Research map / Time series / Anomaly detection
Position: Quo Vadis, Unsupervised Time Series Anomaly 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.
Position paper arguing that unsupervised time series anomaly detection progress is illusory due to point-adjust F1 and inconsistent benchmarks, and showing simple baselines (PCA reconstruction error, 1-NN distance, L2-norm) match or beat SOTA deep models, with distilled linear models performing similarly.
Problem
Unsupervised time series anomaly detection research relies on flawed evaluation metrics (notably point-adjust F1), inconsistent benchmarking, and insufficiently justified complex deep model designs, so reported progress may be illusory.
Approach
Shows that simple baselines (sensor range deviation, L2-norm, 1-NN distance, PCA reconstruction error, and small MLP/Transformer/MLP-Mixer/GCN-LSTM blocks) perform on par with or better than SOTA deep TAD methods under point-wise and range-wise F1, and that distilling deep models into linear models retains their performance, while random predictions score highly under point-adjust F1.
Key results
- PCA reconstruction error achieves the best point-wise F1 on SWaT (0.833), WADI-127 (0.501), WADI-112 (0.655), and SMD (0.572), outperforming listed SOTA deep models in the paper's evaluation.
- A random predictor achieves point-adjust F1 of 0.963 on SWaT and 0.783 on WADI-127, higher than most SOTA methods under F1_PA.
- Linear distilled versions of deep models keep point-wise F1 close to the originals, e.g., GDN on SWaT 0.810 original vs 0.808 linear and GCN-LSTM on SWaT 0.829 vs 0.794.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PCA reconstruction error instead of the evaluated SOTA deep-learning methods | TAD baselines · SWaT, WADI-127, WADI-112, and SMD | F1: 0.810 -> 0.833; 0.347 -> 0.501; 0.571 -> 0.655; 0.526 -> 0.572 | Helped |
| Use PCA reconstruction error instead of the strongest evaluated SOTA method | TAD baselines · SWaT, WADI-127, WADI-112, and SMD | F1_T: 0.425 -> 0.574; 0.434 -> 0.557; 0.589 -> 0.699; 0.570 -> 0.580 | Helped |
| Use PCA Error instead of MERLIN | Univariate TAD · UCR/IB-18 | F1: 0.795 -> 0.990 | Helped |
| Distill GDN to a linear model (ablation) | GDN · SWaT | F1: 0.810 -> 0.808 | No clear effect |
| Distill TranAD to a linear model (ablation) | TranAD · WADI-112 | F1: 0.511 -> 0.572 | Helped |
| Use Median-IQR normalization instead of no normalization (ablation) | PCA Error · WADI-112 | F1: 0.600 -> 0.654 | Helped |
| Use Median-IQR normalization instead of no normalization (ablation) | 1-Layer GCN-LSTM · SWaT and WADI-112 | F1: 0.770 -> 0.829; 0.592 -> 0.596 | Mixed |
| Train neural-network baselines in reconstruction mode instead of forecasting mode (ablation) | 1-Layer GCN-LSTM · SWaT, WADI-127, and WADI-112 | F1: 0.829 -> 0.770; 0.439 -> 0.498; 0.596 -> 0.577 | Mixed |
Surprising findings
- Uniformly random predictions outperform most SOTA methods under point-adjust F1.
- Distilled linear models match the performance of complex deep TAD models such as GDN and TranAD.
Limitations
- Evaluation is limited to a few benchmark datasets, some of which the authors themselves flag as flawed (SWaT, SMAP, MSL).
- Best-threshold F1 is used for main results, which the authors acknowledge is not a realistic protocol; AUPRC only in appendix.
- Published SOTA numbers and hyperparameters partly taken from original papers, with mixed re-implementation and checkpoint use (GDN on WADI-112 not retrained).
Experimental setup
Datasets: SWaT WADI-127 WADI-112 SMD UCR/Internal Bleeding (IB-16 to IB-19) SMAP MSL
Benchmarks: SWaT WADI-127 WADI-112 SMD UCR/Internal Bleeding SMAP MSL
Metrics: F1 with point-adjust point-wise F1 range-wise F1 (F1_T) AUPRC
Baselines: MERLIN DAGMM OmniAnomaly USAD GDN TranAD Anomaly Transformer LOF
Builds on: MERLIN time-series precision and recall (Tatbul et al.) range-based F1 correction (Hwang et al.) MLP-Mixer GDN
Closest papers by meaning
- The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.91)
- Detect, Explain, Interpret: An End-to-End Benchmark for Time Series Anomaly Detection, Explainability and Interpretability. (NeurIPS 2026 · similarity 0.88)
- Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection (NeurIPS 2026 · similarity 0.87)
- Can LLMs Understand Time Series Anomalies? (ICLR 2025 · similarity 0.87)
- When Model Meets New Normals: Test-Time Adaptation for Unsupervised Time-Series Anomaly Detection (AAAI 2024 · similarity 0.87)
- Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders (ICLR 2025 · similarity 0.87)
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.