Graph-level Classification Datasets

This dataset package includes three benchmark datasets (Mutagenicity, BA3-motif, and MNIST-Superpixel) designed for evaluating explainability methods of Graph Neural Networks (GNNs).Mutagenicity: 4,377 molecular graphs (avg. 30.32 nodes\/30.77 edges per graph) with 14 node features, classified into mutagenic\/non-mutagenic labels, used for molecular toxicity prediction.BA3-motif: 3,000 synthetic graphs (avg. 31.44 nodes\/31.24 edges per graph) with 5 node features, constructed by attaching house\/cycle\/grid motifs to BA base graphs, providing ground-truth subgraphs for explanation validation.MNIST-Superpixel: 70,000 graphs derived from MNIST handwritten digits (avg. 66.87 nodes\/725.39 edges per graph) with 2 node features (pixel intensity + spatial location), suitable for image-related GNN explanation tasks.All datasets are split into training (80%), validation (10%), and test (10%) sets. They support evaluating key metrics of GNN explainers (e.g., prediction accuracy, fidelity, precision) and are optimized for causal inference-driven explanation methods.