TonyBagnall/transformers

A scikit-learn compatible Python toolbox for supervised learning with time-series/panel data

sktime

A scikit-learn compatible Python toolbox for learning with time series and panel data. Eventually, we would like to support:

• Time series classification and regression,
• Classical forecasting,
• Supervised/panel forecasting,
• Time series segmentation,
• Time-to-event and event risk modelling,
• Unsupervised tasks such as motif discovery and anomaly detection, and diagnostic visualization,
• On-line and streaming tasks, e.g., in variation of the above.

The package is under active development. Development takes place in the sktime repository on Github.

Currently, modular modelling workflows for supervised learning with time series have been implemented. As next steps, we will move to forecasting and integration of a modified pysf interface for forecasting and supervised forecasting.

Installation

The package is currently not feature stable, and thus not available directly via PyPI. In the interim, please follow these steps to install the development version:

1. Download the repository if you have not done so already: git clone https://github.com/alan-turing-institute/sktime.git
2. Move into the root directory: cd sktime
3. Make sure your local version is up-to-date: git pull
4. Optionally, activate destination environment for package, e.g. with conda: conda activate <env>
5. Install package: pip install .

Overview

High-level interface

There are numerous different time series data related learning tasks, for example

• Time series classification and regression,
• Classical forecasting,
• Supervised/panel forecasting,
• Time series segmentation.

The sktime high-level interface aims to create a unified interface for these different learning tasks (partially inspired by the APIs of mlr and openML) through the following two objects:

• Task object that encapsulates meta-data from a dataset and the necessary information about the particular supervised learning task, e.g. the instructions on how to derive the target/labels for classification from the data,
• Strategy objects that wrap low-level estimators and allows to use fit and predict methods using data and a task object.

Low-level interface

The low-level interface extends the standard scikit-learn API to handle time series and panel data. Currently, the package implements:

• Various state-of-the-art approaches to supervised learning with time series features,
• Transformation of time series, including series-to-series transforms (e.g. Fourier transform), series-to-primitives transforms aka feature extractors, (e.g. mean, variance), sub-divided into fittables (on table) and row-wise applicates,
• Pipelining, allowing to chain multiple transformers with a final estimator,
• Meta-learning strategies including tuning and ensembling, accepting pipelines as the base estimator,
• Off-shelf composites strategies, such as a fully customisable random forest for time-series classification, with interval segmentation and feature extraction.

Documentation

The full API documentation and an introduction can be found here. Tutorial notebooks for currently stable functionality are here.

Development road map

1. Functionality for the advanced time series tasks. For (supervised) forecasting, integration of a modified pysf interface. For time-to-event and event risk modell, integration of an adapted pysf interface.
2. Extension of high-level interface to classical and supervised/panel forecasting, to include reduction strategies in which forecasting or supervised forecasting tasks are reduced to tasks that can be solved with classical supervised learning algorithms or time series classification/regression,
3. Integration of algorithms for classical forecasting (e.g. ARIMA), deep learning strategies, and third-party feature extraction tools,
4. Design and implementation of specialised data-container for efficient handling of time series/panel data in a supervised learning workflow and separation of time series meta-data, re-utilising existing data-containers whenever possible,
5. Automated benchmarking functionality including orchestration of experiments and post-hoc evaluation methods, based on the mlaut design.

Contributors

Former and current active contributors are as follows.

Project management: Jason Lines (@jasonlines), Franz J Kiraly (@fkiraly)

Design: Anthony Bagnall, Sajaysurya Ganesh (@sajaysurya), Jason Lines (@jasonlines), Viktor Kazakov (@viktorkaz), Franz J Kiraly (@fkiraly), Markus Löning (@mloning)

Coding: Sajaysurya Ganesh (@sajaysurya), Jason Lines (@jasonlines), Viktor Kazakov (@viktorkaz), Markus Löning (@mloning)

We are actively looking for contributors. Please contact @fkiraly or @jasonlines for volunteering or information on paid opportunities, or simply raise an issue in the tracker.