Improving Graph Neural Network Interpretability through Conformal-based Graph Sparsification

CORES, a novel training approach for Graph Neural Networks, jointly identifies the most predictive subgraph and optimizes the graph classification performance, leading to more interpretable predictions.