Distilling Multi-Teacher Scoring Principles for Unsupervised Time Series Anomaly Detection

Key points

• Time series anomalies take diverse forms, and individual detectors encode different anomaly scoring principles that rarely cover all anomaly types.1
• Recent unsupervised time-series anomaly detectors provide complementary scoring signals, yet deploying them as an ensemble preserves the inference and maintenance cost of every individual detector.1
• AnoMix addresses this by extracting weak pairwise supervision from the training-set anomaly scores of frozen teachers.1
• A zero-shot variant further shows that the distilled ranking signal transfers to unseen datasets and can outperform large pretrained time-series foundation models for anomaly detection.1

Sources (1)

• [1]Distilling Multi-Teacher Scoring Principles for Unsupervised Time Series Anomaly Detection
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 01:42 AM

  We study whether heterogeneous teacher detectors can be distilled into a single lightweight detector in the unsupervised setting, where no anomaly labels are available to define a shared target.

  Time series anomalies take diverse forms, and individual detectors encode different anomaly scoring principles that rarely cover all anomaly types.

Extractive summary: sentences quoted from the sources.