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.