Research map / Graph learning / GNN architectures
From GNNs to Trees: Multi-Granular Interpretability 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.
Proposes TIF, a tree-like interpretable GNN for graph classification that builds hierarchical trees of progressively coarsened graphs with learnable perturbations and adaptive routing to yield multi-granular explanations.
Problem
Subgraph-based interpretable GNNs overemphasize local structure, and fixed-granularity graph coarsening (e.g., GIP) cannot capture relationships that span multiple granularities in graph classification.
Approach
Transforms a plain GNN into a tree whose levels hold coarsened graphs of different granularity, uses learnable perturbation matrices with similarity and diversity regularizers to diversify branches, and an adaptive router selects informative root-to-leaf paths that give both the prediction and multi-granular explanation.
Key results
- On D&D, TIF reaches 84.19% accuracy and 81.01 F1, above the best listed baseline (GIP at 79.32%).
- On MUTAG, TIF reports 94.44% accuracy, slightly above GSAT's 94.35%, and reports 86.23 F1 vs GIP's 86.73.
- On ENZYMES, TIF (58.66% accuracy) trails the best baseline GSAT (61.34%) and CAL/GIP, so it is not uniformly best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a compression ratio q that is neither too high nor too low (ablation) | TIF · D&D | classification accuracy | Mixed |
| Use 4 paths per node instead of fewer or more paths (ablation) | TIF · D&D | classification performance | Helped |
| Use 4 paths per node instead of fewer or more paths (ablation) | TIF · D&D | explanation accuracy | Helped |
| Replace MLP-based routing with a simpler linear router and remove inter-layer adaptive routing (ablation) | TIF · Graph classification benchmarks | classification performance | Hurt |
| Remove the graph perturbation module (ablation) | TIF · Graph classification benchmarks | classification performance | Hurt |
| Use TIF instead of the strongest listed baseline | Graph classification models · D&D | accuracy (%): 81.56 -> 84.19 | Helped |
| Use TIF instead of the strongest listed baseline | Graph classification models · MUTAG | accuracy (%): 94.35 -> 94.44 | Helped |
| Use TIF instead of the strongest listed baseline | Graph classification models · COLLAB | accuracy (%): 77.49 -> 77.29 | Hurt |
Surprising findings
- On ENZYMES, TIF's accuracy (58.66%) is below several interpretable baselines such as GSAT (61.34%), which the paper does not emphasize.
- Explanation accuracy on ENZYMES is 86.53 for TIF versus 87.34 for XGNN, so TIF is not best on every explanation metric.
Limitations
- Authors state future work on scaling from medium-scale to large-scale graphs at moderate computational cost, implying current scalability limits.
- Explanation metrics are self-designed (explanation accuracy via a trained GNN, consistency via random walk kernel), so comparisons are not standardized.
- Results on ENZYMES are weaker than several interpretable baselines.
Experimental setup
Datasets: ENZYMES D&D PROTEINS MUTAG COLLAB GraphCycle GraphFive MultipleCycle
Benchmarks: ENZYMES D&D PROTEINS MUTAG COLLAB GraphCycle GraphFive MultipleCycle
Metrics: classification accuracy F1 score explanation accuracy consistency path consistency path importance
Baselines: GCN DGCNN DiffPool RWNN GraphSAGE ProtGNN KerGNN GIB GSAT CAL GIP GNNExplainer
Builds on: GIP DiffPool GCN neural decision trees
Closest papers by meaning
- How Interpretable Are Interpretable Graph Neural Networks? (ICML 2024 · similarity 0.90)
- Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks (ICLR 2024 · similarity 0.88)
- Self-Interpretable Graph Learning with Sufficient and Necessary Explanations (AAAI 2024 · similarity 0.88)
- The Intelligible and Effective Graph Neural Additive Network (NeurIPS 2024 · similarity 0.88)
- Stratified GNN Explanations through Sufficient Expansion (AAAI 2024 · similarity 0.87)
- A Flexible, Equivariant Framework for Subgraph GNNs via Graph Products and Graph Coarsening (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.