Research map / Graph learning / Graph self-supervised learning
Unifying Unsupervised Graph-Level Anomaly Detection and Out-of-Distribution Detection: A Benchmark
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.
Presents UB-GOLD, a unified benchmark that combines unsupervised graph-level anomaly detection (GLAD) and graph-level OOD detection (GLOD) across 35 datasets and 18 methods, with analyses of effectiveness, near/far-OOD sensitivity, robustness to training contamination, and efficiency.
Problem
Unsupervised graph-level anomaly detection and graph-level out-of-distribution detection have been studied separately with distinct evaluation setups, hindering cross-application and fair comparison of methods.
Approach
Unify GLAD and GLOD under a generalized graph-level OOD detection formulation and evaluate 18 unsupervised methods (two-step graph kernel/SSL with detectors and end-to-end GNN methods) on 35 datasets spanning intrinsic anomaly, class-based anomaly, inter-dataset shift and intra-dataset shift scenarios.
Key results
- SIGNET achieves the best average AUROC (75.86) and average rank (3.37) among the 18 methods in the AUROC comparison table.
- No single method is best on more than 12 of 35 datasets, e.g., SIGNET scores 64.56 vs 71.42 SOTA on HSE.
- End-to-end methods have average ranks below 10 while all two-step methods rank above 10.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use end-to-end methods instead of two-step methods | GLAD/GLOD detection methods · UB-GOLD 35 datasets | average rank | Helped |
| Evaluate on far-OOD rather than near-OOD samples | GLAD/GLOD detection methods · Near-OOD and far-OOD settings | AUROC | Helped |
| Increase OOD contamination in the ID training set | GLAD/GLOD detection methods · PROTEINS, BZR, and COX2 | AUROC | Hurt |
| Use GLADC under training-set contamination | GLAD/GLOD detection methods · All datasets under perturbation | AUROC | No clear effect |
| Use CVTGAD under training-set contamination | GLAD/GLOD detection methods · BZR under perturbation | AUROC | No clear effect |
| Use GLAD methods for size-based distribution shifts | GLAD methods · Size-based near/far OOD | AUROC | Hurt |
| Use SIGNET instead of the strongest reported method | Graph-level anomaly and OOD detection · HSE | AUROC: 64.56 vs. 71.42 | Hurt |
Surprising findings
- Near-OOD samples are harder to detect than far-OOD samples for most end-to-end methods.
- Most methods degrade as OOD contamination in the training set rises from 0% to 30%.
Limitations
- Only unsupervised methods are included; methods requiring anomaly labels (e.g., i-GAD) and domain-specific pretrained models (e.g., PGR-MOOD) are excluded.
- Hyperparameters are selected by random search on the test set, so reported numbers are performance upper bounds rather than held-out estimates.
- No validation set with anomalies is used during training.
Experimental setup
Datasets: Tox21 COLLAB IMDB-BINARY REDDIT-BINARY ENZYMES PROTEINS DD BZR AIDS COX2 NCI1 DHFR
Benchmarks: UB-GOLD (35 datasets)
Metrics: AUROC AUPRC FPR95
Baselines: PK-SVM PK-iF WL-SVM WL-iF IG-SVM IG-iF GCL-SVM GCL-iF OCGIN GLocalKD OCGTL SIGNET
Builds on: OpenOOD GraphCL InfoGraph GOOD benchmark
Closest papers by meaning
- ARC: A Generalist Graph Anomaly Detector with In-Context Learning (NeurIPS 2024 · similarity 0.90)
- UniGAD: Unifying Multi-level Graph Anomaly Detection (NeurIPS 2024 · similarity 0.90)
- Graph Anomaly Detection as Dynamical Transport: Training-Free Scoring via Empirical Bayes (NeurIPS 2026 · similarity 0.89)
- DiffGAD: A Diffusion-based Unsupervised Graph Anomaly Detector (ICLR 2025 · similarity 0.89)
- Leveraging Diffusion Model as Pseudo-Anomalous Graph Generator for Graph-Level Anomaly Detection (ICML 2025 · similarity 0.89)
- Bridging OOD Detection and Generalization: A Graph-Theoretic View (NeurIPS 2024 · similarity 0.89)
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.