Research map / Graph learning / Graph self-supervised learning
Data-Augmented Curriculum 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 full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DCGAS, a graph NAS method for graph classification under distribution shifts that combines an embedding-guided discrete graph diffusion data generator with a two-factor uncertainty-based curriculum weighting strategy.
Problem
Graph neural architecture search for graph classification fails to generalize when training and test graph distributions shift, and existing methods ignore data augmentation and treat all training graphs equally.
Approach
Learns a graph-to-architecture mapping (architecture customizer over a differentiable supernet) trained on diffusion-generated graphs guided by graph embeddings, with per-graph loss weights derived from architecture-level and data-level uncertainty.
Key results
- On Spurious-Motif with b=0.7, DCGAS reaches 87.68±6.12 test accuracy vs. 65.72±17.47 for GRACES.
- On Ogbg-molbace, DCGAS achieves 81.31±1.94 ROC-AUC vs. 79.46±3.04 for GRACES.
- Ablation on Spurious-Motif (b=0.7) shows None 72.44, data generator only 82.05, curriculum only 80.38, full DCGAS 87.68.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DCGAS instead of GRACES | Graph architecture search · Spurious-Motif (b = 0.7) | accuracy: 65.72±17.47 -> 87.68±6.12 | Helped |
| Use DCGAS instead of GRACES | Graph architecture search · Spurious-Motif (b = 0.8) | accuracy: 59.57±17.37 -> 75.45±17.40 | Helped |
| Use DCGAS instead of GRACES | Graph architecture search · Spurious-Motif (b = 0.9) | accuracy: 50.94±8.14 -> 61.42±16.26 | Helped |
| Use DCGAS instead of GRACES | Graph architecture search · OGBG-MolHIV | ROC-AUC: 77.31±1.00 -> 78.04±0.71 | Helped |
| Use DCGAS instead of GRACES | Graph architecture search · OGBG-MolSIDER | ROC-AUC: 61.85±2.58 -> 63.46±1.42 | Helped |
| Use DCGAS instead of GRACES | Graph architecture search · OGBG-MolBACE | ROC-AUC: 79.46±3.04 -> 81.31±1.94 | Helped |
| Add embedding-guided data generation and curriculum weighting (ablation) | DCGAS · Spurious-Motif (b = 0.7) | accuracy: 72.44±17.13 -> 87.68±6.12 | Helped |
| Add curriculum weighting (ablation) | DCGAS with data generation · Spurious-Motif (b = 0.7) | accuracy: 82.05±5.67 -> 87.68±6.12 | Helped |
| Add embedding-guided data generation (ablation) | DCGAS with curriculum weighting · Spurious-Motif (b = 0.7) | accuracy: 80.38±8.29 -> 87.68±6.12 | Helped |
| Add embedding-guided data generation and curriculum weighting (ablation) | DCGAS · Spurious-Motif (b = 0.8) | accuracy: 61.73±20.23 -> 75.45±17.40 | Helped |
| Add curriculum weighting (ablation) | DCGAS with data generation · Spurious-Motif (b = 0.8) | accuracy: 68.25±11.32 -> 75.45±17.40 | Helped |
| Add embedding-guided data generation (ablation) | DCGAS with curriculum weighting · Spurious-Motif (b = 0.8) | accuracy: 63.99±19.60 -> 75.45±17.40 | Helped |
Surprising findings
- On Spurious-Motif b=0.9, DCGAS (61.42±16.26) is only marginally above GRACES (50.94±8.14), with large standard deviations, indicating high variance across runs.
Limitations
- Large run-to-run variance on Spurious-Motif (e.g., ±17 points at b=0.8), so gains are sometimes within noise.
- Relies on a discrete graph diffusion model trained on the training set, adding cost and complexity to the search.
Experimental setup
Datasets: Spurious-Motif Ogbg-molhiv Ogbg-molbace Ogbg-molsider
Benchmarks: Spurious-Motif Ogbg-molhiv Ogbg-molbace Ogbg-molsider
Metrics: accuracy ROC-AUC
Baselines: GCN GAT GIN SAGE GraphConv MLP ASAP DIR Random search DARTS GNAS PAS
Builds on: DiGress GRACES DARTS DIR
Closest papers by meaning
- AutoGFM: Automated Graph Foundation Model with Adaptive Architecture Customization (ICML 2025 · similarity 0.89)
- Graph Neural Architecture Search Under Distribution Shifts (ICML 2022 · similarity 0.88)
- Graph Differentiable Architecture Search with Structure Learning (NeurIPS 2021 · similarity 0.87)
- Neural Architecture Retrieval (ICLR 2024 · similarity 0.86)
- Multimodal Graph Neural Architecture Search under Distribution Shifts (AAAI 2024 · similarity 0.85)
- DCLP: Neural Architecture Predictor with Curriculum Contrastive Learning (AAAI 2024 · similarity 0.85)
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.