GDCE-I: Discrete Diffusion Inversion for Graph Counterfactual Explanations

A new method for generating counterfactual explanations for Graph Neural Networks (GNNs) has been proposed in a paper on arXiv. The method, called Graph Diffusion Counterfactual Explanation via Inversion (GDCE-I), addresses the challenge of finding minimal structural modifications that change a GNN's prediction while respecting domain constraints. GNNs are widely used in chemistry, biology, and network analysis, but their lack of interpretability limits adoption in high-stakes settings. Counterfactual explanations are crucial for understanding model decisions, but on graphs, the search space is discrete and combinatorial, and valid edits must adhere to categorical node and edge types and domain rules like chemical valency. Existing explainers either fail to keep edits on the data manifold or do not explore the full edit space. GDCE-I overcomes these limitations by using a discrete denoising diffusion model to invert the counterfactual search, ensuring edits are both valid and comprehensive. The paper is available at arXiv:2608.12083 and was announced as a cross-type submission. The method is significant for improving the transparency and trustworthiness of GNNs in critical applications.