Research map / Graph learning / GNN architectures
Improving Graph Neural Networks with Learnable Propagation Operators
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 |
|---|---|---|---|
| learn a separate propagation weight for each layer and channel instead of one global weight (ablation) | ωGCN · Cora, Citeseer, and Pubmed (semi-supervised; 64 layers) | accuracy (%): 84.4 -> 85.9; 71.8 -> 73.3; 80.3 -> 81.1 | Helped |
| add channel-wise weights to the per-layer propagation weights (ablation) | ωGCN · Cora, Citeseer, and Pubmed (semi-supervised; 64 layers) | accuracy (%): 84.8 -> 85.9; 72.6 -> 73.3; 80.8 -> 81.1 | Helped |
| learn a separate propagation weight for each layer and channel instead of one global weight (ablation) | ωGAT · Cora, Citeseer, and Pubmed (semi-supervised; 64 layers) | accuracy (%): 83.9 -> 84.8; 72.9 -> 74.0; 80.9 -> 81.8 | Helped |
| add channel-wise weights to the per-layer propagation weights (ablation) | ωGAT · Cora, Citeseer, and Pubmed (semi-supervised; 64 layers) | accuracy (%): 84.0 -> 84.8; 73.2 -> 74.0; 81.4 -> 81.8 | Helped |
| allow propagation weights to learn mixed-sign smoothing and sharpening operators rather than using the fixed GCN operator | ωGCN · Synthetic node-gradient regression | loss: 1e-1 -> 1e-12 | Helped |
| replace a fixed non-negative propagation operator with learnable weights | deep GCN and GAT · Cora, Citeseer, and Pubmed (deep node classification) | node-feature energy | Helped |
| use more ωGCN layers | ωGCN · Cora, Citeseer, and Pubmed (semi-supervised node classification) | accuracy (%): Cora: 82.6 -> 85.9; Citeseer: 71.3 -> 73.3; Pubmed: 79.7 -> 81.1 | Mixed |
| use more ωGAT layers | ωGAT · Cora, Citeseer, and Pubmed (semi-supervised node classification) | accuracy (%): Cora: 83.4 -> 84.8; Citeseer: 72.5 -> 74.0; Pubmed: 80.3 -> 81.8 | Helped |
| add learnable propagation weights to GCN | GCN · Ogbn-arxiv | accuracy (%): 71.74 -> 73.02 | Helped |
| add learnable propagation weights to GAT | GAT · Ogbn-arxiv | accuracy (%): 71.59 -> 72.76 | Helped |
Closest papers by meaning
- Why Do Attributes Propagate in Graph Convolutional Neural Networks? (AAAI 2021 · similarity 0.91)
- On Provable Benefits of Depth in Training Graph Convolutional Networks (NeurIPS 2021 · similarity 0.89)
- pathGCN: Learning General Graph Spatial Operators from Paths (ICML 2022 · similarity 0.89)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.88)
- Generalizing CNNs to graphs with learnable neighborhood quantization (NeurIPS 2024 · similarity 0.88)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.88)
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.