Research map / Graph learning / GNN architectures
Are GATs Out of Balance?
Nimrah Mustafa, Aleksandar Bojchevski, Rebekka Burkholz
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 balanced LL-orthogonal initialization instead of Xavier | GAT, 10 layers · Cora (SGD) | test accuracy: 39.3% -> 80.2% | Improved |
| use balanced LL-orthogonal initialization instead of Xavier | GAT, 5 layers · Cora (SGD) | test accuracy: 75.5% -> 79.9% | Improved |
| use balanced LL-orthogonal initialization instead of Xavier | GAT, 40 layers · Citeseer (SGD) | test accuracy: 18.72 ± 1.15 -> 63.40 ± 1.43 | Improved |
| use balanced Xavier initialization instead of Xavier | GAT, 10 layers · Citeseer (SGD) | test accuracy: 24.70 ± 8.90 -> 64.13 ± 1.57 | Improved |
| increase width from 64 to 512 with Xavier init (ablation) | GAT, 10 layers · Cora (SGD) | test accuracy: 24.7% -> 68.7% | Improved |
| increase width from 64 to 512 with Bal_O init (ablation) | GAT, 10 layers · Cora (SGD) | test accuracy: 79.7% -> 80.9% | Improved |
| replace ReLU with ELU under Bal_O (ablation) | GAT, 10 layers Bal_O · Cora (SGD) | test accuracy: 79.46 ± 1.34 -> 47.76 ± 7.39 | Hurt |
| add dropout 0.6 to balanced Bal_O (ablation) | GAT, 10 layers · Cora (SGD) | test accuracy: 79.46 ± 1.34 -> 44.40 ± 1.84 | Hurt |
| add dropout 0.6 to Xavier GAT with Adam (ablation) | GAT, 10 layers Xavier · Cora (Adam) | test accuracy: 70.86 ± 3.99 -> 32.48 ± 6.99 | Hurt |
| add weight decay 0.0005 to Xavier GAT with SGD (ablation) | GAT, 10 layers Xavier · Cora (SGD) | test accuracy: 25.48 ± 18.13 -> 18.78 ± 11.96 | No clear effect |
| use 8 attention heads instead of 1 with Bal_O (ablation) | GAT, 10 layers Bal_O · Cora (SGD) | test accuracy: 79.46 ± 1.34 -> 79.06 ± 0.73 | No clear effect |
| turn off weight sharing with Bal_O (ablation) | GAT, 10 layers Bal_O · Cora (SGD) | test accuracy: 79.46 ± 1.34 -> 79.06 ± 0.32 | No clear effect |
Closest papers by meaning
- A Dynamical Systems-Inspired Pruning Strategy for Addressing Oversmoothing in Graph Attention Networks (ICML 2025 · similarity 0.89)
- Dynamic Rescaling for Training GNNs (NeurIPS 2024 · similarity 0.88)
- Mixture of Scope Experts at Test: Generalizing Deeper Graph Neural Networks with Shallow Variants (NeurIPS 2025 · similarity 0.87)
- How Graph Neural Networks Learn: Lessons from Training Dynamics (ICML 2024 · similarity 0.87)
- GATE: How to Keep Out Intrusive Neighbors (ICML 2024 · similarity 0.87)
- On Vanishing Gradients, Over-Smoothing, and Over-Squashing in GNNs: Bridging Recurrent and Graph Learning (NeurIPS 2025 · 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.
