Research map / Graph learning / GNN architectures
MAGE: Model-Level Graph Neural Networks Explanations via Motif-based Graph Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MAGE, a motif-based model-level explainer for molecular GNNs that selects class-specific motifs via attention and generates valid molecular explanations with a VAE junction-tree generator.
Problem
Model-level explanations for GNNs on molecular graphs (XGNN, GNNInterpreter) often produce chemically invalid graphs and miss substructures such as rings.
Approach
Use motifs (bridge-bond decomposition) as generation units; learn class-specific motifs via a single-layer attention over molecule-motif relations, combined with class probabilities; then generate explanations with a motif junction-tree VAE whose property loss aligns with the target GNN.
Key results
- MAGE achieves validity 1.00 on all six datasets, versus XGNN (0.295 to 0.81 average where run) and GNNInterpreter (0.07 to 0.60).
- On average class probability, MAGE reaches 0.9959 on Mutagenicity and beats baselines in 10 of 12 label settings.
- MAGE sampling takes about 0.003 s per explanation versus 12-62 s for XGNN and GNNInterpreter.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use motif-based generation with valid substructures | MAGE · Mutagenicity | validity (Average): 0.295 -> 1.00 | Helped |
| use motif-based generation with valid substructures | MAGE · PTC_MR | validity (Average): 0.45 -> 1.00 | Helped |
| use motif-based generation with valid substructures | MAGE · PTC_MM | validity (Average): 0.51 -> 1.00 | Helped |
| use motif-based generation with valid substructures | MAGE · PTC_FM | validity (Average): 0.42 -> 1.00 | Helped |
| use motif-based generation with valid substructures | MAGE · AIDS | validity (Average): 0.81 -> 1.00 | Helped |
| use motif-based generation with valid substructures | MAGE · NCI-H23 | validity (Average): 0.60 -> 1.00 | Helped |
| add attention-based motif identification instead of individual-motif scoring (ablation) | MAGE · Mutagenicity | Average Probability: 0.9892 -> 0.9959 | Helped |
| add attention-based motif identification instead of individual-motif scoring (ablation) | MAGE · PTC_MR | Average Probability: 0.9771 -> 0.9939 | Helped |
| add attention-based motif identification instead of individual-motif scoring (ablation) | MAGE · PTC_MM | Average Probability: 0.9738 -> 0.9873 | Helped |
| add attention-based motif identification instead of individual-motif scoring (ablation) | MAGE · PTC_FM | Average Probability: 0.9836 -> 0.9934 | Helped |
| add attention-based motif identification instead of individual-motif scoring (ablation) | MAGE · AIDS | Average Probability: 0.9870 -> 0.9893 | Helped |
| add attention-based motif identification instead of individual-motif scoring (ablation) | MAGE · NCI-H23 | Average Probability: 0.9868 -> 0.9935 | Helped |
Surprising findings
- GNNInterpreter's average-probability score on NCI-H23 (0.9940) is higher than MAGE's (0.9935), so MAGE does not win on every dataset.
- Training time for MAGE (up to 13657 s on NCI-H23) is far longer than the baselines' sampling-only cost, though per-sample sampling is faster.
Limitations
- Experiments use only a simple 3-layer GCN as the target model.
- Validity is reported as 1.00 by construction from motif assembly, and the evaluation covers only binary classification datasets.
- The motif extraction used in the main experiments is bridge-bond cutting only; the paper's own text says other extractors were only described, not evaluated.
Experimental setup
Datasets: Mutagenicity PTC_MR PTC_MM PTC_FM AIDS NCI-H23
Benchmarks: Mutagenicity PTC_MR PTC_MM PTC_FM AIDS NCI-H23
Metrics: validity average class probability novelty sampling time
Baselines: XGNN GNNInterpreter Variant (motif scoring without attention) GNNExplainer GEM
Builds on: XGNN GNNInterpreter junction tree VAE GNNExplainer
Closest papers by meaning
- Towards Multiscale Graph-based Protein Learning with Geometric Secondary Structural Motifs (NeurIPS 2025 · similarity 0.88)
- On the Scalability of GNNs for Molecular Graphs (NeurIPS 2024 · similarity 0.88)
- B-XAIC Dataset: Benchmarking Explainable AI for Graph Neural Networks Using Chemical Data (NeurIPS 2026 · similarity 0.88)
- Pre-Training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck (AAAI 2025 · similarity 0.87)
- RISE: Radius of Influence based Subgraph Extraction for 3D Molecular Graph Explanation (ICML 2025 · similarity 0.87)
- LLaMo: Large Language Model-based Molecular Graph Assistant (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.