Empirical studies reveal improved quality and reliability of GNN explanations, suggesting better trust in decision-making.
Key Points
This research investigates the challenges in explaining Graph Neural Networks, particularly regarding distribution shifts between training and explanation subgraphs.
Developed a theoretical framework formalizing explanation subgraphs through sufficiency and minimality criteria.
Conducted theoretical analysis and empirical studies on diverse datasets.
Optimized explanatory information retention through parametric and non-parametric approaches.
Identified a fundamental distributional disparity between explanation subgraphs and original graphs.
Proposed the concept of proxy graphs to maintain essential explanatory information while conforming to original data distribution.
Showed improvements in the quality and reliability of GNN explanations through empirical evaluations.