Reproducibility Study of GNNBoundary: Towards Explain- ing Graph Neural Networks through the Lens of Decision Boundaries
Abstract: This study reproduces and extends GNNBoundary, a method for explaining Graph Neural Networks (GNNs) by analyzing decision boundaries between graph classes. GNNBoundary identifies adjacent class pairs and generates boundary graphs to provide insights into model behavior. We evaluate the reproducibility of key claims from the original work, including the identification of adjacent classes, the generation of accurate boundary graphs, and the effectiveness of an adaptive loss function in achieving faster convergence. Besides partly generating successful boundary graphs, our reproduction mostly highlights challenges with training variability and convergence, particularly with the Enzymes dataset. This suggests that GNNBoundary’s performance is sensitive to hyperparameter settings and random initialization. In addition, we extend GNNBoundary to handle three-class decision boundaries. While it demonstrated its feasibility, it also highlighted limitations in achieving balanced class separability and convergence. By assessing the abilities of GNNBoundary and the extension, this study contributes to improving the transparency and interpretability of GNN decision boundaries. Our findings emphasize the need for refined loss functions, additional baseline comparisons, and methodological extensions to more complex datasets for improved reliability.
Submission Length: Regular submission (no more than 12 pages of main content)
Previous TMLR Submission Url: https://openreview.net/forum?id=Kjsfz0HxtA¬eId=Kjsfz0HxtA
Changes Since Last Submission: The previous version was rejected due to incorrect TMLR formatting. Therefor we adapted the correct TMLR format and changed some minor thing to fit the page limit.
Assigned Action Editor: ~Guillaume_Rabusseau1
Submission Number: 4367
Loading