Research map / Graph learning / GNN architectures
Deformable Graph Convolutional 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 |
|---|---|---|---|
| use Deformable GCN instead of the strongest listed baseline | node classification · six heterophilic datasets | accuracy: Deformable GCN achieves the highest performance on all heterophilic graphs | Helped |
| use Deformable GCN instead of the strongest listed baseline | node classification · Citeseer and Pubmed | accuracy: 76.83 vs 77.48; 89.49 vs 89.51 | Mixed |
| use Deformable GCN instead of the strongest listed baseline | node classification · Cora | accuracy: 87.48 vs 87.48 | No clear effect |
| replace pre-trained node2vec or Poincare embeddings with learned smoothed-feature positional embeddings (ablation) | Deformable GCN · Wisconsin, Actor, Squirrel, and Pubmed | accuracy: Wisconsin: 67.57 / 68.1 -> 87.06; Actor: 35.04 / 35.15 -> 37.07; Squirrel: 45.27 / 46.31 -> 62.56; Pubmed: 88.35 / 87.26 | Helped |
| replace Deformable GConv with a GAT layer (ablation) | Deformable GCN · Wisconsin, Actor, Squirrel, and Pubmed | accuracy: Wisconsin: 70.54 -> 87.06; Actor: 36.26 -> 37.07; Squirrel: 61.62 -> 62.56; Pubmed: 88.04 -> 89.49 | Hurt |
| remove deformation from the convolution kernel (ablation) | Deformable GCN · Wisconsin and Pubmed | accuracy: 2~6 % on average | Hurt |
| remove the separating and focusing regularizers (ablation) | Deformable GCN · Wisconsin, Actor, Squirrel, and Pubmed | accuracy: Wisconsin: 84.12 -> 87.06; Actor: 36.67 -> 37.07; Squirrel: 60.31 -> 62.56; Pubmed: 89.02 -> 89.49 | Hurt |
| add the separating regularizer (ablation) | Deformable GCN without regularizers · Wisconsin and Squirrel | accuracy: Wisconsin: 84.12 -> 86.08; Squirrel: 60.31 -> 60.43 | Helped |
| add the separating regularizer (ablation) | Deformable GCN without regularizers · Pubmed | accuracy: 89.02 -> 88.86 | Hurt |
| add the focusing regularizer (ablation) | Deformable GCN without regularizers · Wisconsin and Squirrel | accuracy: Wisconsin: 84.12 -> 86.08; Squirrel: 60.31 -> 61.83 | Helped |
| add the focusing regularizer (ablation) | Deformable GCN without regularizers · Pubmed | accuracy: 89.02 -> 88.79 | Hurt |
| use both separating and focusing regularizers rather than either one alone (ablation) | Deformable GCN · Wisconsin, Actor, Squirrel, and Pubmed | accuracy: Wisconsin: 86.08 -> 87.06; Actor: 36.70 -> 37.07; Squirrel: 61.83 -> 62.56; Pubmed: 88.86 -> 89.49 | Helped |
Closest papers by meaning
- Simplifying Graph Convolutional Networks (ICML 2019 · similarity 0.90)
- All in a Row: Compressed Convolution Networks for Graphs (ICML 2023 · similarity 0.90)
- L2-GCN: Layer-Wise and Learned Efficient Training of Graph Convolutional Networks (CVPR 2020 · similarity 0.90)
- Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning (AAAI 2018 · similarity 0.90)
- Simple and Deep Graph Convolutional Networks (ICML 2020 · similarity 0.89)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.89)
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.