Research map / Graph learning / Graph self-supervised learning
[Re] GNNInterpreter: A probabilistic generative model-level explanation for Graph Neural Networks
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.
Reproduces GNNInterpreter, a model-level explanation method that learns a Gilbert-graph generative distribution via Concrete relaxation and gradient ascent, and compares it against XGNN on MUTAG and synthetic datasets.
Problem
Model-level explanation of message-passing GNN graph classifiers: generating faithful, valid explanation graphs for a target class without manually specified domain rules, and supporting discrete or continuous node/edge features.
Approach
Learn a probabilistic generative graph distribution (Gilbert random graph with Binary/Categorical Concrete relaxation) by gradient ascent on the target class logit plus a cosine-similarity term to the average training-class graph embedding, instead of RL with hand-crafted rules as in XGNN.
Key results
- On MUTAG, GNNInterpreter explanations reach class probability 1.000 ± 0.000 for both classes vs 0.986 and 0.991 for XGNN, with ~12 s per class vs 128 s for XGNN.
- On Motif, House-class explanations reach 0.918 ± 0.268 probability, while the other classes reach ≥0.998; the verification study found 8 of 9 plausible House-X motifs classified as House-X in large fractions (e.g., 3566/5000 for motif 1).
- On Is_Acyclic, GNNInterpreter gives 0.999 (Cyclic) and 1.000 (Acyclic) probability, whereas XGNN gives 0.076 for Cyclic; the authors could not reproduce XGNN's claimed results on Is_Acyclic.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GNNInterpreter instead of XGNN | MUTAG explanations · MUTAG Mutagen | class probability: 0.986 ± 0.057 -> 1.000 ± 0.000 | Helped |
| Use GNNInterpreter instead of XGNN | MUTAG explanations · MUTAG Nonmutagen | class probability: 0.991 ± 0.083 -> 1.000 ± 0.000 | Helped |
| Use GNNInterpreter instead of XGNN | Is_Acyclic explanations · Is_Acyclic Cyclic | class probability: 0.076 ± 0.000 -> 0.999 ± 0.001 | Helped |
| Use GNNInterpreter instead of XGNN | Is_Acyclic explanations · Is_Acyclic Acyclic | class probability: 0.927 ± 0.000 -> 1.000 ± 0.000 | Helped |
| Use GNNInterpreter instead of XGNN | MUTAG explanations · MUTAG | training time per class: 128 s -> 12 s | Helped |
| Use GNNInterpreter instead of XGNN | Is_Acyclic explanations · Is_Acyclic | training time per class: 45 s -> 20 s | Helped |
| Add the embedding-similarity term to the objective (ablation) | GNNInterpreter · MUTAG Mutagen | logit: 310.14 -> 43.45 | Hurt |
| Add the embedding-similarity term to the objective (ablation) | GNNInterpreter · MUTAG Mutagen | explanation fidelity to data distribution | Helped |
| Use random graphs instead of generated explanations | GNNInterpreter evaluation · MUTAG Mutagen | class probability: 0.068 ± 0.251 vs 1.000 ± 0.000 | Hurt |
Surprising findings
- The Consistent-class explanation on ColorConsistency violates the true class definition, suggesting the GNN may misclassify graphs with few inconsistent edges despite 0.9963 test accuracy.
- The MUTAG Nonmutagen explanation contains no special pattern yet reaches class probability 1, indicating a likely bias attribution issue.
Limitations
- XGNN comparison is limited to MUTAG and Is_Acyclic because XGNN could not be made to produce acceptable explanations on Motif and Shape.
- Explanation graphs for Lollipop class have only 0.742 average class probability, and some explanations deviate from desired patterns.
- Independence assumption for node and edge features is acknowledged as potentially violated.
Experimental setup
Datasets: MUTAG Cyclicity Motif Shape Is_Acyclic ColorConsistency
Benchmarks: MUTAG Cyclicity Motif Shape Is_Acyclic ColorConsistency
Metrics: average predicted class probability over 1000 explanation gr training time per class classification counts on plausible motif verification sets
Baselines: XGNN random graph baseline
Builds on: GNNInterpreter XGNN PGExplainer Gumbel-Softmax GCN GNNExplainer
Closest papers by meaning
- GNNInterpreter: A Probabilistic Generative Model-Level Explanation for Graph Neural Networks (ICLR 2023 · similarity 0.92)
- GraphTrail: Translating GNN Predictions into Human-Interpretable Logical Rules (NeurIPS 2024 · similarity 0.86)
- Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks (ICLR 2024 · similarity 0.86)
- EiG-Search: Generating Edge-Induced Subgraphs for GNN Explanation in Linear Time (ICML 2024 · similarity 0.85)
- Generating In-Distribution Proxy Graphs for Explaining Graph Neural Networks (ICML 2024 · similarity 0.85)
- On Neural Networks as Infinite Tree-Structured Probabilistic Graphical Models (NeurIPS 2024 · similarity 0.85)
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.