Toward Human-Interpretable Explanations in a Unified Framework for GNNs
Keywords: eXplainable AI, Graph Neural Networks
TL;DR: Toward Human-Interpretable Explanations in a Unified Framework for GNNs
Abstract: As Graph Neural Networks (GNNs) are increasingly applied across various domains, explainability has become a critical factor for real-world applications. Existing post-hoc explainability methods primarily focus on estimating the importance of edges, nodes, or subgraphs in the input graph to identify substructures crucial for predictions. However, these methods often lack human interpretability and do not provide a unified framework that incorporates both model-level and instance-level explanations. In this context, we propose leveraging a set of graphlets---small, connected, non-isomorphic induced subgraphs widely used in various scientific fields---and their associated orbits as human-interpretable units to decompose GNN predictions. Domain experts can select the most relevant graphlets as interpretable units and request unified explanations based on these units. To address this problem, we introduce UO-Explainer, the Unified and Orbit-based Explainer for GNNs, which utilizes predefined orbits that are generalizable and universal across graph domains as interpretable units. Our model decomposes GNN weights into orbit units to extract class-specific graph patterns (model-level) and to identify important subgraphs within individual data instances for prediction (instance-level). Extensive experimental results demonstrate that UO-Explainer outperforms existing baselines in providing meaningful and interpretable explanations across both synthetic and real-world datasets.
Primary Area: interpretability and explainable AI
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: 14225
Loading