Research map / Graph learning / GNN architectures
Adversarial Weight Perturbation Improves Generalization in Graph Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add WT-AWP to GCN, GAT, or PPNP | GCN, GAT, and PPNP · Cora, Citeseer, and Polblogs | clean accuracy: GCN: 84.14 ± 0.61 -> 85.16 ± 0.44 (Cora); 73.44 ± 1.35 -> 74.48 ± 1.04 (Citeseer); 95.04 ± 0.66 -> 95.26 ± 0.51 (Polblog | Helped |
| Add WT-AWP to GCN | GCN · Proteins, IMDB-Binary, and IMDB-Multi | accuracy: 75.05 ± 1.40 -> 76.48 ± 0.49; 72.40 ± 2.73 -> 75.80 ± 1.17; 55.53 ± 1.33 -> 57.26 ± 0.63 | Helped |
| Add WT-AWP to GCN | GCN · Cora and Citeseer, 5% PGD and Metattack poisoning | robust accuracy: Cora PGD: 81.26 ± 1.27 -> 82.66 ± 1.07; Citeseer PGD: 72.04 ± 1.60 -> 73.73 ± 1.23; Cora Metattack: 78.61 ± 1.66 -> 79.0 | Helped |
| Include the unperturbed-loss term by using λ below 1 (ablation) | WT-AWP on GCN · Cora | accuracy: 60.29 ± 1.94 (λ=1, ρ=1) -> 85.16 ± 0.44 (λ=0.7, ρ=1) | Helped |
| Truncate perturbation to the first layer rather than the last layer (ablation) | WT-AWP on GCN · Cora | accuracy: 67.37 ± 5.01 (last layer, λ=0.7, ρ=1); first-layer value not tabulated in this table | Helped |
| Use 5-step instead of 1-step PGD for perturbations (ablation) | WT-AWP on GCN · Cora | accuracy: 84.99 ± 0.48 (5-step, λ=0.7, ρ=1); 85.16 ± 0.44 (1-step, λ=0.7, ρ=1) | No clear effect |
| Increase λ and ρ from small values toward λ=0.7 and ρ=1 (ablation) | WT-AWP on GCN · Cora | accuracy: 84.15 ± 0.60 (λ=0.1, ρ=0.05) -> 85.16 ± 0.44 (λ=0.7, ρ=1); 60.29 ± 1.94 (λ=1, ρ=1) | Mixed |
| Use vanilla AWP instead of weighted truncated AWP (ablation) | GCN · Cora | accuracy: 85.16 ± 0.44 (WT-AWP, λ=0.7, ρ=1) vs 29.18 ± 0.07 (AWP, ρ=1) | Hurt |
| Use weighted AWP without truncating the perturbed layers (ablation) | GCN · Cora | accuracy: 84.63 ± 0.51 (W-AWP, ρ=0.5) vs 85.00 ± 0.46 (WT-AWP, λ=0.7, ρ=0.5) | Hurt |
| Perturb all layers using AWP at larger ρ (ablation) | GCN · Linearly separable toy dataset | accuracy: 0.97 (ρ=0.5) -> 0.51 (ρ=2.5) | Hurt |
| Add WT-AWP to GCN | GCN · Cora, Citeseer, and Polblogs under DICE and PGD evasion attacks at 5% and 10% | robust accuracy: DICE 5%, Cora: 82.83 ± 0.87 -> 84.01 ± 0.59; PGD 10%, Polblogs: 72.90 ± 0.73 -> 75.11 ± 1.03 | Helped |
| Train with WT-AWP rather than natural training | GCN · Cora and Citeseer, noisy inputs around adjacency and node features | average input-gradient norm: 37 out of 40 comparisons had both better accuracy and smaller average gradient norm | Helped |
Closest papers by meaning
- Adversarial Robust Generalization of Graph Neural Networks (ICML 2025 · similarity 0.88)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 2025 · similarity 0.87)
- If You Want to Be Robust, Be Wary of Initialization (NeurIPS 2024 · similarity 0.87)
- Covered Forest: Fine-grained generalization analysis of graph neural networks (ICML 2025 · similarity 0.86)
- Generalization Error of Graph Neural Networks in the Mean-field Regime (ICML 2024 · similarity 0.86)
- Generalization of Graph Neural Networks Is Robust to Model Mismatch (AAAI 2025 · similarity 0.86)
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.