Absynthe-Motifs

Overview Absynthe-Motifs is a synthetic graph benchmark for evaluating Graph Neural Network (GNN) explainers under controlled, ground-truth conditions, with first-class support for absence-aware explanations — explanations that account for what is missing from a graph, not only for what is present. The dataset is generated with the Absynthe framework (companion to the paper Absynthe: A Framework for Absence-Aware Synthetic Graph Benchmarks andGNN Explainer Evaluation). Composition The release contains four sub-datasets, each obtained by composing instances of one or more structural motifs with a sequential composition pattern, plus 2 extra random vertices per graph (no extra random edges): Sub-dataset Motifs Motif counts (per graph) house house ~ Normal(μ=3, σ=1), ≥ 1 cycle_5 5-node cycle ~ Normal(μ=3, σ=1), ≥ 1 star_5 star with 5 leaves ~ Normal(μ=3, σ=1), ≥ 1 house_cycle5_star5 house + cycle_5 + star_5 each ~ Normal(μ=2, σ=1), ≥ 1 Each sub-dataset is built from 300 base graphs. Every base graph is then perturbed under three independent perturbation regimes, producing three perturbed variants per base graph (≈ 900 variants per sub-dataset, ≈ 3,600 variants in total): Perturbation folder Type Parameters remove_nodes Node removal 1 node edge_perturbation Joint edge add/remove ρ_remove = 0.10, ρ_add = 0.05 remove_edges Edge removal ρ_remove = 0.15