johannfaouzi/pyts

pyts: a Python package for time series classification

pyts is a Python package for time series classification. It aims to make time series classification easily accessible by providing preprocessing and utility tools, and implementations of state-of-the-art algorithms. Most of these algorithms transform time series, thus pyts provides several tools to perform these transformations.

Installation

Dependencies

pyts requires:

• Python (>= 3.11, < 3.16)
• NumPy (>= 1.24.0)
• SciPy (>= 1.15.0)
• Scikit-Learn (>= 1.6.0)
• Joblib (>= 1.3.0)
• Numba (>= 0.60.0)

To run the examples Matplotlib (>= 3.7) is required.

User installation

If you already have a working installation of numpy, scipy, scikit-learn, joblib and numba, you can easily install pyts using pip

pip install pyts

or conda via the conda-forge channel

conda install -c conda-forge pyts

You can also get the latest version of pyts by cloning the repository

git clone https://github.com/johannfaouzi/pyts.git
cd pyts
pip install .

Testing

After installation, you can launch the test suite from outside the source directory using pytest:

pytest pyts

Changelog

See the changelog for a history of notable changes to pyts.

Development

The development of this package is in line with the one of the scikit-learn community. Therefore, you can refer to their Development Guide. A slight difference is the use of Numba instead of Cython for optimization.

Documentation

The section below gives some information about the implemented algorithms in pyts. For more information, please have a look at the HTML documentation available via ReadTheDocs.

Citation

If you use pyts in a scientific publication, we would appreciate citations to the following paper:

Johann Faouzi and Hicham Janati. pyts: A python package for time series classification.
Journal of Machine Learning Research, 21(46):1−6, 2020.

Bibtex entry:

@article{JMLR:v21:19-763,
  author  = {Johann Faouzi and Hicham Janati},
  title   = {pyts: A Python Package for Time Series Classification},
  journal = {Journal of Machine Learning Research},
  year    = {2020},
  volume  = {21},
  number  = {46},
  pages   = {1-6},
  url     = {http://jmlr.org/papers/v21/19-763.html}
}

Implemented features

Note: the content described in this section corresponds to the main branch (i.e., the latest version), and not the latest released version. You may have to install the latest version to use some of these features.

pyts consists of the following modules:

• approximation: This module provides implementations of algorithms that approximate time series. Implemented algorithms are Piecewise Aggregate Approximation, Symbolic Aggregate approXimation, Discrete Fourier Transform, Multiple Coefficient Binning and Symbolic Fourier Approximation.

• bag_of_words: This module provide tools to transform time series into bags of words. Implemented algorithms are WordExtractor and BagOfWords.

• classification: This module provides implementations of algorithms that can classify time series. Implemented algorithms are KNeighborsClassifier, SAXVSM, BOSSVS, LearningShapelets, TimeSeriesForest and TSBF.

• datasets: This module provides utilities to make or load toy datasets, as well as fetching datasets from the UEA & UCR Time Series Classification Repository.

• decomposition: This module provides implementations of algorithms that decompose a time series into several time series. The only implemented algorithm is Singular Spectrum Analysis.

• image: This module provides implementations of algorithms that transform time series into images. Implemented algorithms are Recurrence Plot, Gramian Angular Field and Markov Transition Field.

• metrics: This module provides implementations of metrics that are specific to time series. Implemented metrics are Dynamic Time Warping with several variants and the BOSS metric.

• multivariate: This modules provides utilities to deal with multivariate time series. Available tools are MultivariateTransformer and MultivariateClassifier to transform and classify multivariate time series using tools for univariate time series respectively, as well as JointRecurrencePlot and WEASEL+MUSE.

• preprocessing: This module provides most of the scikit-learn preprocessing tools but applied sample-wise (i.e. to each time series independently) instead of feature-wise, as well as an imputer of missing values using interpolation. More information is available at the pyts.preprocessing API documentation.

• transformation: This module provides implementations of algorithms that transform a data set of time series with shape (n_samples, n_timestamps) into a data set with shape (n_samples, n_extracted_features). Implemented algorithms are BagOfPatterns, BOSS, ShapeletTransform, WEASEL and ROCKET.

• utils: a simple module with utility functions.

License

The contents of this repository is under a BSD 3-Clause License.