[Re] GNNBoundary: Towards explaining Graph Neural Networks through the lens of decision boundaries
Abstract: Graph Neural Networks (GNNs) have been successfully applied to machine-learning tasks for graph structured data. However, their decision-making process remains difficult to interpret. The GNNBoundary method proposed by Wang & Shen (2024) is a model-level explanation method designed to analyze GNN decision boundaries. This study aims to reproduce and verify the claims made in the original paper: (1) GNNBoundary method can identify the adjacent classes, (2) GNNBoundary method can generate faithful near-boundary graphs, and (3) these graphs can be used to analyze the decision boundary. Experiments were conducted on four datasets, including the Proteins dataset, which extends the original work. To reproduce the results, we followed the authors’ open-sourced implementation. Our findings only partially support Claim 1, due to variations found in adjacent classes. Generally, we were able to generate faithful near-boundary graphs, mostly supporting Claim 2. The boundary analysis differed from the original results, but it was in line with results for adjacent classes and confusion matrices, partially verifying Claim 3. Further support for this claim was found on Proteins through a PCA visualization of the data.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Shuiwang_Ji1
Submission Number: 4287
Loading