BOCPD

BOCPD performs Bayesian online change-point detection for a Gaussian signal with unknown mean and known observation variance. It tracks a posterior over run length, the number of samples since the latest change.

The current implementation stores a quadratic run-length matrix, so memory and runtime grow quickly with signal length.

Parameters

run_length : int

Prior expected segment length. The implementation uses a constant hazard of 1 / run_length.

prior_mean : float, optional

Gaussian prior mean. Estimated from sliding-window means during fit() when omitted.

prior_var : float, optional

Variance of the prior over the unknown mean. Estimated during fit() when omitted.

signal_var : float, optional

Observation variance. Estimated from sliding-window variances when omitted.

change_length_threshold : int, optional

Maximum run length counted as a recent change. Defaults to int(0.1 * run_length).

Methods

fit(time_series)

Estimate any omitted prior parameters from a one-dimensional signal. The signal must contain at least run_length samples.

transform(time_series)

time_series : numpy.ndarray, shape (n_samples,)

Signal to process in chronological order. Calls fit() automatically when needed.

Returns

score : numpy.ndarray, shape (n_samples,)

Posterior probability assigned to run lengths at or below change_length_threshold.

Minimal Example

import numpy as np
from changepoynt.algorithms.bocpd import BOCPD

rng = np.random.default_rng(4)
signal = np.r_[rng.normal(0, 1, 100), rng.normal(3, 1, 100)]
score = BOCPD(run_length=40).transform(signal)

Reference: Adams and MacKay, Bayesian Online Changepoint Detection.