Research map / Label-efficient and robust learning / OOD and anomaly detection
Harnessing Feature Resonance under Arbitrary Target Alignment for Out-of-Distribution Node 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 RSL, a label-free graph OOD node detector that scores nodes by the one-step representation movement of unlabeled nodes when known ID nodes are aligned to a random target, then trains a classifier with SGLD-generated synthetic OOD nodes.
Problem
Out-of-distribution node detection in graphs without multi-class labels or a well-defined pretext classification task.
Approach
Feature Resonance: when known ID node representations are optimized toward an arbitrary fixed target, unlabeled ID nodes move more per training step than OOD nodes. A micro-level proxy (one-step representation change, norm of Δh) is used to score nodes, with the resonance period picked on a binary ID/OOD validation set, and the scores seed Langevin-based synthetic OOD nodes for training a classifi
Key results
- RSL reaches FPR@95 of 39.58% ... on Squirrel/WikiCS/YelpChi etc.; on WikiCS RSL attains AUROC 84.01 vs 78.46 for GRASP (label-based) with label-free setting.
- Resonance score alone (RSL w/o classifier) reduces FPR95 by an average of 9.70% vs SOTA, per the paper.
- Paper reports an average 15.20% FPR95 reduction over state-of-the-art in the introduction.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the trained OOD classifier to the resonance-score filter (ablation) | RSL · Squirrel, WikiCS, YelpChi, Amazon, and Reddit | AUROC: 61.52 -> 64.12; 79.15 -> 84.01; 65.42 -> 66.11; 87.43 -> 90.03; 52.37 -> 64.83 | Helped |
| add synthetic OOD nodes (ablation) | RSL · Squirrel, WikiCS, YelpChi, Amazon, and Reddit | AUROC: 60.46 -> 64.12; 81.21 -> 84.01; 65.15 -> 66.11; 87.81 -> 90.03; 61.36 -> 64.83 | Helped |
| align ID representations to a single random vector instead of true multi-label targets (ablation) | RSL w/o classifier · Squirrel and WikiCS | AUROC: 61.63 -> 61.52; 71.03 -> 79.15 | Mixed |
| increase the available label proportion for supervised-contrastive representation pretraining (ablation) | RSL · WikiCS | AUROC: 84.01 -> 89.07 | Helped |
| increase graph homophily by replacing heterophilous edges with homophilous edges (ablation) | RSL · Citeseer | AUROC: 90.44 -> 98.23 | Helped |
| use RSL instead of the strongest listed baseline | OOD node detection · Squirrel, WikiCS, YelpChi, Amazon, Reddit, Cora, Citeseer, Pubmed, and Chameleon | AUROC: 64.15 -> 64.12; 78.46 -> 84.01; 62.04 -> 66.11; 86.57 -> 90.03; 63.32 -> 64.83; 93.50 -> 94.14; 89.75 -> 90.44; 88.43 -> | Mixed |
Surprising findings
- Feature resonance still appears when all known ID nodes are aligned to a single random fixed target vector, without any labels.
- Heterophilous graphs show larger gains, since RSL does not rely on homophily.
Limitations
- The resonance period must be selected using a binary ID/OOD validation set, which is a form of OOD supervision.
- Macroscopic trajectory-based scores are unreliable due to error accumulation, so the method depends on a micro-level measure.
- Theoretical bounds rely on strong assumptions (smoothness, bounded gradients, small learning rate).
Experimental setup
Datasets: Squirrel WikiCS Cora Citeseer Pubmed Chameleon YelpChi Amazon Reddit ENZYMES PROTEINS ClinTox
Benchmarks: Squirrel WikiCS YelpChi Amazon Reddit Cora Citeseer Pubmed Chameleon ENZYMES PROTEINS ClinTox
Metrics: AUROC AUPR FPR@95
Baselines: LOF-KNN MLPAE GCNAE GAAN DOMINANT ANOMALOUS SL-GAD GOAD NeuTral AD MSP Energy GKDE
Builds on: EnergyDef GNNSafe SSD GRASP
Closest papers by meaning
- GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation (ICLR 2025 · similarity 0.88)
- Decoupled Graph Energy-based Model for Node Out-of-Distribution Detection on Heterophilic Graphs (ICLR 2025 · similarity 0.88)
- Graph Out-of-Distribution Detection Goes Neighborhood Shaping (ICML 2024 · similarity 0.87)
- Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.87)
- ML-GOOD: Towards Multi-Label Graph Out-Of-Distribution Detection (AAAI 2025 · similarity 0.87)
- Bridging OOD Detection and Generalization: A Graph-Theoretic View (NeurIPS 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.