[RE] GNNBoundary: Finding Boundaries and Going Beyond Them

Abstract

Graph classification models are becoming increasingly popular, while explainability methods face challenges due to the discrete nature of graphs and other factors. However, investigating model decision-making, such as through decision-boundary regions, helps prevent misclassification and improve model robustness. This study aims to reproduce the findings of GNNBoundary: Towards Explaining Graph Neural Networks Through the Lens of Decision Boundaries (Wang & Shen, 2024). Their work supports 3 main claims: (1) their proposed algorithm can identify adjacent class pairs reliably, (2) their GNNBoundary can effectively and consistently generate near-boundary graphs outperforming the cross entropy baseline and (3) the generated near-boundary graphs can be used to accurately assess key properties of the decision boundary; margin, thickness, and complexity. We reproduce the experiments on the same datasets and extended them to two additional real-world datasets. Beyond that, we test different boundary probability ranges and their effect on decision boundary metrics, develop an additional baseline, and conduct hyperparameter tuning. We confirm the first claim regarding the adjacency discovery as well as the second claim that GNNBoundary outperforms the cross-entropy baseline under the limitation that it requires intensive hyperparameter tuning for convergence. The third claim is partially accepted as we observe a high variance between reported and obtained results, disproving the reliability and precision of the boundary statistics. Code and instructions are available at: https://github.com/jhb300/re_gnnboundary.