AutoML for time-series outliers, from smoothing to scoring

What it does TODS is a Python-based AutoML system for finding outliers in multivariate time-series data. It bundles preprocessing, time-series smoothing, feature extraction from time and frequency domains, and detection algorithms into modular components that can be assembled into end-to-end pipelines. The system targets three distinct flavors of anomaly: individual point outliers, anomalous subsequences (pattern-wise), and outlier sets across entire systems (system-wise).

The interesting bit Rather than treating detection as a single-model problem, TODS treats the whole workflow as a combinatorial search space. Its AutoML layer—exemplified by a brute-force search backend—tries to stitch together the best sequence of modules for a given dataset and metric, effectively treating pipeline architecture as a hyperparameter. It also incorporates a human-in-the-loop “reinforcement module,” though the README is light on exactly how that calibration works.

Key highlights

• Covers the full stack: preprocessing, smoothing, feature extraction, and detection.
• Reuses the full catalog of point-wise detectors from PyOD and adds pattern-wise methods like DeepLog and Telemanon.
• Supports system-wise detection via ensemble algorithms.
• Includes ready-made Colab notebooks for general usage, fraud detection, and blockchain analysis.
• Published at AAAI 2021 by Rice University’s DATA Lab.

Caveats

• Installation requires several system-level dependencies (libssl-dev, ffmpeg, etc.), so it is not a pure pip experience.
• The AutoML search example uses a BruteForceSearch backend, which suggests the module search space may be manageable but is not described in detail.
• The human-in-the-loop reinforcement module is mentioned as a feature, but the README does not clarify its interface or mechanism.

Verdict Data scientists and engineers who need to detect anomalies in multivariate time series without manually orchestrating every preprocessing and modeling step should take a look. If you already have a favorite, finely tuned single-model detector and don’t need pipeline search, it may be overkill.