Welcome to the host repository of the Turing Change Point Dataset, a set of time series specifically collected for the evaluation of change point detection algorithms on real-world data. This dataset was introduced in this paper. For the repository containing the code used for the experiments, see TCPDBench.

Useful links:

• Turing Change Point Dataset on GitHub.
• Turing Change Point Detection Benchmark
• An Evaluation of Change Point Detection Algorithms by Gertjan van den Burg and Chris Williams.
• Annotation Tool

Change point detection focuses on accurately detecting moments of abrupt change in the behavior of a time series. While many methods for change point detection exists, past research has paid little attention to the evaluation of existing algorithms on real-world data. This work introduces a benchmark study and a dataset (TCPD) that are explicitly designed for the evaluation of change point detection algorithms. We hope that our work becomes a proving ground for the evaluation and development of change point detection algorithms that work well in practice.

This repository contains the code needed to obtain the time series in the dataset. For the benchmark study, see TCPDBench. Note that work based on the dataset should cite our paper:

@article{vandenburg2020evaluation,
        title={An Evaluation of Change Point Detection Algorithms},
        author={{Van den Burg}, G. J. J. and Williams, C. K. I.},
        journal={arXiv preprint arXiv:2003.06222},
        year={2020}
}

The annotations are stored in the annotations.json file, which is the same as that used in the experiments (see here). Annotations are organised in a JSON object by dataset name and annotator id, and use 0-based indexing. See the TCPDBench repository for more information on extending the benchmark with your own methods or datasets.

Many of the time series in the dataset are included in this repository. However, due to licensing restrictions, some series can not be redistributed and need to be downloaded locally. We've added a Python script and a Makefile to make this process as easy as possible. There is also a Dockerfile to facilitate reproducibility.

To build the dataset using Docker, first build the docker image:

$ docker build -t tcpd https://github.com/alan-turing-institute/TCPD.git

then build the dataset:

$ docker run -i -t -v /path/to/where/you/want/the/dataset:/TCPD/export tcpd

To obtain the dataset, please run the following steps:

1. Clone the GitHub repository and change to the new directory:

   $ git clone https://github.com/alan-turing-institute/TCPD
   $ cd TCPD
   

2. Make sure you have Python (v3.2 or newer) installed, as well as virtualenv:

   $ pip install virtualenv
   

3. Next, use either of these steps:

   • To obtain the dataset using Make, simply run:

     $ make
     

     This command will download all remaining datasets and verify that they match the expected checksums.

   • If you don't have Make, you can obtain the dataset by manually executing the following commands:

     $ virtualenv ./venv
     $ source ./venv/bin/activate
     $ pip install -r requirements.txt
     $ python build_tcpd.py -v collect
     

     If you wish to verify the downloaded datasets you can run:

     $ python ./utils/check_checksums.py -v -c ./checksums.json -d ./datasets
     

4. It may be convenient to export all dataset files to a single directory. This can be done using Make as follows:

   $ make export
   

All datasets are stored in individual directories inside the datasets directory and each has its own README file with additional metadata and sources. The data format used is JSON and each file follows the JSON Schema provided in schema.json.

For your convenience, example code to load a dataset from the JSON format to a data frame is provided in the examples directory in the following languages:

• Python
• R

Implementations of various change point detection algorithms that use these datasets are available in TCPDBench. A script to plot the datasets and detection results from TCPDBench is also provided in utils/plot_dataset.py.

The annotations are included in the annotations.json file. They are in the format:

{
  "<dataset>": {
      "annotator_id": [
          <change point index>
          <change point index>
          ...
          ],
      ...
  },
  ...
}

where the annotator_id is a unique ID for the annotator and the change point indices are 0-based. Please also see the documentation in TCPDBench for more information about using the dataset and benchmark in your own work.

The code in this repository is licensed under the MIT license. See the LICENSE file for more details. Individual data files are often distributed under different terms, see the relevant README files for more details. Work that uses this dataset should cite our paper.

If you find any problems or have a suggestion for improvement of this repository, please let us know as it will help us make this resource better for everyone. You can open an issue on GitHub or send an email to gertjanvandenburg at gmail dot com.