MovingWindow

MovingWindow is a deliberately simple baseline. It compares the mean, variance, or both between adjacent windows and is useful as a sanity check before applying a more complex detector.

Parameters

window_length : int

Number of samples in each adjacent comparison window.

method : {"mean", "var", "meanvar"}, default "mean"

Statistic used in the absolute-difference score.

Methods

fit(time_series)

Validate that the input is one-dimensional and longer than twice window_length.

transform(time_series)

time_series : numpy.ndarray, shape (n_samples,)

Signal to compare. fit() is called automatically when necessary.

Returns

score : numpy.ndarray, shape (n_samples,)

Absolute mean and/or variance difference with zero-padded boundaries.

Minimal Example

import numpy as np
from changepoynt.algorithms.baseline import MovingWindow

signal = np.r_[np.zeros(100), np.ones(100)]
score = MovingWindow(window_length=20, method="meanvar").transform(signal)

Background: Wu and Keogh, Current Time Series Anomaly Detection Benchmarks are Flawed.