Improved GraphSVX for GNN Explanations Based on Cross Entropy
Xin‐Yao Yu, Liang Dong, Qinfeng Li · 2023
Graph neural networks (GNNs) are a type of neural networks that can operate on graph data structures. And GNNs are difficult to explain. This lack of interpretability is a significant challenge in domains where transparency, accountability, and fairness are essential. This paper proposes an improved version of the GraphSVX method for explaining graph neural networks (GNNs) based cross-entropy. The proposed method combines the strengths of the existing explanation techniques and outperforms the original methods in terms of accuracy. It was evaluated on several benchmark datasets and a self-collected circuit dataset. The improved GraphSVX method provides a promising approach to explain GNNs for both graph classification tasks and node classification tasks, which can increase trust in the model’s decisions and facilitate the development of more transparent and accountable AI systems.