Chapydette contains fast Cython implementations of kernel-based change-point detection algorithms and feature generation methods. There are currently two main algorithms implemented:

• The kernel change-point algorithm of Harchaoui and Cappé (2007), which detects a fixed number of change points in a sequence of observations. We also provide an implementation of the data-driven model selection procedure based on Arlot, Celisse, and Harchaoui (2019), which allows one to automatically select the number of change points.
• The change-point analysis algorithm of Harchaoui, Bach, and Moulines (2009), which detects the presence of and estimates the location of a single change point based on the Maximum Mean Discrepancy (Gretton et al., 2012).

See the full documentation for more details.

This code was written using Python 3.7 and requires Cython, Faiss, Jupyter, Matplotlib, Numba, Numpy, PyTorch, Scipy, and Scikit-learn. If you are using Anaconda you can install these in a new conda environment called chpt via

conda create -y --name=chpt python=3.7
conda activate chpt
conda install cython jupyter matplotlib nb_conda numba numpy scipy scikit-learn
conda install pytorch torchvision cpuonly -c pytorch
conda install faiss-cpu -c pytorch 

In order to compile the code you will need to have gcc installed. If you are using Ubuntu, you can install this via
sudo apt install build-essential
If you are using a Mac, you should install llvm, gcc, and libgcc:
conda install llvm gcc libgcc

To then install this package, run
python setup.py install
from the installation directory.

This code has not been tested on a Windows operating system.

The installer will check whether you have the remainder of the required dependencies. There are two optional dependencies:

• Yael http://yael.gforge.inria.fr/
• Pomegranate https://pomegranate.readthedocs.io/en/latest/index.html

If you have them installed and use them, they can greatly speed up the feature generation code.

There are two tasks that Chapydette can perform:

1. Feature generation
2. Change-point estimation

The functions for the former are contained in feature_generation.py, while the functions for the latter are in cp_estimation.py. Examples of how to use Chapydette are provided in the Jupyter notebooks in the examples directory.

Corinne Jones
Zaid Harchaoui

If you use this code in your work please cite the following paper:

C. Jones and Z. Harchaoui. End-to-End Learning for Retrospective Change-Point Estimation. In Proceedings of the IEEE International Workshop on Machine Learning for Signal Processing (to appear), 2020.

@inproceedings{JH2020,
  author    = {Jones, Corinne and Harchaoui, Zaid},
  title     = {End-to-End Learning for Retrospective Change-Point Estimation},
  booktitle = {30th {IEEE} International Workshop on Machine Learning for Signal Processing},
  year      = {2020},
}

This code has a GPLv3 license.

• S. Arlot, A. Celisse, and Z. Harchaoui, "A kernel multiple change-point algorithm via model selection," Journal of Machine Learning Research, 2019.
• A. Gretton, K.M. Borgwardt, M.J. Rasch, B. Schölkopf, and A. Smola, "A kernel two-sample test," Journal of Machine Learning Research, vol. 13, pp. 723–773, 2012.
• Z. Harchaoui and O. Cappé, "Retrospective mutiple change-point estimation with kernels," in IEEE Workshop on Statistical Signal Processing, 2007, pp. 768–772.
• Z. Harchaoui, F.R. Bach, and E. Moulines, "Kernel change-point analysis," in Advances in Neural Information Processing Systems, 2008, pp. 609–616.