tsseg.algorithms.clap package

ClaSP and CLaP — classification-based time series segmentation.

Description

This package contains two related algorithms that share core routines:

ClaSP (Classification Score Profile) is a parameter-free method for change point detection. It hierarchically splits a time series by training a binary classifier (k-NN by default) for each possible split point and selecting the point that best separates the two halves. The number of change points and the window size are both learned automatically from the data. ClaSP is fast, domain-agnostic, and requires no distributional assumptions.

CLaP (Classification Label Profile) extends ClaSP from change point detection to state detection. Instead of a k-NN classifier, CLaP uses a time series classifier (e.g. ROCKET) and cross-validates segment-labelled subsequences. Segments with high confusion are iteratively merged, producing a set of state labels that can be reused across non-contiguous time regions.

Both algorithms support three usage modes:

• Fully unsupervised — the number of segments is learned automatically.

• Guided semi-supervised — provide n_segments or n_change_points.

• Exact semi-supervised — provide pre-computed change_points.

Type: change point detection (ClaSP) / state detection (CLaP)

Supervision: unsupervised or semi-supervised

Scope: univariate and multivariate

ClaSP parameters

Name	Type	Default	Description
n_segments	str / int	"learn"	Number of segments ("learn" = auto-detect).
n_change_points	int / None	None	Exact number of change points (semi-supervised mode).
n_estimators	int	10	Number of ClaSPs in the ensemble.
window_size	str / int	"suss"	Window size or auto method ("suss", "fft", "acf").
k_neighbours	int	3	Number of nearest neighbours for ClaSP.
excl_radius	int	5	Exclusion radius (multiples of window size).
distance	str	"znormed_euclidean_distance"	Distance function for k-NN.
score	str	"roc_auc"	Scoring function for the profile.
early_stopping	bool	True	Stop early when no significant split is found.
validation	str	"significance_test"	Validation method for change point significance.
threshold	str / float	"default"	Threshold for the validation test.
n_jobs	int	-1	Number of parallel jobs.

CLaP parameters

Name	Type	Default	Description
window_size	str / int	"suss"	Window size or auto-detection method.
classifier	str	"rocket"	Time series classifier ("rocket", "knn", "svm").
merge_score	str	"cgain"	Scoring function for segment merging.
n_splits	int	5	Cross-validation folds.
n_jobs	int	1	Parallel jobs (-1 = all CPUs).
sample_size	int	1000	Samples for classifier training.
n_change_points	int / None	None	Number of change points (overrides n_segments).
change_points	list / None	None	Pre-computed change points for exact mode.
n_segments	int / None	None	Number of segments.

Usage

from tsseg.algorithms import ClaspDetector, ClapDetector

# --- ClaSP (change point detection, fully unsupervised) ---
clasp = ClaspDetector()
cps = clasp.fit_predict(X)                    # dense labels

# --- CLaP (state detection, guided) ---
clap = ClapDetector(n_segments=4, classifier="rocket")
states = clap.fit_predict(X)                  # sparse state labels

Implementation: Adapted from original ClaSP code by Arik Ermshaus. BSD 3-Clause.

Reference: Ermshaus, Schäfer & Leser (2023), ClaSP: parameter-free time series segmentation, DMKD; Ermshaus, Schäfer & Leser (2024), CLaP — State Detection from Time Series.

Submodules

tsseg.algorithms.clap.clap module

tsseg.algorithms.clap.clap_detector module

tsseg.algorithms.clap.clasp_detector module

tsseg.algorithms.clap.clasp_knn module

tsseg.algorithms.clap.distance module

tsseg.algorithms.clap.nearest_neighbour module

tsseg.algorithms.clap.scoring module

tsseg.algorithms.clap.segmentation module

tsseg.algorithms.clap.utils module

tsseg.algorithms.clap.validation module

tsseg.algorithms.clap.window_size module

Module contents