Time series anomaly detection benchmarks

In this page, we describe all datasets that can be loaded with dtaianomaly. The table below provides a brief summary of these datasets. In addition, dtaianomaly provides the ability to load custom time series data and synthetic data. How to do this is also described on this page.

Dataset

Size

Download

UCR

500MB

UCR Time Series Anomaly Archive

Note

You can also read custom data by implementing a custom LazyDataLoader, as described in the documentation.

Synthetic data

Within dtaianomaly, it is possible to generate synthetic data for testing purposes. First of all, it is possible to load the demonstration time series used throughout the documentation of dtaianomaly. This is done as follows:

>>> from dtaianomaly.data import demonstration_time_series
>>> X, y = demonstration_time_series()

Alternatively, it is possible to generate a synthetic sine wave with specified amplitude, frequency, noise, … via the dtaianomaly.data.make_sine_wave() method.

UCR time series anomaly archive

The UCR time series anomaly archive consists of 250 time series, which have been published to mitigate common issues in existing time series anomaly detection benchmarks [20]:

  1. Triviality: many benchmarks are easily solved without any fancy algorithms;

  2. Unrealistic anomaly density: the number of ground truth anomalies is relatively high, even though anomalies should be rare observations;

  3. Mislabeling: the ground truth labels might not be perfectly aligned with the actual anomalies in the data;

  4. Run-to-failure bias: most anomalies are located near the end of the time series.