Time Series classification through transformation and ensembles

Lines, Jason (2015) Time Series classification through transformation and ensembles. Doctoral thesis, University of East Anglia.

[img]
Preview
PDF
Download (9MB) | Preview

    Abstract

    The problem of time series classification (TSC), where we consider any real-valued ordered data a time series, offers a specific challenge. Unlike traditional classification
    problems, the ordering of attributes is often crucial for identifying discriminatory features between classes. TSC problems arise across a diverse range of domains, and this
    variety has meant that no single approach outperforms all others.
    The general consensus is that the benchmark for TSC is nearest neighbour (NN) classifiers using Euclidean distance or Dynamic Time Warping (DTW). Though conceptually simple, many have reported that NN classifiers are very diffi�cult to beat and new work is often compared to NN classifiers. The majority of approaches have focused on classification in the time domain, typically proposing alternative elastic similarity measures
    for NN classification. Other work has investigated more specialised approaches, such as building support vector machines on variable intervals and creating tree-based
    ensembles with summary measures.
    We wish to answer a specific research question: given a new TSC problem without any prior, specialised knowledge, what is the best way to approach the problem? Our thesis is that the best methodology is to first transform data into alternative representations where discriminatory features are more easily detected, and then build ensemble
    classifiers on each representation. In support of our thesis, we propose an elastic ensemble classifier that we believe is the first ever to significantly outperform DTW on the widely used UCR datasets. Next, we propose the shapelet-transform, a new data transformation that allows complex classifiers to be coupled with shapelets, which outperforms the original algorithm and is competitive with DTW. Finally, we combine these two works with with heterogeneous ensembles built on autocorrelation and spectral-transformed data to propose a collective of transformation-based ensembles (COTE). The results of COTE are, we believe, the best ever published on the UCR datasets.

    Item Type: Thesis (Doctoral)
    Faculty \ School: Faculty of Science > School of Computing Sciences
    Depositing User: Brian Watkins
    Date Deposited: 15 Jun 2015 15:07
    Last Modified: 15 Jun 2015 15:07
    URI: https://ueaeprints.uea.ac.uk/id/eprint/53360
    DOI:

    Actions (login required)

    View Item