changepoint_online

• A Collection of Methods for Online Changepoint Detection
• Installation
• Examples
  • Simple Univariate Gaussian Change-in-mean
  • Change in one-parameter exponential family distributions
  • Non-Parametric changepoint detection
  • Multivariate Data
  • Real-data examples
• License
• GitHub Repository
• Experimental
• Contributors
• How to Cite This Work
• References

A Collection of Methods for Online Changepoint Detection

The changepoint_online package provides efficient algorithms for detecting changes in data streams, based on the Focus algorithm. The Focus algorithm solves the CUSUM likelihood-ratio test exactly in $O(\log(n))$ time per iteration, where n represents the current iteration. The method is equivalent to running a rolling window (MOSUM) simultaneously for all sizes of windows or the Page-CUSUM for all possible values of the size of change (an infinitely dense grid).

Key Features

• Contains all Focus exponential family algorithms as well as the NPFocus algorithm for non-parametric changepoint detection.

• It’s versatile enough to be applied in scenarios where the pre-change parameter is either known or unknown.

• It is possible to apply constraints to detect specific types of changes (such as increases or decreases in parameter values).

Installation

installing from PyPI

pip install changepoint-online

installing from github with pip

python -m pip install 'git+https://github.com/grosed/changepoint_online/#egg=changepoint_online&subdirectory=python/package'

Testing the package

To run package tests:

python -m changepoint_online.tests.runtests

Examples

Simple Univariate Gaussian Change-in-mean

from changepoint_online import Focus, Gaussian
import numpy as np

# generating some data with a change at 50,000
np.random.seed(0)
Y = np.concatenate((np.random.normal(loc=0.0, scale=1.0, size=50000), np.random.normal(loc=2.0, scale=1.0, size=5000)))

# initialize a Focus Gaussian detector
detector = Focus(Gaussian())
threshold = 13.0

for y in Y:
    # update your detector sequentially with
    detector.update(y)
    if detector.statistic() >= threshold:
        break

print(detector.changepoint())

{'stopping_time': 50013, 'changepoint': 50000}

If the pre-change location is known (in case of previous training data), this can be specified with:

# initialize a Focus Gaussian detector (with pre-change location known)
detector = Focus(Gaussian(loc=0))
threshold = 13.0

for y in Y:
    # update your detector sequentially with
    detector.update(y)
    if detector.statistic() >= threshold:
        break

print(detector.changepoint())

{'stopping_time': 50013, 'changepoint': 50000}

See help(FamilyName) for distribution specific parameters, e.g. help(Gaussian).

Change in one-parameter exponential family distributions

As in the Gaussian case, we can detect:

• Poisson change-in-rate

• Gamma change-in-scale (or rate). Exponential change-in-rate implemented as a Gamma shape=1.

• Bernoulli change-in-probability

For example:

from changepoint_online import Focus, Gamma
import numpy as np 

np.random.seed(0)
Y = np.concatenate((np.random.gamma(4.0, scale=3.0, size=50000), 
                    np.random.gamma(4.0, scale=6.0, size=5000)))

# initialize a Gamma change-in-scale detector (with shape = 4)
detector = Focus(Gamma(shape=4.0))
threshold = 12.0
for y in Y:
    detector.update(y)
    if detector.statistic() >= threshold:
        break
        
print(detector.changepoint())

{'stopping_time': 50012, 'changepoint': 50000}

Non-Parametric changepoint detection

If we do not know the underlying distribution, or if the nature of the change is unkown a priori, we can then use NPFocus.

from changepoint_online import NPFocus
import numpy as np

# Define a simple Gaussian noise function
def generate_gaussian_noise(size):
    return np.random.normal(loc=0.0, scale=1.0, size=size)

# Generate mixed data with change in gamma component
gamma_1 = np.random.gamma(4.0, scale=3.0, size=5000)
gamma_2 = np.random.gamma(4.0, scale=6.0, size=5000)
gaussian_noise = generate_gaussian_noise(10000)
Y = np.concatenate((gamma_1 + gaussian_noise[:5000], gamma_2 + gaussian_noise[5000:]))

# Create and use NPFocus detector
## One needs to provide some quantiles to track the null distribuition over
quantiles = [np.quantile(Y[:100], q) for q in [0.25, 0.5, 0.75]]
## the detector can be initialised with those quantiles
detector = NPFocus(quantiles)

stat_over_time = []

for y in Y:
    detector.update(y)
    # we can sum the statistics over to get a detection
    # see  (Romano, Eckley, and Fearnhead 2024) for more details
    if np.sum(detector.statistic()) > 25:
        break


changepoint_info = detector.changepoint()
print(changepoint_info["stopping_time"])

5014

Multivariate Data

It is possible to run a multivariate analysis via MDFocus by feeding, at each iteration, a numpy array.

