Research map / Graph learning / GNN architectures
Improving Breadth-Wise Backpropagation in Graph Neural Networks Helps Learning Long-Range Dependencies.
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 |
|---|---|---|---|
| Attention-based aggregation with residual message function and breadth-wise skip connections (ResRGAT) instead of baseline GNNs | ResRGAT · Conditional recall N=15 | accuracy: 95.4 vs 30.2 best baseline (RGAT) | Helped |
| Removing the skip connection over attention (transformer-style skip) (ablation) | ResRGAT · Conditional recall N=15 | accuracy: 33.3 vs 95.4 full model | Hurt |
| Adding a linear value transformation in attention (ablation) | ResRGAT · Conditional recall N=15 | accuracy: 82.8 vs 95.4 | Hurt |
| Adding SGRU node update to ResRGAT (ablation) | ResRGAT · Conditional recall N=15 | accuracy: 92.4 vs 95.4 | Hurt |
| Adding skip connection over attention is harmful; removing skip-over-attention lowers Tree Max accuracy (ablation) | ResRGAT · Tree max | graph-level accuracy: 79.0 vs 83.8 | Hurt |
| Not concatenating edge features into key vector (ablation) | ResRGAT · Tree max | node-level accuracy: 92.5 vs 93.3 | Hurt |
| Adding linear value transformation (ablation) | ResRGAT · Tree max | node-level accuracy: 92.4 vs 93.3 (graph-level 81.8 vs 83.8) | Hurt |
| Adding SGRU node update to ResRGAT (ablation) | ResRGAT · ZINC | MAE: 0.240 vs 0.327 (4 layers); 0.212 vs 0.314 (16 layers) | Helped |
| Adding SGRU node update to ResRGAT (ablation) | ResRGAT · OGBG-CODE2 | test F1: 0.1676 vs 0.1715 | Hurt |
| Depth-wise-only skip connection baselines (GGNN, GatedGCN, Transformer) vs skip-less RGCN/RGAT | baseline GNNs · Conditional recall N=15 | accuracy: GGNN 3.3, GatedGCN 4.4, Rel. TM 4.6 vs RGAT 30.2 | Hurt |
| Using SGRU instead of GRU as node update in GNNs (ablation) | GNN with node-wise update · Conditional recall N=15 | accuracy: 84.4 vs 4.6 (GRU at N=15 is 4.6 ± 1.9) | Helped |
| Attention-based ResRGAT on ZINC without SGRU vs baselines | ResRGAT · ZINC | MAE: 0.327 (4-layer) vs GatedGCN 0.375 but PNA 0.188 | Mixed |
Closest papers by meaning
- Sketch-Augmented Features Improve Learning Long-Range Dependencies in Graph Neural Networks (NeurIPS 2025 · similarity 0.89)
- Diffusing to the Top: Boost Graph Neural Networks with Minimal Hyperparameter Tuning (ICLR 2025 · similarity 0.89)
- DFA-GNN: Forward Learning of Graph Neural Networks by Direct Feedback Alignment (NeurIPS 2024 · similarity 0.88)
- Forward Learning of Graph Neural Networks (ICLR 2024 · similarity 0.88)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · similarity 0.88)
- Mixture of Scope Experts at Test: Generalizing Deeper Graph Neural Networks with Shallow Variants (NeurIPS 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.