Locating Distributional Changes in Highly Dependent Time Series.

Locating distributional changes in piece-wise stationary time series with long-range dependencies

Implementations of the change-point estimation algorithms proposed in the following two papers.

• A. Khaleghi, D. Ryabko, Asymptotically Consistent Estimation of the Number of Change Points in Highly Dependent Time Series In Proceedings of the International Conference on Machine Learning, 2014.

• A. Khaleghi, D. Ryabko, Locating Changes in Highly-Dependent Data with an Unknown Number of Change-Points, In Proceedings of Neural Information Processing Systems, 2012.

written in Go and distributed in Python, R and Go.

Python

Installing

The Python package is precompiled for Windows (32/64 bit), MacOS (64 bit) and many Linux (32/64 bit) distributions. If you want to use the code on system for which there are no precompiled binaries like Raspberry Pis or Android phones head to the Compiling section in the Appendix.

If Python is installed on your system simply run the following code

pip install PyChest

How to use

To find changepoint positions in a sequence (given as either a python list or a numpy array), with a minimum distance of 0.03 * len(sequence) between changepoints which have been generated by 2 different processes simply call:

import PyChest
estimates = PyChest.find_changepoints(sequence, 0.03, 2)

estimates will now be a python list containing a list of changepoints in increasing order. The list can be empty.

If you do not know the number of generating processes you can call the list_estimator instead like following

import PyChest
estimates = PyChest.list_estimator(sequence, 0.03)

Now estimates is a list of changepoint estimates at least 0.03 * len(sequence) apart in decreasing order of a "score" associated to each estimate by the algorithm, which reflects the quality of the estimate; the list can be empty.

R

Installing

The R implementation uses the Python package which is precompiled for Windows (32/64 bit), MacOS (64 bit) and many Linux (32/64 bit) distributions. If you want to use the code on system for which there are no precompiled binaries like Raspberry Pis or Android phones head to the Compiling section in the Appendix.

You will need to have Python installed on your system in order to run the R-package.

Install the Package by running the following code in any R console:

devtools::install_github("azalk/GoChest")
library("RChest")
install_PyChest()

install_PyChest is used to install PyChest onto your local Python environment, call it to update the PyChest version R uses. You should not need to do this every time you use the R-package.

How to use

To find changepoint positions in a sequence (given as either a Matrix/Array or a Multi-element vector), with a minimum distance of 0.03 * length(sequence) between changepoints which have been generated by 2 different processes simply call:

library("RChest")
estimates <- find_changepoints(sequence, 0.03, 2L)

Now estimates will now be a Multi-element vector containing a list of changepoints at least 0.03 * length(sequence) apart in increasing order. The list can be empty.

If you do not know the number of generating processes you can call the list_estimator instead like following

library("RChest")
estimates <- list_estimator(sequence, 0.03)

Now estimates is a list of changepoint estimates at least 0.03 * length(sequence) apart in decreasing order of a "score" associated to each estimate by the algorithm, which reflects the quality of the estimate; the list can be empty.

Go

Installing

Install the Go source code by entering the following line in the console:

go get github.com/azalk/GoChest

Go will now complain that the directory layout is unexpected. That is because of the duality of it being a Python and Go package and can safely be ignored.

How to use

To find changepoint positions in a sequence (given as a slice), with a minimum distance of 0.03 * len(sequence) between changepoints which have been generated by 2 different processes simply call:

import (
    "github.com/azalk/GoChest/GoChest"
)
func main() {
    sequence := make([]float64, 0)

    // fill sequence somehow

    estimates = GoChest.FindChangepoints(sequence, 0.03, 2)
}

estimates will now be a slice containing a list of changepoints at least 0.03 * len(sequence) apart in increasing order. The slice can be empty.

If you do not know the number of generating processes you can call the ListEstimator instead like following

import (
    "github.com/azalk/GoChest/GoChest"
)

func main() {
    sequence := make([]float64, 0)

    // fill sequence somehow

    estimates = GoChest.ListEstimator(sequence, 0.03)
}

Now estimates is a list of changepoint estimates at least 0.03 * len(sequence) apart in decreasing order of a "score" associated to each estimate by the algorithm, which reflects the quality of the estimate; the list can be empty.

How to cite

If you use this package, please cite the following papers:

• A. Khaleghi, D. Ryabko, Asymptotically Consistent Estimation of the Number of Change Points in Highly Dependent Time Series In Proceedings of the International Conference on Machine Learning, 2014.

• A. Khaleghi, D. Ryabko, Locating Changes in Highly-Dependent Data with an Unknown Number of Change-Points, In Proceedings of Neural Information Processing Systems, 2012.

Appendix

Compiling the Go code

You should not need to compile the Go code yourself if you just want to use the package on Windows, MacOS or a common Unix distribution, you can find instructions how to install the code here: Python.

If you want to compile the Go code yourself you need to install Go first, then clone and enter the repository by typing the following code into a command line:

git clone https://github.com/azalk/GoChest
cd GoChest

Next you will need to actually compile the code, for Windows enter the following line in the console to do that

go build -buildmode=c-shared -o PyChestBuild/GoChest.dll CWrapper.go

And for MacOS and Linux

go build -buildmode=c-shared -o PyChestBuild/GoChest.so CWrapper.go

Your compiled c-shared library is now situated in the PyChestBuild folder. To complete the installation with simply type:

pip install .

The package will now prefer your compiled library over the precompiled libraries.

Compiling for R

If you want to use the R-package with your own compiled binary, first install the R-Package by typing the following into an R-command-line:

devtools::install_github("azalk/GoChest")

Next, compile the Go code by following the [Compiling the Go code](#Compiling the Go code) chapter. Finally, install the Python package into the virtual environment used by reticulate.

Do not run the init_RChest command as that might override your self-compiled binary, if the Python or Go code updates you will need to recompile the code.

Cross Compiling

Cross compiling, for example to make the package available on Android phones, is a little more complicated. In order to build the code you will need to enable cgo specifically and provide a c-compiler for your target platform. Your final command should look something like this:

CGO_CFLAGS="-g -O2 -w" CGO_ENABLED=1 GOOS=windows GOARCH=amd64 CC=ValidCCompilerForThePlatform go build -buildmode=c-shared -o PyChestBuild/GoChest.dll CWrapper.go

Which would compile for windows/amd64 architecture given that the CC=ValidCCompilerForThePlatform is replaced by an actual c-compiler for windows/amd64.

I recommend compiling on system wherever possible.