Research map / Graph learning / GNN architectures
On the Robustness of Graph Neural Diffusion to Topology Perturbations
Yang Song, QIYU KANG, Sijie Wang, Zhao Kai, Wee Peng Tay
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| Use Beltrami diffusion rather than standard non-PDE GNNs | Node classification models · PubMed under SPEIT and TDGIA | accuracy (%): 84.36 $\pm$ 0.58 -> 86.66 $\pm$ 0.68; 84.00 $\pm$ 1.12 -> 85.56 $\pm$ 0.91 | Improved |
| Use Beltrami diffusion rather than GNNGuard | Node classification models · Cora under SPEIT and TDGIA | accuracy (%): 78.50 $\pm$ 2.27 vs 61.87 $\pm$ 0.49; 78.92 $\pm$ 1.80 vs 62.84 $\pm$ 1.17 | Hurt |
| Combine Beltrami with GNNGuard | Beltrami · Citeseer and PubMed under SPEIT and TDGIA | accuracy (%): Citeseer: 66.46 $\pm$ 1.33 -> 69.90 $\pm$ 0.44; PubMed SPEIT: 86.66 $\pm$ 0.68 -> 87.68 $\pm$ 0.03 | Improved |
| Replace Beltrami flow with heat flow (ablation) | Graph neural PDE · Cora and Citeseer under SPEIT | accuracy (%): Cora: 61.87 $\pm$ 0.49 -> 38.66 $\pm$ 2.40; Citeseer: 66.46 $\pm$ 1.33 -> 22.95 $\pm$ 5.07 | Hurt |
| Replace heat flow with mean curvature or Beltrami flow (ablation) | Graph neural PDE · Cora and Citeseer under SPEIT | accuracy (%): Cora heat 38.66 $\pm$ 2.40; mean curvature 60.67 $\pm$ 1.31; Beltrami 61.87 $\pm$ 0.49. Citeseer heat 22.95 $\pm$ 5.07; | Improved |
| Replace Beltrami flow with mean curvature flow (ablation) | Graph neural PDE · Citeseer TDGIA | accuracy (%): 65.77 $\pm$ 1.28 -> 66.83 $\pm$ 1.59 | Improved |
| Replace Beltrami flow with GRAND/BLEND (ablation) | Graph neural PDE · Cora SPEIT and PubMed SPEIT | accuracy (%): Cora: 61.87 $\pm$ 0.49 -> 55.67 $\pm$ 3.60; PubMed: 86.66 $\pm$ 0.68 -> 39.39 $\pm$ 0.26 | Hurt |
| Remove spectral normalization/Lipschitz constraint (ablation) | Beltrami · Cora SPEIT, 4 layers | accuracy (%): 61.87 $\pm$ 0.49 -> 56.79 $\pm$ 1.52 | Hurt |
| Reduce the model to 1, 2, or 3 layers rather than 4 (ablation) | Beltrami · Cora SPEIT | accuracy (%): 4 layers: 61.87 $\pm$ 0.49; 1 layer: 59.10 $\pm$ 3.20; 2 layers: 63.58 $\pm$ 1.39; 3 layers: 60.52 $\pm$ 1.56 | Mixed |
| Use Dopri5 instead of implicit Adam with τ=2 (ablation) | Beltrami solver · Citeseer SPEIT | accuracy (%): 66.46 $\pm$ 1.33 -> 66.96 $\pm$ 1.46 | Improved |
| Increase injection attack strength (ablation) | Beltrami · Cora, Citeseer, and PubMed | accuracy (%) | No clear effect |
| Make heat diffusivity a trainable shared edge weight (ablation) | Heat flow · Citeseer SPEIT | accuracy (%): Heat+ 54.61 $\pm$ 2.92; Heat 22.95 $\pm$ 5.07 | Improved |
Closest papers by meaning
- Coupling Graph Neural Networks with Fractional Order Continuous Dynamics: A Robustness Study (AAAI 2024 · similarity 0.91)
- Hyperbolic-PDE GNN: Spectral Graph Neural Networks in the Perspective of A System of Hyperbolic Partial Differential Equations (ICML 2025 · similarity 0.90)
- Graph Neural PDE Solvers with Conservation and Similarity-Equivariance (ICML 2024 · similarity 0.88)
- Uncertainty Modeling in Graph Neural Networks via Stochastic Differential Equations (ICLR 2025 · similarity 0.88)
- A Generalized Neural Diffusion Framework on Graphs (AAAI 2024 · similarity 0.87)
- On the Stability of Graph Convolutional Neural Networks: A Probabilistic Perspective (NeurIPS 2025 · 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.
