tsseg.algorithms.kcpd package

KCPD — Kernel Change Point Detection.

Description

KCPD operates in a Reproducing Kernel Hilbert Space (RKHS) and detects mean-shifts in the mapped signal. The time series is implicitly mapped by a kernel function \(k(\cdot,\cdot)\) and the algorithm minimises:

\[V(t_1,\dots,t_K) = \sum_{k=0}^{K}\sum_{t=t_k}^{t_{k+1}-1} \|\phi(y_t) - \bar\mu_{t_k..t_{k+1}}\|_{\mathcal{H}}^2\]

When the number of changes K is known, the exact minimum is found by dynamic programming; otherwise a penalised formulation (PELT) is used.

Available kernels: "linear" (L2 cost), "rbf" (Gaussian), "cosine".

Type: change point detection

Supervision: unsupervised or semi-supervised

Scope: univariate and multivariate

Parameters

Name	Type	Default	Description
n_cps	int / None	None	Number of change points. None = use pen.
pen	float / None	10	Penalty for the penalised optimisation.
kernel	str	"rbf"	Kernel to use ("linear", "rbf", "cosine").
min_size	int	2	Minimum segment length.
jump	int	1	Sub-sampling factor for candidates.
cost_params	dict / None	None	Extra arguments for the kernel cost.
axis	int	0	Time axis.

Usage

from tsseg.algorithms import KCPDDetector

detector = KCPDDetector(kernel="rbf", n_cps=3)
labels = detector.fit_predict(X)

# Unknown K — use a penalty
detector = KCPDDetector(kernel="rbf", pen=10)
labels = detector.fit_predict(X)

Implementation: Vendored from ruptures v1.1.8 (C implementation). BSD 2-Clause.

Reference: Celisse, Marot, Pierre-Jean & Rigaill (2018), Computational Statistics and Data Analysis; Arlot, Celisse & Harchaoui (2019), JMLR.

Submodules

tsseg.algorithms.kcpd.detector module

Module contents