Research map / Graph learning / GNN architectures
Gradient Gating for Deep Multi-Rate Learning on Graphs
T. Konstantin Rusch, Benjamin Paul Chamberlain, Michael W. Mahoney, Michael M. Bronstein, Siddhartha Mishra
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 |
|---|---|---|---|
| Add gradient gating to GCN or GAT | GCN and GAT · Chameleon and Squirrel multi-scale node regression | normalized test MSE: Chameleon: 0.207 ± 0.039 -> 0.137 ± 0.033 (GCN); 0.207 ± 0.038 -> 0.136 ± 0.029 (GAT). Squirrel: 0.143 ± 0.039 -> 0.070 | Improved |
| Add gradient gating to message passing | GCN and GAT · 1000-layer grid propagation | Dirichlet energy | Improved |
| Use gradient gating in deeper GCNs | GCN on Cora · Cora test classification, 2 to 128 layers | accuracy | Improved |
| Add gradient gating to GCN or GAT | GCN and GAT · Synthetic Cora at strongest heterophily | accuracy: over 80% versus approximately 20% | Improved |
| Add gradient gating to GraphSAGE | GraphSAGE · Texas, Wisconsin, Film, Squirrel, Chameleon, Cornell | accuracy: average improvement of 18.6% | Mixed |
| Add gradient gating to GAT | GAT · Texas, Wisconsin, Film, Squirrel, Chameleon, Cornell | accuracy: average improvement of 45.75% | Mixed |
| Add gradient gating to GCN | GCN · Texas, Wisconsin, Film, Squirrel, Chameleon, Cornell | accuracy: average improvement of 45.4% | Mixed |
| Use gradient-gated GraphSAGE | GraphSAGE · snap-patents and arXiv-year | accuracy: by up to 13% | Improved |
| Increase the multi-rate strength α (ablation) | G²-GCN and G²-GAT · Chameleon node regression | test NMSE | Improved |
| Tune p across values from 1 to 5 (ablation) | G²-GraphSAGE · Texas, Squirrel, Film, Wisconsin, Chameleon | accuracy | No clear effect |
| Remove the additional rate-generating GNN (ablation) | G²-GraphSAGE · Film | accuracy: 37.14 ± 1.01 -> 36.83 ± 1.26 | Hurt |
| Remove the additional rate-generating GNN (ablation) | G²-GraphSAGE · Squirrel | accuracy: 64.26 ± 2.38 -> 55.78 ± 1.61 | Hurt |
Closest papers by meaning
- Optimization of Graph Neural Networks: Implicit Acceleration by Skip Connections and More Depth (ICML 2021 · similarity 0.89)
- Are GATs Out of Balance? (NeurIPS 2023 · similarity 0.88)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.87)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.87)
- Locality-Aware Graph Rewiring in GNNs (ICLR 2024 · similarity 0.86)
- On the Interplay between Graph Structure and Learning Algorithms in Graph Neural Networks (ICML 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.
