Out-of-Distribution-Resistant Evaluations for Explanations of Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Explainability in Graph Neural Networks (GNNs) is crucial for trustworthiness and transparency.
• Current evaluation metrics for GNN explainability often suffer from out-of-distribution (OOD) challenges.
• These challenges arise when explanatory subgraphs do not match real-world data distributions, impacting explanation reliability.

Purpose of the Study:

• To develop a novel evaluation metric for GNN explainability that is robust to out-of-distribution data.
• To enhance the reliability and trustworthiness of GNN explanation techniques in practical applications.
• To establish a standardized framework for benchmarking GNN explainability metrics.

Main Methods:

• Introduced OOD-resistant Adversarial Robustness (OAR), inspired by adversarial robustness to assess subgraph resilience.
• Incorporated an OOD reweighting mechanism to maintain alignment with original data distributions.
• Developed a counterfactual attack module and utilized a conditional graph diffusion model for perturbed subgraphs, creating the OAR+ paradigm.

Main Results:

• Demonstrated the effectiveness of the OAR and OAR+ metrics through extensive experiments.
• The proposed metrics address OOD challenges, improving the reliability of GNN explainability assessments.
• The OAR+ paradigm offers versatility across various evaluation tasks.

Conclusions:

• OAR and OAR+ provide a robust and reliable method for evaluating GNN explainability.
• The developed metrics and standardized framework advance the field of trustworthy AI.
• The research contributes to more dependable real-world applications of GNNs.