Research map / Graph learning / Graph self-supervised learning
Graph Neural Network Generalization With Gaussian Mixture Model Based Augmentation
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 GRATIN, which fits per-class Gaussian Mixture Models via EM on GNN graph-level embeddings, samples augmented representations, and fine-tunes the post-readout layer, backed by a Rademacher-complexity generalization bound and influence-function analysis.
Problem
GNNs generalize poorly to unseen or out-of-distribution graphs, especially with small or low-diversity training sets, and existing graph augmentations are costly or weakly grounded in generalization theory.
Approach
Augment graph-level hidden representations rather than raw graphs: fit a class-conditional GMM (EM) to the trained GNN's readout embeddings, sample new embeddings, and fine-tune only the post-readout MLP; a Rademacher-based bound motivates controlling the representation-space deviation via KL divergence to a GMM that can approximate any density.
Key results
- GRATIN with GCN reaches 71.90% on DD (±2.81) versus 69.69% without augmentation, and 49.82% on IMDB-MULTI versus 47.73%.
- GRATIN with GIN achieves 88.83% on MUTAG (±5.02), versus 83.42% without augmentation, and 71.33% on PROTEINS.
- Under 20% training-edge corruption, GRATIN gets 73.10% on IMDB-BIN, outperforming the other augmentation baselines in the robustness table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply GMM-based augmentation with GRATIN instead of the strongest listed baseline | GCN · IMDB-BIN | accuracy: 74.00 ± 3.44 -> 71.00 ± 4.40 | Hurt |
| Apply GMM-based augmentation with GRATIN instead of the strongest listed baseline | GCN · IMDB-MUL | accuracy: 49.80 ± 4.71 -> 49.82 ± 4.26 | Helped |
| Apply GMM-based augmentation with GRATIN instead of the strongest listed baseline | GCN · MUTAG | accuracy: 88.77 ± 5.71 -> 76.05 ± 6.74 | Hurt |
| Apply GMM-based augmentation with GRATIN instead of the strongest listed baseline | GCN · PROTEINS | accuracy: 70.07 ± 3.86 -> 70.97 ± 5.07 | Helped |
| Apply GMM-based augmentation with GRATIN instead of the strongest listed baseline | GCN · DD | accuracy: 70.11 ± 4.48 -> 71.90 ± 2.81 | Helped |
| Apply GMM-based augmentation with GRATIN instead of the strongest listed baseline | GIN · IMDB-BIN | accuracy: 72.50 ± 4.98 -> 71.70 ± 4.24 | Hurt |
| Apply GMM-based augmentation with GRATIN instead of no augmentation | GIN · IMDB-MUL | accuracy: 48.53 ± 4.05 -> 49.20 ± 2.06 | Helped |
| Apply GMM-based augmentation with GRATIN instead of the strongest listed baseline | GIN · MUTAG | accuracy: 87.77 ± 7.48 -> 88.83 ± 5.02 | Helped |
| Apply GMM-based augmentation with GRATIN instead of the strongest listed baseline | GIN · PROTEINS | accuracy: 70.44 ± 2.58 -> 71.33 ± 5.04 | Helped |
| Apply GMM-based augmentation with GRATIN instead of the strongest listed baseline | GIN · DD | accuracy: 69.01 ± 3.95 -> 68.61 ± 4.62 | Hurt |
| Use GRATIN rather than other listed augmentation strategies under edge corruption | GCN · IMDB-BIN, IMDB-MUL, PROTEINS, and DD at 10% and 20% structure corruption | accuracy | Helped |
| Filter candidate augmentations by validation influence and retain the high-influence subset | Fisher-Guided GMM Augmentation · graph classification test accuracy | accuracy | Helped |
Surprising findings
- On DD with GIN, GRATIN has no measurable effect, which the authors attribute to softmax saturation and vanishing loss gradients with respect to input graphs.
- G-Mixup attains the best MUTAG score with GCN (88.77%) but performs poorly on PROTEINS and DD, so no single augmentation dominates across datasets.
Limitations
- Baseline results show high standard deviations, and some gains are within one standard deviation of the baselines.
- Influence-based filtering uses validation graphs rather than test labels, and the paper reports only a single training run for the filtering experiment.
- Augmentation effectiveness depends on the GNN backbone and dataset; GRATIN has no effect on DD with GIN.
Experimental setup
Datasets: IMDB-BINARY IMDB-MULTI MUTAG PROTEINS DD
Benchmarks: IMDB-BINARY IMDB-MULTI MUTAG PROTEINS DD
Metrics: classification accuracy
Baselines: DropEdge DropNode SubMix G-Mixup GeoMix NoisyGNN
Builds on: GCN GIN G-Mixup GeoMix Manifold-Mixup
Closest papers by meaning
- Data Augmentation for Graph Neural Networks (AAAI 2021 · similarity 0.89)
- Towards Understanding Generalization of Graph Neural Networks (ICML 2023 · similarity 0.89)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 2025 · similarity 0.87)
- Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization (NeurIPS 2025 · similarity 0.87)
- On the Rademacher Complexity of Graph Neural Networks: Unifying Expressivity and Geometry (NeurIPS 2026 · similarity 0.87)
- Generalization Error of Graph Neural Networks in the Mean-field Regime (ICML 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.