Graph-Explain

An explainability library for graph-based models (GNNs).

Features

• Unified API: a single Explainer object for every method.

• 15 algorithms: GNNExplainer, PGExplainer, SubgraphX, Saliency, Integrated Gradients, GNNGatedLRP, DeepLift, Attention/GAT, GradXInput, GraphLIME, NodeMask, GuidedBackprop, Random baseline and Counterfactual.

• Node-level and graph-level (graph classification) explanations.

• Metrics: fidelity+ / fidelity-, GEA (node and graph), sparsity and stability.

• PyTorch Geometric and DGL backends, static and interactive visualization, natural-language narration and a full CLI with comparative benchmarking.

Contents

• API Reference
  • Core
  • Metrics
  • Comparative benchmark
  • Backends
  • Methods
  • Synthetic benchmarks
  • Narration
  • Visualization
  • CLI

Quick start

from graph_explain import Explainer, GNNExplainer, describe

# node-level
expl = Explainer(algorithm=GNNExplainer(epochs=120)).explain_node(
    data, model, node_idx=42)
print(describe(expl, data=data))

# graph-level (models with task_level="graph")
expl_g = Explainer(algorithm=GradXInput()).explain_graph(graph, graph_model)
from graph_explain import evaluate_gea_graph
gea = evaluate_gea_graph(expl_g, data=graph, top_k=13)

CLI

graph-explain explain --model model.pt --data data.pt \
    --method gnn_explainer --node 42 --describe --json report.json
graph-explain bench --model model.pt --data data.pt --node 42 \
    --json bench.json --html bench.html

# graph-level: omit --node
graph-explain explain --model model.pt --data graph.pt --method grad_x_input