[Paper Review] sktime: A Unified Interface for Machine Learning with Time Series

Markus Löning, Anthony Bagnall+4 more|arXiv (Cornell University)|Sep 17, 2019

Time Series Analysis and Forecasting49 references176 citations

TL;DR

sktime introduces a scikit-learn compatible Python library that unifies time series tasks under a single API, enabling reductions between forecasting, classification, and other tasks via meta-estimators and time series transformations.

ABSTRACT

We present sktime -- a new scikit-learn compatible Python library with a unified interface for machine learning with time series. Time series data gives rise to various distinct but closely related learning tasks, such as forecasting and time series classification, many of which can be solved by reducing them to related simpler tasks. We discuss the main rationale for creating a unified interface, including reduction, as well as the design of sktime's core API, supported by a clear overview of common time series tasks and reduction approaches.

Motivation & Objective

• Motivate the creation of a unified API to handle diverse time series learning tasks.
• Provide a scikit-learn compatible core API that supports reduction, composition, and modular transformations for time series.
• Showcase state-of-the-art time series classification, forecasting, and benchmarking tools within a single framework.

Proposed method

• Describe time series tasks and their relationships via reductions to simpler tasks.
• Present a data model and a nested, pandas-compatible data container to handle multivariate, panel, and time-heterogeneous data.
• Define task-specific estimators and transformers that extend scikit-learn interfaces for time series use cases.
• Introduce meta-estimators for reductions and multivariate ensembling to enable flexible model composition.
• Provide an API overview including classification, forecasting, transformers, composition, and benchmarking tools.

Experimental results

Research questions

• RQ1 How can time series tasks be unified under a single API that remains compatible with scikit-learn?
• RQ2 Can reductions and meta-estimators enable flexible solving of complex time series tasks by composing simpler components?
• RQ3 What data representations and transformation frameworks best support diverse time series data (multivariate, panel, time-heterogeneous)?

Key findings

• sktime provides a unified API that extends scikit-learn for time series tasks, enabling reductions from complex tasks to simpler ones.
• It introduces a nested/pandas-based data container to support multivariate, panel, and time-heterogeneous data while reusing scikit-learn functionality.
• The library includes state-of-the-art time series classification methods, classical forecasting interfaces, transformers, and benchmarking tools.
• It features composition capabilities like pipelines, ensembles, and novel reduction meta-estimators for configurable modeling choices.
• sktime supports multivariate column-wise ensembling and column concatenation for time series data, enabling flexible model construction.