Research map / Time series / Anomaly detection

Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing

NeurIPS 2026 · Time series › Anomaly detection · approach: Time series anomaly detection evaluation

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 CAPS, which learns anomaly-semantic codes from simulated normal-anomalous pairs (reconstruction, background consistency, counterfactual structure swapping), then conditionally diffuses residual anomaly effects onto target reference windows to build pseudo-labeled training pairs for a lightweight TCN detector.

Problem

Time series anomaly detection without target-domain anomaly labels, where surrogate or predefined synthetic anomalies provide supervision that may not match the target temporal context.

Approach

Treat ideal anomaly supervision as matched normal-anomalous outcomes under the same context; learn transferable anomaly semantics from simulated pairs via within-pair counterfactual recombination, model them with a mode-wise Gaussian prior, and realize them as context-conditioned residual diffusion (hard-routed residual MoE) on target references to produce target-matched counterparts.

Key results

• CAPS achieves the best aggregate average rank of 1.32 across four metrics, with 26 first-place and 35 top-two results out of 36 dataset-metric pairs.
• On the Affiliation-F metric CAPS scores 85.33 versus 80.30 for the strongest TCN baseline TCN-Injection.
• Ablation: removing the counterfactual loss lowers VUS-PR from 46.82 to 43.11 and shuffled residuals lowers Affiliation-F to 79.27.

Surprising findings

• CAPS outperforms the label-supervised TCN-Supervised reference in aggregate, though the authors note this is not an upper bound on supervised detection.

Limitations

• Performance depends on how well the simulated semantic prior covers target-relevant anomaly variation.
• Focuses on temporal anomaly effects without explicit modeling of cross-variable dependencies.
• Theoretical bounds do not account for contamination from anomalous target references labeled as normal.

Experimental setup

Datasets: IOPS MGAB NAB Power SED UCR NEK TODS YAHOO

Benchmarks: IOPS MGAB NAB Power SED UCR NEK TODS YAHOO

Metrics: Affiliation-F F1_T Standard-F1 VUS-PR

Baselines: LOF IForest OmniAnomaly USAD TranAD TimesNet FITS TCN-AE TCN-Injection TCN-Supervised TCN-Forecast

Builds on: TCN-Injection diffusion models Anomaly Transformer

Closest papers by meaning

• Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.90)
• Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability. (NeurIPS 2026 · similarity 0.89)
• The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.89)
• AR-Pro: Counterfactual Explanations for Anomaly Repair with Formal Properties (NeurIPS 2024 · similarity 0.88)
• When Model Meets New Normals: Test-Time Adaptation for Unsupervised Time-Series Anomaly Detection (AAAI 2024 · similarity 0.87)
• COMET: Codebook-based Online-adaptive Multi-scale Embedding for Time-series Anomaly Detection (NeurIPS 2026 · similarity 0.87)