Gaussian Change-in-mean

from changepoint_online import MDFocus, MDGaussian
import numpy as np

np.random.seed(123)

# Define means and standard deviations for pre-change and post-change periods (independent dimensions)
mean_pre = np.array([0.0, 0.0, 5.0])
mean_post = np.array([1.0, 1.0, 4.5])

# Sample sizes for pre-change and post-change periods
size_pre = 5000
size_post = 500

# Generate pre-change data (independent samples for each dimension)
Y_pre = np.random.normal(mean_pre, size=(size_pre, 3))

# Generate post-change data (independent samples for each dimension)
Y_post = np.random.normal(mean_post, size=(size_post, 3))

# Concatenate data with a changepoint in the middle
changepoint = size_pre
Y = np.concatenate((Y_pre, Y_post))

# Initialize the Gaussian detector 
# (for change in mean known, as in univariate case, write MDGaussian(loc=mean_pre))
detector = MDFocus(MDGaussian(), pruning_params = (2, 1))
threshold = 25
for y in Y:
    detector.update(y)
    if detector.statistic() >= threshold:
        break
print(detector.changepoint())

{'stopping_time': 5014, 'changepoint': 5000}

Poisson Change-in-rate

from changepoint_online import MDFocus, MDPoisson
import numpy as np

np.random.seed(123)

# Define rates (lambda) for pre-change and post-change periods (independent dimensions)
rate_pre = np.array([1.0, 2.0, 3.0])
rate_post = np.array([2.0, 3.0, 4.0])

# Sample sizes for pre-change and post-change periods
size_pre = 5000
size_post = 500

# Generate pre-change data (independent samples for each dimension)
Y_pre = np.random.poisson(rate_pre, size=(size_pre, 3))

# Generate post-change data (independent samples for each dimension)
Y_post = np.random.poisson(rate_post, size=(size_post, 3))

# Concatenate data with a changepoint in the middle
changepoint = size_pre
Y = np.concatenate((Y_pre, Y_post))

# Initialize the Poisson detector
detector = MDFocus(MDPoisson(), pruning_params = (2, 1))
threshold = 45  # Adjust the threshold as needed for the Poisson distribution
for y in Y:
    detector.update(y)
    if detector.statistic() >= threshold:
        break
print(detector.changepoint())

{'stopping_time': 5056, 'changepoint': 5000}

Real-data examples

More examples, including real-world applications, are found in the examples folder, including:

• Change in the tails of Energy Wholesale Price. (markdown, quarto notebook)

• Constrained Spike inference in calcium imaging data (markdown, quarto notebook)

License

Copyright (C) 2023 Gaetano Romano, Daniel Grose

This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.

This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with this program. If not, see https://www.gnu.org/licenses/gpl-3.0.txt.

GitHub Repository

Source files for the packages can be found at https://github.com/grosed/changepoint_online

Experimental

The package includes the nunc algorithm for non-parametric changepoint detectino from Austin et al. (2023) as an experimental function. It is possible to call nunc as:

from changepoint_online.nunc import nunc, nunc_default_quantiles
import numpy as np

np.random.seed(1)
# Generate data with change
X1 = np.random.normal(1, 1, 3000)
X2 = np.random.normal(-2, 4, 5000)
Y = np.concatenate((X1, X2))

# Create and use NUNC detector
detector = nunc(300, 5, nunc_default_quantiles)

stat_over_time = []

for y in Y:
    detector.update(y)
    stat_over_time.append(detector.max_cost)
    if detector.statistic() > 30:
        break


changepoint_info = detector.changepoint()

print(f"Output information:\n{changepoint_info}")

Output information:
{'stopping_time': 300, 'changepoint': 284, 'max_cost': 31.492069041110202}

Contributors

• Gaetano Romano: email (Author) (Maintainer) (Creator) (Translator)

• Daniel Grose: email (Author) (Maintainer) (Creator)

• Kes Ward: email (Author)

• Austin Edward: email (Author) (Maintainer)

• Liudmila Pishchagina: email (Author)

• Guillem Rigaill: email (Author) (Thesis Advisor)

• Vincent Runge: email (Author) (Thesis Advisor)

• Paul Fearnhead: email (Author) (Thesis Advisor)

• Idris A. Eckley: email (Author) (Thesis Advisor)

How to Cite This Work

A possible BibTeX entry for this package could be:

@software{changepoint_online,
  author       = {Gaetano Romano, Daniel Grose, Kes Ward, Austin Edward, Liudmila Pishchagina, Guillem Rigaill, Vincent Runge, Paul Fearnhead, Idris A. Eckley},
  title        = {changepoint_online: A Collection of Methods for Online Changepoint Detection.},
  month        = Apr,
  year         = 2024,
  version      = {v1.2.1},
  url          = {https://https://github.com/grosed/changepoint_online}
}

For citing the methodologies:

• Gaussian FOCuS: (Romano et al. 2023)

• Other Exponential Family detectors: (Ward et al. 2024)

• Multi-dimentional FOCuS (Pishchagina et al. 2023)

• NPFocus: (Romano, Eckley, and Fearnhead 2024)

• NUNC: (Austin et al. 2023)

See references below.

References