs, C2Explainer, which produces higher-quality counterfactuals and allows flexible user customization to meet diverse needs or constraints across a wide range of specific domain tasks.We empirically demonstrate that C2Explainer significantly outperforms state-of-the-art baselines in the standard benchmark datasets.Furthermore, we present several use cases in real-world scenarios showcasing the flexibility of C2Explainer.