Skip to contents

Kolmogorov-Smirnov Windowing for univariate time series. The detector keeps a sliding window, compares an early sample against the most recent observations, and flags a changepoint when the two empirical distributions differ significantly.

This implementation is restricted to univariate numeric series and is intended to capture virtual drift on the signal directly, without any classifier.

Usage

hcp_kswin(window_size = 100, stat_size = 30, alpha = 0.005, data = NULL)

Arguments

window_size

Size of the sliding window.

stat_size

Size of the statistic subwindow used for the KS test.

alpha

Significance level for the KS test.

data

Optional initial window content.

Value

An hcp_kswin object.

References

  • Raab C, Heusinger M, Schleif FM (2020). Reactive Soft Prototype Computing for Concept Drift Streams. Neurocomputing.

  • Bifet A, Gavaldà R (2007). Learning from time-changing data with adaptive windowing. SIAM International Conference on Data Mining.