changepoint 0.2.0

A Bayesian change point library

changepoint

Python bindings for important functionality of the rust library changepoint, a library for doing change point detection for steams of data.

Installation

Install via pip with

$ python3 -m pip install "changepoint"

Note: If there is no binary distribution for your OS, architecture, and Python version, you will need the Rust compiler to build the package and install a Python tool called Maturin:

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
python3 -m pip install maturin

See rustup.rs for instructions on installing Rust.

Quick Docs

By convention in these docs and examples,

import changepoint as chp

Models

Bocpd

The Bayesian change point detector, Bocpd, takes a prior distribution, aka one of

chp.BetaBernoulli
chp.NormalGamma
chp.NormalInvChiSquared
chp.NormalInvGamma
chp.NormalInvWishart
chp.PoissonGamma

Then, a Bocpd may be created:

cpd = chp.Bocpd(
    prior=chp.NormalGamma(),
    lam=12,
)

where the prior is a NormalGamma and the characteristic run length is 12.

Each step of the data stream, data, can be processed by

n = len(data)
change_point_history = np.zeros((n, n))
for i, x in enumerate(data):
    change_point_history[i, : i + 1] = cpd.step(x)

ArgpCpd

ArgpCpd has an implicit prior as it is a Gaussian Process of the form X_{i+1} = c X_{i-l-1, ..., i-1} + ε where c is the scale, X_{i-l-1, ..., i-1} is the previous vales in the sequence (l is the max-lag parameter), and ε is a white noise parameter. It behaves similarity to the Bocpd class; for example,

argp = chp.ArgpCpd(logistic_hazard_h=-2, scale=3, noise_level=0.01)
n = len(data)
change_point_history = np.zeros((n + 1, n + 1))
xs = []
ys = []
for i, x in enumerate(data):
    cps = argp.step(x)
    change_point_history[i, : len(cps)] = cps

Example

An example IPython notebook can be found here.