Research map / Graph learning / GNN architectures
Simplifying 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 |
|---|---|---|---|
| remove intermediate nonlinearities and collapse layer weights, retaining graph propagation | GCN · Citeseer | test accuracy (%): 70.9± 0.5 -> 71.9 ± 0.1 | Helped |
| remove intermediate nonlinearities and collapse layer weights, retaining graph propagation | GCN · Reddit | training time: 270.45 -> 2.70 | Helped |
| replace the four-layer highway GCN with SGC using K=3 | GCN+H · TWITTER-WORLD | Acc.@161: 53.6 ± 0.2 -> 54.1 ± 0.2 | Helped |
| replace the GCN in C-GCN with SGC | C-GCN · TACRED | test accuracy (%): 66.4 ±{0.4} -> 67.0 ±{0.4} | Helped |
| replace GCN with an MLP followed by SGC | GCNZ · ImageNet zero-shot, unseen categories | top-1 accuracy (%): 20.9 ± 0.2 -> 21.2 ± 0.2; 4.3 ± 0.0 -> 4.4 ± 0.1 | Helped |
| replace GCN with SGC in graph classification | DCGCN · QM8 | MAE: 0.03 MAE vs 0.01 MAE | Hurt |
| use propagation matrices with added self-loops (ablation) | SGC · citation networks, validation sets | accuracy | Helped |
| increase propagation depth K with the first-order Chebyshev matrix (ablation) | SGC · citation networks, validation sets | accuracy | Hurt |
| use odd propagation powers without self-loops (ablation) | SGC with normalized-adjacency or random-walk propagation · citation networks, validation sets | accuracy | Hurt |
| use odd propagation powers with self-loops (ablation) | SGC with augmented propagation · citation networks, validation sets | accuracy | No clear effect |
| reduce the number of training samples (ablation) | SGC compared with GCN · Cora validation set | accuracy: 33.16 vs 32.94; 63.74 vs 60.68; 72.04 vs 71.46; 80.30 vs 80.16; 85.56 vs 85.38; 90.08 vs 90.44 | Mixed |
| replace a learned GCN feature encoder with randomly initialized DGI features | DGI · Reddit | test micro F1: 93.3 vs 94.9 | Hurt |
Closest papers by meaning
- Simple and Deep Graph Convolutional Networks (ICML 2020 · similarity 0.95)
- L2-GCN: Layer-Wise and Learned Efficient Training of Graph Convolutional Networks (CVPR 2020 · similarity 0.94)
- Dissecting the Diffusion Process in Linear Graph Convolutional Networks (NeurIPS 2021 · similarity 0.93)
- Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning (AAAI 2018 · similarity 0.93)
- On Provable Benefits of Depth in Training Graph Convolutional Networks (NeurIPS 2021 · similarity 0.92)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.90)
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.