bclr 0.3.3

A package for Bayesian Changepoint detection via Logistic Regression.

These methods were developed in Thomas, Jauch, and Matteson (2025).

Installation

The simplest way to install the package, is to install from PyPI via the code:

pip install bclr

Alternatively, if you would like to install the source code, you can clone the repository and then install via pip (Linux/Mac):

git clone https://github.com/manilasoldier/bclr.git
cd bclr
pip install .

or run

pip install git+https://github.com/manilasoldier/bclr.git

Running the code

import bclr
import numpy as np

np.random.seed(800037)
a1 = np.random.standard_t(10, size=(320, 6))
a2 = np.random.standard_t(1.5, size=(480, 6))

# "Gaussian" type embedding
Xs = np.exp(-np.r_[a1, a2]**2)

Let's first investigate the single changepoint setup, before moving the multiple one.

First, let's specify a prior for the $\beta$ coefficients (we need to provide a mean and covariance for the single changepoint setup).

Single changepoint setup

rng1 = np.random.default_rng(800037) #for reproducibility
prior_covS = np.diag(np.repeat(3, 6))
prior_mean = np.repeat(0, 6)

bclrS = bclr.BayesCC(X = Xs, prior_mean = prior_mean, prior_cov = prior_covS, 
                     n_iter = 2000)
bclrS.fit(rng = rng1)
bclrS.transform(verbose=False)

bclrS.plot_k()

Multiple changepoint setup

Now we consider the multiple changepoint setup and the whole series $X$, where specify $J = 10$ (i.e. cps=10).

Then, we will go ahead and fit bclr to the data and estimate (predict) the changes.

Note: there is no need to specify a prior_mean for the multiple changepoint setup.

np.random.seed(199203)
b1 = np.random.randn(100, 10)
b2 = np.random.randn(80, 10) + np.broadcast_to(np.random.randn(10), (80, 10))
b3 = np.random.randn(140, 10) + np.broadcast_to(np.random.randn(10), (140, 10))
b4 = np.random.randn(80, 10) + np.broadcast_to(np.random.randn(10), (80, 10))

Xm = np.r_[b1, b2, b3, b4]

rng2 = np.random.default_rng(199203) #for reproducibility
prior_covM = np.diag(np.repeat(3, 10))
bclrM = bclr.MultiBayesCC(X = Xm, cps = 10, prior_cov = prior_covM, 
                          rng = rng2, warnings=False)
cps = bclrM.fit_predict()

Now we can look at

• where the changes are estimated to be located,
• the posterior probability of the estimates, and
• the normalized entropy of the segment distribution.

print(cps)

   Location  Posterior Probability  Normalized Entropy
0      99.0                  0.946            0.047544
1     180.0                  1.000            0.000000
2     320.0                  0.996            0.005354