tsseg.algorithms.window package

Window — sliding-window discrepancy change point detection.

Description

Window-based detection slides two adjacent windows along the signal and computes a discrepancy measure \(d(\cdot,\cdot)\) derived from the cost function:

\[d(y_{u..v},\,y_{v..w}) = c(y_{u..w}) - c(y_{u..v}) - c(y_{v..w})\]

When both windows fall inside the same segment their statistical properties are similar and the discrepancy is low. When they straddle a change point the discrepancy spikes. A sequential peak search on the discrepancy curve produces the final change points.

Complexity is \(O(n\,w)\) where w is the window width — making Window one of the fastest methods. The jump parameter further speeds up prediction at the expense of positional precision.

Type: change point detection

Supervision: unsupervised or semi-supervised

Scope: univariate and multivariate

Complexity: \(O(n\,w)\)

Parameters

Name	Type	Default	Description
width	int	100	Width of each sliding window (in samples).
n_cps	int / None	None	Number of change points.
pen	float / None	None	Penalty threshold.
epsilon	float / None	None	Reconstruction-error tolerance.
model	str	"l2"	Ruptures cost model.
min_size	int	2	Minimum segment length.
jump	int	5	Sub-sampling factor for candidates.
cost_params	dict / None	None	Extra keyword arguments for the cost factory.
axis	int	0	Time axis.

Usage

from tsseg.algorithms import WindowDetector

detector = WindowDetector(width=50, model="l2", n_cps=3)
labels = detector.fit_predict(X)

# Penalty-based stopping
import numpy as np
detector = WindowDetector(width=50, model="l2", pen=np.log(n) * d * sigma**2)
labels = detector.fit_predict(X)

Implementation: Vendored from ruptures v1.1.8. BSD 2-Clause.

Reference: Basseville & Nikiforov (1993), Detection of Abrupt Changes, Prentice Hall.

Submodules

tsseg.algorithms.window.detector module

Module contents