Research map / Label-efficient and robust learning / OOD and anomaly detection

SetAD: Semi-Supervised Anomaly Learning in Contextual Sets

NeurIPS 2026 · Label-efficient and robust learning › OOD and anomaly detection · approach: Anomaly detection with contaminated data · 0 citations (OpenAlex)

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.

Reframes semi-supervised anomaly detection as set-level graded learning: an attention-based set encoder regresses the number of labeled anomalies in sampled sets, and inference averages context-normalized scores over random contexts.

Problem

Existing semi-supervised anomaly detection methods score individual points or pairs, missing the contextual, group-level nature of anomalies and underusing supervision from labeled anomalies.

Approach

Train a self-attention set encoder with MAE regression to predict the count of known anomalies in randomly composed sets (size k=8), then score test points by averaging their deviation from reference normals across multiple random contexts (context-calibrated anomaly scoring).

Key results

• SetAD achieves average AUC-ROC 0.9405 vs 0.9205 for the best baseline REPEN, ranking first on 7 of 10 datasets.
• SetAD achieves average AUC-PR 0.7008 vs 0.6614 for DevNet, ranking first on 9 of 10 datasets.
• Ablation removing context-calibrated scoring lowers average AUC-ROC from 0.9405 to 0.9353 and AUC-PR from 0.7008 to 0.6943.

Surprising findings

• On Cardiotocography, the variant without context-calibrated scoring slightly outperforms the full model.
• On Celeba, performance peaks at set size k=3 and then declines, unlike other datasets.

Limitations

• Inference cost scales with number of contexts (n_C=60) and reference points, and set-encoder runs per test point.
• Training labels for sets assume unlabeled pool is mostly normal; contamination introduces label noise.
• Evaluated only on tabular benchmarks with the baselines' default hyperparameters.

Experimental setup

Datasets: Cardiotocography Mammography SpamBase Mnist Shuttle Celeba Cover Campaign Fraud Census

Benchmarks: Cardiotocography Mammography SpamBase Mnist Shuttle Celeba Cover Campaign Fraud Census

Metrics: AUC-ROC AUC-PR

Baselines: REPEN DeepSAD GANomaly DevNet FEAWAD PReNet SSIF

Builds on: DevNet PReNet REPEN DeepSAD

Closest papers by meaning

• Anomaly Heterogeneity Learning for Open-set Supervised Anomaly Detection (CVPR 2024 · similarity 0.89)
• ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.86)
• Open-Set Graph Anomaly Detection via Normal Structure Regularisation (ICLR 2025 · similarity 0.86)
• Beyond Normal References: Discriminative Few-Shot Anomaly Detection (NeurIPS 2026 · similarity 0.86)
• Long-Tailed Anomaly Detection with Learnable Class Names (CVPR 2024 · similarity 0.85)
• Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.85)