Research map / Label-efficient and robust learning / OOD and anomaly detection
Redundancy-Aware 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 RedOUT, a test-time graph OOD detection framework that builds a structural-entropy-minimizing coding tree as an essential view and optimizes a redundancy-aware graph information bottleneck (ReGIB) via contrastive and conditional-MI losses, without modifying the pre-trained GNN.
Problem
Graph OOD detection at test time suffers because structural redundancy in test graphs induces semantic shifts, so pre-trained GNNs cannot reliably separate ID from OOD graphs.
Approach
Decompose the graph information bottleneck into essential information and redundancy (ReGIB), instantiate the essential view as a coding tree obtained by minimizing structural entropy, and calibrate OOD scores with a tree encoder while keeping the pre-trained model frozen.
Key results
- RedOUT achieves average AUC 87.46 across 10 OOD pairs, about 6.7% above the second-best GOOD-D (80.73), with average rank 1.3.
- On the ClinTox/LIPO pair RedOUT reaches 86.56 AUC, surpassing the best competitor by 17.3 points.
- Anomaly detection on ENZYMES, DHFR, BZR, IMDB-B and REDDIT-B shows RedOUT best on all five datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RedOUT instead of the strongest listed baseline | Graph OOD detection · 10 OOD dataset pairs | average AUC (%): +6.73 | Helped |
| Use RedOUT instead of GOOD-D | Graph OOD detection · ClinTox/LIPO | AUC (%): 86.56±0.76 vs 69.18±3.61; +17.38 | Helped |
| Use RedOUT instead of InfoGraph-MD | Graph OOD detection · IMDB-M/IMDB-B | AUC (%): 79.54±0.72 vs 81.38±1.14 | Hurt |
| Add contrastive loss L_Cl (ablation) | RedOUT · OOD detection, 10 dataset pairs | AUC (%): 93.04±1.83 -> 95.06±0.54; 86.38±4.72 -> 94.45±1.66 | Helped |
| Add conditional redundancy-eliminated loss L_CRI (ablation) | RedOUT · OOD detection, 10 dataset pairs | AUC (%): 92.73±1.71 -> 95.06±0.54; 89.40±2.84 -> 94.45±1.66 | Helped |
Surprising findings
- RedOUT does not achieve the best result on the IMDB-B/IMDB-M pair, which the authors attribute to similar structural patterns across these social network datasets.
- Ablation shows removing the contrastive loss L_Cl or the conditional redundancy loss L_CRI each reduces average performance, with L_Cl-only ablation giving higher AUC on some pairs than the full model is not observed.
Limitations
- Only graph-level tasks are considered; node-level test-time OOD detection is left for future work.
- Performance is sensitive to the coding tree height k and the trade-off hyperparameter lambda.
- Weak on social network datasets such as IMDB-B/IMDB-M where structure does not separate classes.
Experimental setup
Datasets: TUDataset OGB
Benchmarks: BZR/COX2 PTC-MR/MUTAG AIDS/DHFR ENZYMES/PROTEIN IMDB-M/IMDB-B Tox21/SIDER FreeSolv/ToxCast BBBP/BACE ClinTox/LIPO Esol/MUV
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 OCGIN GLocalKD
Builds on: GOODAT GOOD-D graph information bottleneck SEGO InfoNCE
Closest papers by meaning
- Structural Entropy Guided Unsupervised Graph Out-Of-Distribution Detection (AAAI 2025 · similarity 0.92)
- Graph Out-of-Distribution Detection Goes Neighborhood Shaping (ICML 2024 · similarity 0.89)
- GOODAT: Towards Test-Time Graph Out-of-Distribution Detection (AAAI 2024 · similarity 0.89)
- Bridging OOD Detection and Generalization: A Graph-Theoretic View (NeurIPS 2024 · similarity 0.89)
- Unifying Unsupervised Graph-Level Anomaly Detection and Out-of-Distribution Detection: A Benchmark (ICLR 2025 · similarity 0.88)
- Revisiting Score Propagation in Graph Out-of-Distribution Detection (NeurIPS 2024 · 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.