RuLSIF

RuLSIF compares the distributions of subsequences before and after a candidate position. It directly estimates a relative density ratio and converts the relative Pearson divergence into a change score.

Parameters

window_length : int, default 10

Length of each subsequence used as one density-estimation sample.

n_windows : int, default 50

Number of subsequences in each compared group.

lag : int, optional

Separation between comparison regions. Defaults to n_windows.

estimation_lag : int, optional

Reserved for controlling hyperparameter re-estimation. It is currently stored but not used.

scoring_step : int, default 1

Distance between evaluated positions.

n_kernels : int, default 100

Reserved kernel-count setting. It is currently stored but not passed to the estimator.

alpha : float, default 0.01

Relative-density smoothing in [0, 1). Values above zero bound the influence of very large ordinary density ratios.

symmetric : bool, default True

Sum forward and reverse density-ratio scores.

parallel : bool, default False

Compute the two symmetric directions in separate worker processes.

Methods

transform(time_series)

time_series : numpy.ndarray, shape (n_samples,)

One-dimensional signal.

Returns

score : numpy.ndarray, shape (n_samples,)

Density-change score with zero padding where a full comparison is unavailable.

Minimal Example

import numpy as np
from changepoynt.algorithms.rulsif import RuLSIF

rng = np.random.default_rng(4)
signal = np.r_[rng.normal(0, 1, 40), rng.normal(0, 2, 40)]
detector = RuLSIF(window_length=3, n_windows=5,
                  scoring_step=5, symmetric=False)
score = detector.transform(signal)

Reference: Liu et al., Change-Point Detection in Time-Series Data by Relative Density-Ratio Estimation.