Keywords: Trustworthy, Robust, Graph Neural Network
Abstract: Explaining the prediction process of Graph Neural Network (GNN) is crucial for enhancing network transparency. However, real-world networks are predominantly heterogeneous and often beset with noise. The presence of intricate relationships in heterogeneous graphs necessitates a consideration of semantics during the explanation process, while mitigating the impact of noise remains unexplored. For GNN explainers heavily reliant on graph structure and raw features, erroneous predictions may lead to misguided explanations under the influence of noise. To address these challenges, we propose a Robust Heterogeneous Graph Neural Network Explainer with Graph Information Bottleneck, named RHGIB. We theoretically analyze the power of different heterogeneous GNN architectures on the propagation of noise information and exploit denoising variational inference. Specifically, we infer the latent distributions of both graph structure and features to alleviate the influence of noise. Subsequently, we incorporate heterogeneous edge types into the generation process of explanatory subgraph and utilize Graph Information Bottleneck framework for optimization, allowing the Explainer to learn heterogeneous semantics while enhancing robustness. Extensive experiments on multiple real-world heterogeneous graph datasets demonstrate the superior performance of RHGIB compared to state-of-the-art baselines.
Supplementary Material: zip
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.
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: 3728
Loading