A graph neural network explainability strategy driven by key subgraph connectivity.

Abstract

Current explainability strategies for Graph Neural Networks (GNNs) often focus on individual nodes or edges, neglecting the significance of key subgraphs in decision-making processes. This limitation can result in dispersed and less reliable explanatory outcomes, particularly for complex tasks. This paper proposes a key subgraph retrieval method based on Euclidean distance, leveraging node representations obtained through training on the BA3 and Mutagenicity datasets to interpret GNN decisions. The proposed method achieves accuracies of 99.25% and 82.40% on the respective datasets. Performance comparison experiments with other mainstream explainability strategies, along with visualization analyses, demonstrate the effectiveness and robustness of this approach.