Research map / Graph learning / GNN architectures
Training Graph Neural Networks with 1000 Layers
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 reversible connections to train a deep GNN | RevGNN on ogbn-proteins · ogbn-proteins | ROC-AUC: 86.91 -> 87.74 | Helped |
| Increase RevGNN capacity with a wider model | RevGNN on ogbn-proteins · ogbn-proteins | ROC-AUC: 87.74 -> 88.24 | Helped |
| Apply reversible connections to residual GNN operators | GNNs on ogbn-arxiv · ogbn-arxiv | accuracy: ResGCN 72.46 -> RevGCN 73.01; ResSAGE 72.46 -> RevSAGE 72.69; ResGEN 72.32 -> RevGEN 72.34; ResGAT 73.76 -> RevGAT 74.02 | Helped |
| Increase RevGCN channel size from 128 to 180 | RevGCN on ogbn-arxiv · ogbn-arxiv | accuracy: 73.01 -> 73.22 | Helped |
| Use reversible layers with tied weights (ablation) | RevGNN on ogbn-proteins · ogbn-proteins | ROC-AUC: 87.41 -> 85.28 | Hurt |
| Increase group size from 2 to 4 (ablation) | RevGNN on ogbn-proteins, 3–56 layers · ogbn-proteins | ROC-AUC: 85.09 -> 84.86; 85.68 -> 85.25; 86.62 -> 85.79; 86.68 -> 86.30; 86.90 -> 86.76 | Hurt |
| Increase group size from 2 to 4 (ablation) | RevGNN on ogbn-proteins, 112 layers · ogbn-proteins | ROC-AUC: 87.02 -> 87.09 | No clear effect |
| Use more inference views (ablation) | RevGNN-Deep on ogbn-proteins · ogbn-proteins | ROC-AUC: 87.29 -> 87.68 -> 87.74 | Helped |
| Use more inference views (ablation) | RevGNN-Wide on ogbn-proteins · ogbn-proteins | ROC-AUC: 87.84 -> 88.20 -> 88.24 | Helped |
| Increase inference partitions from 3 to 5 or 10 (ablation) | RevGNN-Wide on ogbn-proteins · ogbn-proteins | ROC-AUC: 87.84 -> 87.62 -> 87.23 | Hurt |
| Train RevGNN with mini-batches rather than full batches (ablation) | RevGNN on ogbn-products · ogbn-products | accuracy: 78.77 -> 82.16 | Helped |
| Use GraphNorm instead of BatchNorm (ablation) | RevGNN on ogbg-molhiv · ogbg-molhiv | ROC-AUC: 77.82 -> 78.62 | Helped |
Closest papers by meaning
- Towards Deepening Graph Neural Networks: A GNTK-based Optimization Perspective (ICLR 2022 · similarity 0.89)
- Fast and Deep Graph Neural Networks (AAAI 2020 · similarity 0.89)
- InvGNN: Learning Invertible Node Representations on Graphs (ICML 2026 · similarity 0.88)
- Efficient Learning of Linear Graph Neural Networks via Node Subsampling (NeurIPS 2023 · similarity 0.88)
- MuseGNN: Forming Scalable, Convergent GNN Layers that Minimize a Sampling-Based Energy (ICLR 2025 · similarity 0.87)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · 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.