Addressing Structural Distribution Shift in Explanations for Graph Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Theory

Background:

• Graph Neural Networks (GNNs) are crucial for graph-structured data analysis.
• Explainability is vital for GNNs in critical applications, often achieved by identifying influential subgraphs.
• Existing methods overlook distribution shifts between training data and explanation subgraphs, impacting GNN generalization.

Purpose of the Study:

• To investigate the Out-Of-Distribution (OOD) problem in GNN explainability.
• To develop a theoretical framework for formalizing explanation subgraphs using sufficiency and minimality.
• To propose a novel approach for generating reliable GNN explanations.

Main Methods:

• Formalized explanation subgraph criteria (sufficiency, minimality).
• Introduced a novel concept of 'proxy graphs' to bridge distribution gaps.
• Utilized parametric and non-parametric optimization for proxy graph generation.

Main Results:

• Identified a fundamental distributional disparity between original and explanation subgraphs.
• Demonstrated that proxy graphs maintain explanatory information while aligning with original data distributions.
• Empirically validated improved quality and reliability of GNN explanations on diverse datasets.

Conclusions:

• The proposed proxy graph method effectively addresses the OOD challenge in GNN explainability.
• This work enhances the trustworthiness and transparency of GNN decision-making in critical domains.
• Advances the state-of-the-art in GNN explainability research.