Graph Distributional Analytics: Enhancing GNN Explainability through Scalable Embedding and Distribution Analysis

27 Sept 2024 (modified: 13 Nov 2024)ICLR 2025 Conference Withdrawn SubmissionEveryoneRevisionsBibTeXCC BY 4.0
Keywords: Graph Neural Networks, Explainability, Graph Distributional Analytics, Weisfeiler-Leman graph kernel, Graph embeddings, Distributional analysis, Out-of-distribution data, Model transparency, Structural features, Machine learning, Graph classification, Scalable methods, GNN interpretability, Model robustness
TL;DR: We introduce Graph Distributional Analytics, combining Weisfeiler-Leman graph kernels with distributional analysis to enhance GNN explainability by quantifying graph data distributions and identifying structural causes of misclassifications.
Abstract: Graph Neural Networks (GNNs) have achieved significant success in processing graph-structured data but often lack interpretability, limiting their practical applicability. We introduce the Graph Distributional Analytics (GDA) framework, leveraging novel combinations of scalable techniques to enhance GNN explainability. The integration of Weisfeiler-Leman (WL) graph kernels with distributional distance analysis enables GDA to efficiently quantify graph data distributions, while capturing global structural complexities without significant computational costs. GDA creates high-dimensional embeddings employing WL kernels, measures the distribution of distances from measures of categorical central tendency, and assigns distribution scores to quantify each graph's deviation from this vector We evaluate GDA on the ENZYMES, ogbg-ppa, and MalNet-Tiny datasets. Our experiments demonstrate GDA not only accurately characterizes graph distributions but also outperforms baseline methods in identifying specific structural features responsible for misclassifications. This comprehensive analysis provides deeper insights into how training data distributions affect model performance, particularly with out-of-distribution (OOD) data. By revealing the underlying structural causes of GNN predictions through a novel synergy of established techniques, GDA enhances transparency and offers a practical tool for practitioners to build more interpretable and robust graph-based models. Our framework's scalability, efficiency, and ability to integrate with various embedding methods make it a valuable addition to the suite of tools available for GNN analysis.
Primary Area: learning on graphs and other geometries & topologies
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12275
Loading