TS-QUAD: A Smaller Elastic Ensemble For Time Series Classification

Lines, Jason ORCID: https://orcid.org/0000-0002-1496-5941 and Oastler, George (2022) TS-QUAD: A Smaller Elastic Ensemble For Time Series Classification. In: 3rd International Conference on Pattern Recognition and Artificial Intelligence, 2022-06-01 - 2022-06-03, Campus des Cordeliers.

[thumbnail of TS_QUAD]
PDF (TS_QUAD) - Accepted Version
Download (307kB) | Preview


The Elastic Ensemble (EE) is a time series classification (TSC) ensemble that includes eleven nearest neighbour (NN) classifiers that use variations of eight elastic distance measures. While EE offers an accurate solution for TSC in the time domain, its relatively slow run-time is a weakness. This has led to new algorithms, such as Proximity Forest and TS-CHIEF, that have iterated on the design of EE by taking the same elastic measures and incorporating them into tree-based ensembles. These enhancements were implemented successfully and led to faster and more accurate time domain classifiers and, as such, development on the original EE algorithm subsided. However, in this work we make the simple hypothesis that the original design of EE contains distance measures that capture the same discriminatory features, and as such, the ensemble includes redundant classifiers. If this were true, EE could perform to the same level in terms of accuracy with significantly less computation. If proven true this would have interesting implications to the design of algorithms such as Proximity Forest and TS-CHIEF that are based on the original EE implementation. To investigate this, we form a simple categorisation of the distance measures within EE and form four groups. We take one measure from each group, building an ensemble of four 1-NN classifiers that we call TS-QUAD: the Time Series QUARtet of distance-based classifiers. We demonstrate that this ensemble is able to match EE in terms of accuracy over 10 resamples of 85 datasets while containing fewer than 50% of the original EE constituents, implying that other elastic distance-based TSC ensembles could benefit from the design philosophy of TS-QUAD.

Item Type: Conference or Workshop Item (Paper)
Uncontrolled Keywords: classification,elastic distance measures,time series,theoretical computer science,computer science(all) ,/dk/atira/pure/subjectarea/asjc/2600/2614
Faculty \ School: Faculty of Science > School of Computing Sciences
UEA Research Groups: Faculty of Science > Research Groups > Data Science and Statistics
Faculty of Science > Research Groups > Smart Emerging Technologies
Related URLs:
Depositing User: LivePure Connector
Date Deposited: 13 May 2022 14:56
Last Modified: 20 Mar 2023 09:30
URI: https://ueaeprints.uea.ac.uk/id/eprint/85039
DOI: 10.1007/978-3-031-09282-4_19


Downloads per month over past year

Actions (login required)

View Item View Item