Research map / Graph learning / Graph self-supervised learning
Graph Neural Architecture Search Under Distribution Shifts
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 |
|---|---|---|---|
| Add GRACES architecture customization (vs. no customization, equal operation weights) (ablation) | GRACES on Spurious-Motif · Spurious-Motif b=0.8 | accuracy: 59.57 vs 49.24 | Helped |
| Use self-supervised disentangled graph encoder instead of vanilla GCN encoder (ablation) | GRACES · Spurious-Motif b=0.8 | accuracy: 59.57 vs 51.62 | Helped |
| Use prototype-based customization instead of fully-connected layer (ablation) | GRACES · Spurious-Motif b=0.8 | accuracy: 59.57 vs 57.11 | Helped |
| Add cosine distance regularizer on prototype vectors (ablation) | GRACES · Spurious-Motif b=0.8 | accuracy: 59.57 vs 51.65 | Helped |
| Discretize architecture to one-hot operation choice (vs. continuous mixture) (ablation) | GRACES · Spurious-Motif b=0.8 | accuracy: 39.81 vs 59.57 | Hurt |
| Use continuous customized architecture with prototype-based weights instead of fixed single architecture (GRACES vs best manual GNN) | GRACES vs manual GNNs · Spurious-Motif b=0.7/0.8/0.9 | accuracy: 65.72 vs 50.75 (b=0.7); 59.57 vs 46.73 (b=0.8); 50.94 vs 46.41 (b=0.9) | Helped |
| GRACES on real-world OGBG-Mol* datasets vs best baselines | GRACES · OGBG-MolHIV, OGBG-MolSIDER, OGBG-MolBACE | ROC-AUC: +0.26 HIV, +1.21 SIDER, +0.59 BACE over best baseline | Helped |
| Adding the self-supervised disentangled encoder with customized super-network on OGBG-MolHIV (vs encoder alone) (ablation) | GRACES · OGBG-MolHIV | ROC-AUC | Helped |
| Standalone self-supervised disentangled encoder without customization (on Spurious-Motif b=0.8) (ablation) | SDGE encoder alone · Spurious-Motif b=0.8 | accuracy | No clear effect |
| Search time comparison GRACES vs DARTS | GRACES · Spurious-Motif and OGBG-Mol* | search time (s): 972S vs 872S on Spurious-Motif; 2325S vs 2049S on HIV | No clear effect |
Closest papers by meaning
- Disentangled Graph Self-supervised Learning for Out-of-Distribution Generalization (ICML 2024 · similarity 0.89)
- AutoGFM: Automated Graph Foundation Model with Adaptive Architecture Customization (ICML 2025 · similarity 0.89)
- Data-Augmented Curriculum Graph Neural Architecture Search under Distribution Shifts (AAAI 2024 · similarity 0.88)
- Multimodal Graph Neural Architecture Search under Distribution Shifts (AAAI 2024 · similarity 0.87)
- Graph Metanetworks for Processing Diverse Neural Architectures (ICLR 2024 · similarity 0.87)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 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.