Research map / Graph learning / GNN architectures
Mirage: Model-agnostic Graph Distillation for Graph Classification
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 Mirage, a model-agnostic graph distillation method for graph classification that mines frequently co-occurring computation trees with FP-Growth and trains GNNs on the resulting compact summary, running on CPU.
Problem
Graph distillation for GNN graph classification: existing methods need full-dataset training and gradient matching tied to a specific GNN architecture and hyperparameters, causing high cost and storage overhead.
Approach
Decompose each graph into a multiset of L-hop computation trees, exploit the skewed (power-law) tree frequency distribution, mine frequently co-occurring tree sets per class with FP-Growth, and train any message-passing GNN by sampling these tree sets, avoiding gradient replication entirely.
Key results
- Mirage ranks top-2 in AUC-ROC across all dataset-GNN combinations and is best in 8 of 17 combinations (e.g., NCI1 GCN 68.20 vs. full 87.03).
- Distilled datasets are about 4x smaller than DosCond and 5x smaller than KiDD on average (e.g., 318 bytes on NCI1 vs. 70,168 for DosCond GAT).
- Distillation is about 150x faster than DosCond and about 500x faster than KiDD on average, running on CPU only.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Mirage instead of the strongest reported baseline | Graph classification with GAT, GCN, or GIN · Across the reported dataset-architecture combinations | AUC-ROC | Mixed |
| Use Mirage instead of the strongest reported baseline | GCN graph classification · ogbg-molbace | AUC-ROC: 76.03 -> 77.03 | Helped |
| Use Mirage instead of the strongest reported baseline | GAT graph classification · ogbg-molbace | AUC-ROC: 73.75 -> 70.77 | Hurt |
| Use Mirage instead of the strongest reported baseline | GIN graph classification · ogbg-molbace | AUC-ROC: 77.09 -> 76.18 | Hurt |
| Use Mirage instead of the strongest reported baseline | GAT, GCN, or GIN graph classification · NCI1 | AUC-ROC | Helped |
| Use Mirage instead of the strongest reported baseline | Graph classification · ogbg-molbbbp | AUC-ROC | Mixed |
| Use Mirage instead of the strongest reported baseline | Graph classification · ogbg-molhiv | AUC-ROC | Mixed |
| Use Mirage instead of the strongest reported baseline | GAT, GCN, or GIN graph classification · DD | AUC-ROC | Helped |
| Use Mirage instead of the strongest reported baseline | Graph classification · IMDB-B | AUC-ROC | Mixed |
| Use Mirage instead of the strongest reported baseline | GAT, GCN, or GIN graph classification · IMDB-M | AUC-ROC | Helped |
| Increase the number of hops used for distillation (ablation) | Mirage · Graph classification; parameter variation experiments | AUC-ROC | Hurt |
| Train DosCond to convergence rather than use its random-initialized condensed dataset (ablation) | DosCond on ogbg-molbace with GCN · ogbg-molbace | AUC-ROC: 55.04±9.07 -> 67.34±1.84 | Helped |
Surprising findings
- The distilled dataset outperforms the full dataset on DD (e.g., Mirage GAT 76.08 vs. full 76.36 is close, and some baselines exceed full-data results), which the authors hypothesize is due to outlier removal.
- Random selection with SumPool (Random(sum)) is surprisingly strong and sometimes beats all graph distillation baselines.
Limitations
- Assumes the distribution of computation trees is skewed, which may not hold for heterophilous graphs.
- Does not address generalization to unseen tasks, temporal networks, or graph transformers and equivariant GNNs.
- Requires choosing a maximum number of GNN layers L in advance, and GAT results are omitted on IMDB datasets because GAT performs near chance there.
Experimental setup
Datasets: ogbg-molbace NCI1 ogbg-molbbbp ogbg-molhiv DD IMDB-B IMDB-M
Benchmarks: ogbg-molbace NCI1 ogbg-molbbbp ogbg-molhiv DD IMDB-B IMDB-M
Metrics: AUC-ROC distilled dataset size (bytes) distillation time (seconds)
Baselines: DosCond KiDD Herding Random
Builds on: DosCond KiDD GCond
Closest papers by meaning
- Diffusing to the Top: Boost Graph Neural Networks with Minimal Hyperparameter Tuning (ICLR 2025 · similarity 0.89)
- Graph Distillation with Eigenbasis Matching (ICML 2024 · similarity 0.89)
- GC4NC: A Benchmark Framework for Graph Condensation on Node Classification with New Insights (NeurIPS 2025 · similarity 0.87)
- Bonsai: Gradient-free Graph Condensation for Node Classification (ICLR 2025 · similarity 0.87)
- Distilling Graph Geometry: Knowledge Gap from GNNs to MLPs (NeurIPS 2026 · similarity 0.87)
- VQGraph: Rethinking Graph Representation Space for Bridging GNNs and MLPs (ICLR 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.