Research map / Graph learning / GNN architectures
Test-Time Graph Neural Dataset Search With Generative Projection
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 test-time graph neural dataset search, PGNDS, which uses a GNN-guided score-based graph diffusion model to project unseen test graphs toward the training distribution, then searches and ensembles adapted graphs for inference on a frozen GIN.
Problem
Performance of well-trained GNNs degrades on test graphs under distribution shift; existing test-time adaptation either fine-tunes deployed GNN weights or does per-graph node/edge modifications that are inflexible and costly.
Approach
Learn a parameterized test-time graph distribution via a dual-conditional score-based graph diffusion guided by a frozen well-trained GNN (task pseudo-labels, structure MSE, fused Gromov-Wasserstein distance), search over all reverse-diffusion steps for the best adapted graphs, and ensemble predictions from original and adapted graphs, without updating GNN weights.
Key results
- On Ogbg-BACE, PGNDS reaches 0.8873 ROC-AUC versus 0.8338 for the second-best method (TENT).
- On QM9-alpha, PGNDS reduces RMSE from 4.3220 (ERM) to 4.2881; on QM9-A from 2.2022 to 2.1985.
- PGNDS has the best ROC-AUC on all four classification datasets in Table 1, but its runtime on 5 epochs (2.2-2.6 s) is well above GTRANS (about 0.23 s).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PGNDS instead of TENT | GIN · Ogbg-BBBP | ROC-AUC: 0.6988 -> 0.7014 | Helped |
| Use PGNDS instead of TENT | GIN · Ogbg-BACE | ROC-AUC: 0.8338 -> 0.8873 | Helped |
| Use PGNDS instead of EERM | GIN · Ogbg-ClinTox | ROC-AUC: 0.6314 -> 0.6371 | Helped |
| Use PGNDS instead of TENT | GIN · Enzymes | ROC-AUC: 0.7187 -> 0.7245 | Helped |
| Use PGNDS instead of ERM | GIN · QM9-A | RMSE: 2.2022 -> 2.1985 | Helped |
| Use PGNDS instead of ERM | GIN · QM9-B | RMSE: 0.4415 -> 0.4283 | Helped |
| Use PGNDS instead of ERM | GIN · QM9-C | RMSE: 0.3924 -> 0.3926 | Hurt |
| Use PGNDS instead of GTRANS | GIN · QM9-alpha | RMSE: 1.9289 -> 4.2881 | Hurt |
| Use PGNDS instead of ERM | GIN · Ogbg-FreeSolv | RMSE: 1.9308 -> 1.9252 | Helped |
| Add dual conditional guidance with task-specific guidance (ablation) | PGNDS · Ogbg-BBBP | ROC-AUC | Helped |
| Add structural guidance (ablation) | PGNDS · Ogbg-BBBP | ROC-AUC | Helped |
| Add graph diversity guidance (ablation) | PGNDS · Ogbg-BBBP | ROC-AUC | Helped |
Surprising findings
- GTRANS produces RMSE above 1000 on QM9-alpha, indicating unstable optimization for data-centric test-time graph transformation on regression.
- Data-centric GTRANS performs worse than ERM on Enzymes (0.5037 vs 0.7161 ROC-AUC) and on several regression tasks.
Limitations
- Runtime is higher than GTRANS, the fastest data-centric baseline.
- Regression on test graphs is harder; gains on regression tasks are small (e.g., QM9-A improves only from 2.2022 to 2.1985).
- Evaluated only on graph-level tasks with GIN as the backbone; dynamic graphs not addressed.
Experimental setup
Datasets: Enzymes Ogbg-BACE Ogbg-BBBP Ogbg-ClinTox Ogbg-FreeSolv QM9
Benchmarks: Enzymes Ogbg-BACE Ogbg-BBBP Ogbg-ClinTox Ogbg-FreeSolv QM9
Metrics: ROC-AUC RMSE
Baselines: ERM EERM TENT DropEdge FeatMask GTRANS
Builds on: score-based graph diffusion (Jo et al., 2022) GTRANS TENT
Closest papers by meaning
- Online GNN Evaluation Under Test-time Graph Distribution Shifts (ICLR 2024 · similarity 0.88)
- Test-time Adaptation on Graphs via Adaptive Subgraph-based Selection and Regularized Prototypes (ICML 2025 · similarity 0.87)
- GOODAT: Towards Test-Time Graph Out-of-Distribution Detection (AAAI 2024 · similarity 0.86)
- Can Classic GNNs Be Strong Baselines for Graph-level Tasks? Simple Architectures Meet Excellence (ICML 2025 · similarity 0.86)
- Generative Graph Pattern Machine (NeurIPS 2025 · similarity 0.85)
- How Graph Neural Networks Learn: Lessons from Training Dynamics (ICML 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.