streamchange

A package for segmenting streaming time series data into homogenous segments. The segmentation is based on statistical change-point detection (aka online/sequential/iterative change-point detection). Inspired by river and follows its API.

OBS

The functionality in this package is about to be moved into skchange: https://github.com/NorskRegnesentral/skchange.

Quickstart

>>> from river.stream import iter_pandas
>>> 
>>> from streamchange.amoc import UnivariateCUSUM, WindowSegmentor
>>> from streamchange.data import simulate
>>>
>>> df = simulate([0, 10, 0], [100], p=1)
>>> test = UnivariateCUSUM(minsl=1).set_default_threshold(10 * df.size)
>>> detector = WindowSegmentor(test, min_window=4, max_window=100)
>>> cpts = []
>>> for t, (x, _) in enumerate(iter_pandas(df)):
...     detector.update(x)
...     if detector.change_detected:
...         cpts.append((t, detector.changepoints))
print(cpts)
[(100, [1]), (200, [1])]

Throughout this package, a change-point is defined as the end of a segment, and it is stored relative to the time of detection by its index backwards in time. For example, if t is the index of the current observation, the changepoint in the external data set is given by t - changepoint. In the example above, 100 and 200 are the times of detection, and the changepoints are located at indices 99 and 199.

Installation

pip install git+https://github.com/NorskRegnesentral/streamchange

Dependencies

• pandas >= 1.3
• numpy >= 1.19
• numba >= 0.56

You also need Python >= 3.8.

Optional dependencies:

• Online estimation of segment parameters: river >= 0.14
• Penalty tuning: optuna >= 3.1.1

License

Streamchange is a free and open-source software licensed under the BSD 3-clause license.