Research map / Label-efficient and robust learning / OOD and anomaly detection
ML-GOOD: Towards Multi-Label 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 ML-GOOD, an energy-based OOD detector for multi-label node classification on graphs that aggregates per-label energies (mean plus max-label term) with an energy-gap regularizer.
Problem
OOD detection for multi-label node classification on graph-structured data, which existing graph OOD methods (designed for single-label multi-class settings) do not address.
Approach
Defines a label-specific energy per label from sigmoid logits of a GCN, combines the mean label energy with the typical (max-energy) label energy into a unified node OOD score, and trains with BCE plus an energy-gap margin loss using partitioned OOD training data.
Key results
- ML-GOOD reaches 95.31 AUROC on PCG vs 90.86 for GNNSafe++ in the feature-interpolation setting (Table 2).
- On cross-dataset Din HumLoc / Dout EukLoc, ML-GOOD achieves 72.65 AUROC vs 61.32 for GNNSafe++ (Table 3).
- On PPI, ML-GOOD has AUROC 49.58 and AUPR 54.50, with FPR95 of 47.83 versus GNNSafe++'s 44.09 AUROC on that dataset, showing mixed results; the abstract claims 5.26% intra-domain and 6.54% cross-domain AUROC gains.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add label-specific energy score calculation (ablation) | ML-GOOD · OGB-Proteins | AUROC: 49.91 -> 50.71 | Helped |
| add label-specific energy score calculation (ablation) | ML-GOOD · PPI | AUROC: 50.98 -> 54.50 | Helped |
| add label-specific energy score calculation (ablation) | ML-GOOD · DBLP | AUROC: 81.45 -> 81.84 | Helped |
| add label-specific energy score calculation (ablation) | ML-GOOD · PCG | AUROC: 66.31 -> 95.31 | Helped |
| add label-specific energy score calculation (ablation) | ML-GOOD · HumLoc | AUROC: 63.98 -> 82.07 | Helped |
| use ML-GOOD instead of ODIN | OGB-Proteins OOD detection · OGB-Proteins | AUROC: 50.60 -> 50.71 | Helped |
| use ML-GOOD instead of OE | PPI OOD detection · PPI | AUROC: 50.38 -> 54.50 | Helped |
| use ML-GOOD instead of GNNSafe++ | PCG OOD detection · PCG | AUROC: 90.05 -> 95.31 | Helped |
| use ML-GOOD instead of GNNSafe++ | HumLoc OOD detection · HumLoc | AUROC: 75.10 -> 82.07 | Helped |
| use ML-GOOD instead of MSP | cross-dataset OOD detection · HumLoc → EukLoc | AUROC: 66.20 -> 72.65 | Helped |
| use ML-GOOD instead of Mahalanobis | DBLP OOD detection · DBLP | AUROC: 83.06 -> 81.84 | Hurt |
Surprising findings
- Many baselines, including MSP and ODIN, perform near chance (around 50 AUROC) on OGB-Proteins and PPI.
- On PPI, ML-GOOD does not beat GNNSafe++ on AUROC (49.58 vs 44.09 is in its favor, but FPR95 and some metrics are worse), and its FPR95 on OGB-Proteins is 94.97 versus 94.80 for GNNSafe++.
Limitations
- Memory constraints on large-scale OGB-Proteins forced random sampling of 20% of the dataset.
- Performance is sensitive to margin hyperparameters and the regularization weight lambda.
Experimental setup
Datasets: OGB-Proteins PPI DBLP PCG HumLoc EukLoc
Benchmarks: OGB-Proteins PPI DBLP PCG HumLoc HumLoc+EukLoc cross-dataset PPI+PCG cross-domain
Metrics: AUROC AUPR FPR95 IND ACC
Baselines: MSP ODIN Mahalanobis Energy Fine-Tune OE GNNSafe++
Builds on: Energy-Based Out-of-Distribution Detection GNNSafe Graph Convolutional Networks Outlier Exposure
Closest papers by meaning
- EDGE: Unknown-aware Multi-label Learning by Energy Distribution Gap Expansion (AAAI 2025 · similarity 0.88)
- Multi-Label Node Classification with Label Influence Propagation (ICLR 2025 · similarity 0.88)
- Structural Entropy Guided Unsupervised Graph Out-Of-Distribution Detection (AAAI 2025 · similarity 0.87)
- HGDL: Heterogeneous Graph Label Distribution Learning (NeurIPS 2024 · similarity 0.87)
- Harnessing Feature Resonance under Arbitrary Target Alignment for Out-of-Distribution Node Detection (NeurIPS 2025 · similarity 0.87)
- GOODAT: Towards Test-Time Graph Out-of-Distribution Detection (AAAI 2024 · similarity 0.86)
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.