paper/paper.md

In compstat-lmu/TSClassification: Interface to various time series classifiers from the "The Great Time Series Classification Bake Off"

title: 'Time Series Classification in R' tags: - R - Time Series Classification - Functional Data authors: - name: Florian Pfisterer orcid: 0000-0001-8867-762X affiliation: 1 - name: Yu Wang orcid: 0000-0002-8319-2997 affiliation: 2 - name: Xudong Sun orcid: 0000-0001-9234-4932 affiliation: 1 - name: Bernd Bischl orcid: 0000-0001-6002-6980 affiliation: 1 affiliations: - name: LMU Munich index: 1 - name: TU Munich index: 2 date: 31.10.2019 bibliography: paper.bib output: pdf_document

Summary

Time series classification problems arise in many different domains such as speech recognition, medical signals, gesture recognition [@geurts2001pattern], or in industry often as a byproduct of measuring production processes. A prominent example for time-series data are ECG data, where the heart's electrical activity is measured over time.

Due to the widespread application domains, algorithms for time-series classification have been widely studied. Yet many modern methods are not available in R [@team2013r].

In order to compare different algorithms, 20 different TSC algorithms were implemented in a common Java framework integrated with the WEKA toolkit [@hallweka] by [@Bagnall2017]. This made a benchmark comparison between different algorithms feasible. [@Bagnall2017] implement many current state-of-the-art algorithms, such as HIVE, HIVE-COTE, BOSS and others. They find that several methods, such as HIVE-COTE [@hivecote] outperform many other methods including convolutional neural networks. Because of a focus on a reproducible benchmark, no real user-interface for training and testing on datasets is made available. We aim to make these algorithms available to a wider audience. Currently, state-of-the-art algorithms are only available in Java. The R ecosystem on the other hand already offers several implementations for pre-processing, smoothing and imputation of time-series data, and users might often need the best of both worlds. To achieve this, we adapt the codebase from [@Bagnall2017] in order to enable training and prediction on user-supplied data and hyperparameters.

With this software, we aim to make state-of-the-art algorithms for time-series classification available to scientists and users. A benchmark of the various algorithms can be obtained from the Time Series Classification Repository. Users can train over 30 time-series classification methods with specified hyper-parameters and use the learned models to predict future data. Training models is made possible from data.frames as well as arff files. The software currently focuses on 1-dimensional time-series measured on evenly spaced grids. We aim to extend the software, once implementations of multi-dimensional algorithms are available.

We interface the Java methods using rJava [@urbanek2013rjava] in order to provide an easily accessible interface to the R computing language. Additionally, we provide documentation for the integrated methods which is not easily available from [@Bagnall2017]. Our software enables seamless access to the Time Series Classification algorithms implemented in Java from R, and allows users to combine many different methods implemented in R e.g. for time-series preprocessing with several state-of-the-art time-series classifiers.

The software is available at TSClassification, with its Java source code at jTSC4R repository.

nocite: tsclassif

References

compstat-lmu/TSClassification documentation built on Nov. 18, 2019, 6:56 a.m.