Research map / Graph learning / Graph self-supervised learning
GOODAT: Towards Test-Time Graph Out-of-Distribution Detection
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 GOODAT, a test-time, unsupervised, plug-and-play graph OOD detector that trains a lightweight graph masker on test graphs using three GIB-based losses, without training data or modifying the GNN.
Problem
Graph OOD detection for well-trained GNNs when training data is unavailable and the GNN cannot be modified, with no labels at test time.
Approach
A lightweight parameterized graph masker extracts informative subgraphs from each test graph under a surrogate ID label, optimized with three unsupervised graph-information-bottleneck losses (subgraph GIB, masked-graph GIB, and a copula-based distribution separating loss); the subgraph GIB loss serves as the OOD score.
Key results
- GOODAT achieves the best average rank (1.4) on the 10 graph OOD detection dataset pairs, outperforming all 12 baselines on 8 datasets.
- On graph anomaly detection, GOODAT reaches 95.50% AUC on AIDS and 77.92% on PROTEINS-full, with the best average rank (2.1) across 15 TU datasets.
- Ablation shows the full combination of L_s, L_m and L_d gives the best results on most datasets, e.g., 96.43% AUC on DHFR-style AIDS/DHFR pair.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GOODAT instead of the strongest reported baseline | Graph OOD detection with GIN · 10 graph OOD dataset pairs | AUC (%) | Mixed |
| Use GOODAT instead of the strongest reported baseline | Graph anomaly detection · 15 TU datasets | AUC (%) | Mixed |
| Add masked graph GIB loss to subgraph GIB loss (ablation) | GOODAT · BZR/COX2 | AUC (%): 39.10{\scriptstyle\pm1.59} -> 83.06{\scriptstyle\pm1.45} | Helped |
| Add graph distribution separating loss to subgraph GIB loss (ablation) | GOODAT · PTC-MR/MUTAG | AUC (%): 67.49{\scriptstyle\pm3.48} -> 67.09{\scriptstyle\pm2.58} | No clear effect |
| Add masked graph GIB loss and graph distribution separating loss to subgraph GIB loss (ablation) | GOODAT · AIDS/DHFR | AUC (%): 72.28{\scriptstyle\pm1.03} -> 70.64{\scriptstyle\pm3.28} | Hurt |
| Add graph distribution separating loss to subgraph GIB loss and masked graph GIB loss (ablation) | GOODAT · ENZYMES/PROTEIN | AUC (%): 65.67{\scriptstyle\pm2.31} -> 66.29{\scriptstyle\pm1.54} | Helped |
| Use only masked graph GIB loss instead of all three losses (ablation) | GOODAT · BZR/COX2 | AUC (%): 82.16{\scriptstyle\pm0.15} -> 73.53{\scriptstyle\pm0.04} | Hurt |
| Use only graph distribution separating loss instead of all three losses (ablation) | GOODAT · AIDS/DHFR | AUC (%): 96.43{\scriptstyle\pm0.25} -> 92.52{\scriptstyle\pm0.02} | Hurt |
Surprising findings
- Test-time training method GTrans, when its loss is used as OOD score, is weaker than GOODAT on all datasets, suggesting generalization methods do not directly transfer to OOD detection.
- Subgraph GIB loss alone gives very low AUC on some pairs (e.g., 39.10% on BZR/COX2), showing the masked-graph and separation losses are needed.
Limitations
- Performance is slightly below best on two datasets, attributed to high edge density reducing the effect of GIB subgraph compression.
- Relies on the assumption that all test graphs are ID to create surrogate labels.
- Evaluated only with a GIN encoder and on molecule, bioinformatics and social datasets.
Experimental setup
Datasets: BZR COX2 PTC-MR MUTAG AIDS DHFR ENZYMES PROTEIN IMDB-M IMDB-B Tox21 SIDER
Benchmarks: BZR/COX2 PTC-MR/MUTAG AIDS/DHFR ENZYMES/PROTEIN IMDB-M/IMDB-B Tox21/SIDER FreeSolv/ToxCast BBBP/BACE ClinTox/LIPO Esol/MUV TU graph anomaly detection (15 datasets)
Metrics: AUC average rank
Baselines: PK-LOF PK-OCSVM PK-iF WL-LOF WL-OCSVM WL-iF InfoGraph-iF InfoGraph-MD GraphCL-iF GraphCL-MD GTrans AAGOD
Builds on: graph information bottleneck AAGOD GTrans GIN
Closest papers by meaning
- GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation (ICLR 2025 · similarity 0.90)
- Redundancy-Aware Test-Time Graph Out-of-Distribution Detection (NeurIPS 2025 · similarity 0.89)
- Pruning Spurious Subgraphs for Graph Out-of-Distribution Generalization (NeurIPS 2025 · similarity 0.89)
- Bridging OOD Detection and Generalization: A Graph-Theoretic View (NeurIPS 2024 · similarity 0.89)
- Revisiting Score Propagation in Graph Out-of-Distribution Detection (NeurIPS 2024 · similarity 0.89)
- Bounded and Uniform Energy-based Out-of-distribution Detection for Graphs (ICML 2024 · 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.