Research map / Label-efficient and robust learning / OOD and anomaly detection
MOOD: Benchmarking Post-Hoc OOD Detection for Materials Property Prediction
Shirley X Yu, Amey Pasarkar, Adji Bousso Dieng
Read the paperCheck your idea against 137,217 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.
Introduces MOOD, a benchmark testing whether post-hoc out-of-distribution detectors built for image classification transfer to graph neural networks for materials property prediction.
Problem
Post-hoc OOD detection (flagging OOD inputs without retraining) has been validated almost only on image classification, yet is needed for high-stakes scientific regression such as materials property prediction with GNNs.
Approach
A benchmark that evaluates existing post-hoc OOD detectors on GNNs for materials property prediction, rather than proposing a new detector.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply vision post-hoc OOD detectors to frozen materials GNN embeddings | 15 post-hoc detectors · MOOD (264 settings) | mean AUROC: best 0.531 (Cosine), chance 0.5 | No clear effect |
| use supervised linear probe instead of unsupervised detectors (ablation) | frozen GNN embeddings · dielectric | AUROC: 0.522 -> 0.599 | Improved |
| use supervised linear probe instead of unsupervised detectors (ablation) | frozen GNN embeddings · log_gvrh | AUROC: 0.524 -> 0.600 | Improved |
| use supervised linear probe instead of unsupervised detectors (ablation) | frozen GNN embeddings · log_kvrh, perovskites, phonons | AUROC: probe 0.494 / 0.500 / 0.499 | No clear effect |
| replace linear probe with MLP probe (ablation) | frozen GNN embeddings · MOOD encoder-bottlenecked tasks | probe AUROC: mean |delta| = 0.023, r = 0.971 | No clear effect |
| restrict to settings where OOD error exceeds ID error (ablation) | Cosine detector · MOOD degraded subset (N=153) | OLS slope of AUROC on MAE gap: beta +0.322 -> +0.598; R2 0.27 -> 0.45 | Improved |
| use isolation or boundary detectors instead of distance/density (ablation) | post-hoc OOD detectors · MOOD degraded subset | slope of AUROC vs MAE gap: IForest +0.185, ViM +0.230 vs Cosine +0.598 | Hurt |
| use distance-based family instead of other families (ablation) | post-hoc OOD detectors · MOOD (264 settings) | win rate: 40% vs 20% random baseline | Improved |
| abstain on samples flagged by detector (risk-coverage) | post-hoc detectors · MOOD seed 42 encoders | selective prediction effectiveness: SPE 0.14 to 0.37 (oracle = 1) | Mixed |
Experimental setup
Benchmarks: MOOD
Closest papers by meaning
- GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation (ICLR 2025 · similarity 0.87)
- ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms (ICLR 2024 · similarity 0.87)
- Image Background Serves as Good Proxy for Out-of-distribution Data (ICLR 2024 · similarity 0.87)
- Detecting Out-of-Distribution Through the Lens of Neural Collapse (CVPR 2025 · similarity 0.86)
- Bridging OOD Detection and Generalization: A Graph-Theoretic View (NeurIPS 2024 · similarity 0.86)
- Structural Entropy Guided Unsupervised Graph Out-Of-Distribution Detection (AAAI 2025 · 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.
