TopInG: Topologically Interpretable Graph Learning via Persistent Rationale Filtration
Keywords: topological data analysis, persistent homology, graph neural network, interpretability, explainability, filtration learning
TL;DR: A topologically interpretable GNN with a novel topological discrepancy loss is proved to be uniquely optimized by ground truth.
Abstract: Graph Neural Networks (GNNs) have shown remarkable performance in various scientific domains, but their lack of interpretability limits their applicability in critical decision-making processes. Recently, intrinsic interpretable GNNs have been studied to provide insights into model predictions by identifying rationale substructures in graphs. However, existing methods face challenges when the underlying rationale subgraphs are complicated and variable. To address this challenge,
we propose TopIng, a novel topological framework to interpretable GNNs that leverages persistent homology to identify persistent rationale subgraphs.
Our method introduces a rationale filtration learning technique that models the generating procedure of rationale subgraphs, and enforces the persistence of topological gap between rationale subgraphs and complement random graphs by a novel self-adjusted topological constraint, topological discrepancy. We show that our topological discrepancy is a lower bound of a Wasserstein distance on graph distributions with Gromov-Hausdorff metric.
We provide theoretical guarantees showing that our loss is uniquely optimized by the ground truth under certain conditions.
Through extensive experiments on varaious synthetic and real datasets, we demonstrate that TopIng effectively addresses key challenges in interpretable GNNs including handling variiform rationale subgraphs, balancing performance with interpretability, and avoiding spurious correlations.
Experimental results show that our approach improves state-of-the-art methods up to 20%+ on both predictive accuracy and interpretation quality.
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: 13674
Loading