STExplainer: Global Explainability of GNNs via Frequent SubTree Mining

Shengyao Lu, Yakun Yu, Jiuding Yang, Baochun Li, Di Niu

Primary Area: general machine learning (i.e., none of the above)

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.

Keywords: Graph Neural Networks, Interpretability, Global-level explanation，

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose extracting high-quality global GNN explanations with rooted subtrees, offering insights into the relative importance of significant concepts.

Abstract: The need for transparency and interpretability in critical domains has led to an increasing interest in understanding the inner workings of Graph Neural Networks (GNNs). While local-level GNN explainability has been extensively studied to find important features within individual graph samples, recent research has emphasized the importance of global explainability of GNNs by uncovering global graphical concepts in a dataset underlying GNN behaviors. In this paper, we look into the intrinsic message-passing mechanism of standard GNNs and introduce a new method, STExplainer, to directly extract global explanations of GNNs using rooted subtrees on a dataset level instead of per instance. Unlike existing global explainers, which typically identify clusters of instance-level explanations or aggregate local graphical patterns into prototypes represented as latent vectors or rely on human-defined natural language rules, our approach extracts more intuitive global explanations through rooted subtree patterns and subgraph patterns, along with their associated relative importance scores, without relying on any instance-level explainers. We empirically demonstrate the effectiveness of our approach in extracting meaningful and high-quality global explanations on both synthetic and real-world datasets. The global explanations extracted by STExplainer are faithful to the original GNNs and distinguishable among different classes.

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: 2814