Research map / Time series / Anomaly detection
CATCH: Channel-Aware Multivariate Time Series Anomaly Detection via Frequency Patching
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 CATCH, a frequency-patching reconstruction framework for multivariate time series anomaly detection that uses a learned patch-wise channel mask (between channel-independent and channel-dependent) and bi-level optimization.
Problem
Detecting heterogeneous point and subsequence anomalies in multivariate time series, where existing frequency-domain reconstruction methods miss fine-grained high-frequency details and channel correlations, and CI/CD strategies are either non-robust or lack capacity.
Approach
Patchify the FFT spectrum into frequency bands, and use a Linear-based mask generator with Gumbel-softmax Bernoulli sampling to produce patch-wise binary channel masks for masked attention, trained with clustering and regularization losses in a bi-level optimization that alternates mask-generator and model updates; anomaly scores combine time-domain error with patch-wise frequency error at point g
Key results
- CATCH reports the best Affiliated-F1 on most of the 12 real-world datasets, e.g. 0.994 on NYC and 0.896 on Genesis (Aff-F).
- Average AUC-ROC ablation: full CATCH 0.816 vs. 0.755 for channel-independent and 0.790 for channel-dependent masks on five datasets.
- Removing the frequency reconstruction loss causes the largest ablation drop (avg AUC-ROC 0.760 vs 0.816).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CATCH instead of the strongest listed baseline | CICIDS anomaly detection · CICIDS | AUC-ROC: 0.787 -> 0.795 | Helped |
| Use CATCH instead of the strongest listed baseline | SWAT anomaly detection · SWAT | AUC-ROC: 0.831 -> 0.345 | Hurt |
| Use CATCH instead of the strongest listed baseline | SMAP anomaly detection · SMAP | AUC-ROC: 0.585 -> 0.504 | Hurt |
| Replace learned channel correlation with CI, CD, or random masking (ablation) | CATCH · CICIDS, CalIt2, GECCO, MSL, SMD (Avg) | AUC-ROC: 0.816 -> 0.755; 0.790; 0.782 | Hurt |
| Add time reconstruction loss (ablation) | CATCH · CICIDS, CalIt2, GECCO, MSL, SMD (Avg) | AUC-ROC: 0.796 -> 0.816 | Helped |
| Add frequency reconstruction loss (ablation) | CATCH · CICIDS, CalIt2, GECCO, MSL, SMD (Avg) | AUC-ROC: 0.760 -> 0.816 | Helped |
| Add clustering loss (ablation) | CATCH · CICIDS, CalIt2, GECCO, MSL, SMD (Avg) | AUC-ROC: 0.798 -> 0.816 | Helped |
| Add regularization loss (ablation) | CATCH · CICIDS, CalIt2, GECCO, MSL, SMD (Avg) | AUC-ROC: 0.809 -> 0.816 | Helped |
| Add frequency patching during training (ablation) | CATCH · CICIDS, CalIt2, GECCO, MSL, SMD (Avg) | AUC-ROC: 0.781 -> 0.816 | Helped |
| Use point-aligned patch-wise frequency score instead of window frequency score (ablation) | CATCH anomaly scoring · CICIDS, CalIt2, GECCO, MSL, SMD (Avg) | AUC-ROC: 0.775 -> 0.816 | Helped |
| Add point-granularity temporal score (ablation) | CATCH anomaly scoring · CICIDS, CalIt2, GECCO, MSL, SMD (Avg) | AUC-ROC: 0.770 -> 0.816 | Helped |
| Add patch-wise frequency score (ablation) | CATCH anomaly scoring · CICIDS, CalIt2, GECCO, MSL, SMD (Avg) | AUC-ROC: 0.787 -> 0.816 | Helped |
Surprising findings
- Random channel masking performs worse than the full channel-dependent strategy in ablation.
- On SWaT, CATCH cannot effectively capture long continuous anomaly segments exceeding the look-back window, yielding low AUC-ROC (0.345).
Limitations
- Cannot effectively capture continuous anomalies longer than the look-back window (e.g. SWaT).
- Smaller patch sizes increase memory cost and patch count; best patch size is dataset-dependent.
- Hyperparameter sensitivity: score weight must be tuned (stable range 0.01-0.1 reported).
Experimental setup
Datasets: MSL SMAP PSM SMD SWaT CICIDS CalIt2 NYC Creditcard GECCO Genesis ASD
Benchmarks: MSL SMAP PSM SMD SWaT CICIDS CalIt2 NYC Creditcard GECCO Genesis ASD
Metrics: Affiliated-F1 AUC-ROC Accuracy AUC-PR Range-F1 VUS-ROC
Baselines: ModernTCN iTransformer DualTF Anomaly Transformer DCdetector TimesNet PatchTST DLinear NLinear TFAD AutoEncoder One-Class SVM
Builds on: Anomaly Transformer PatchTST iTransformer DualTF Gumbel softmax
Closest papers by meaning
- The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.86)
- Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.86)
- Channel Matters: Estimating Channel Influence for Multivariate Time Series (NeurIPS 2025 · similarity 0.86)
- TimeCHEAT: A Channel Harmony Strategy for Irregularly Sampled Multivariate Time Series Analysis (AAAI 2025 · similarity 0.85)
- COMET: Codebook-based Online-adaptive Multi-scale Embedding for Time-series Anomaly Detection (NeurIPS 2026 · similarity 0.85)
- Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.85)
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.