Shapelet transforms for univariate and multivariate time series classification

• Main Author: Bostrom, Aaron
• Published: University of East Anglia 2018
• Subjects: 004
• Online Access: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.743360

Time Series Classification (TSC) is a growing field of machine learning research. One particular algorithm from the TSC literature is the Shapelet Transform (ST). Shapelets are a phase independent subsequences that are extracted from times series to form discriminatory features. It has been shown that using the shapelets to transform the datasets into a new space can improve performance. One of the major problems with ST, is that the algorithm is O(n2m4), where n is the number of time series and m is the length of the series. As a problem increases in sizes, or additional dimensions are added, the algorithm quickly becomes computationally infeasible. The research question addressed is whether the shapelet transform be improved in terms of accuracy and speed. Making algorithmic improvements to shapelets will enable the development of multivariate shapelet algorithms that can attempt to solve much larger problems in realistic time frames. In support of this thesis a new distance early abandon method is proposed. A class balancing algorithm is implemented, which uses a one vs. all multi class information gain that enables heuristics which were developed for two class problems. To support these improvements a large scale analysis of the best shapelet algorithms is conducted as part of a larger experimental evaluation. ST is proven to be one of the most accurate algorithms in TSC on the UCR-UEA datasets. Contract classification is proposed for shapelets, where a fixed run time is set, and the number of shapelets is bounded. Four search algorithms are evaluated with fixed run times of one hour and one day, three of which are not significantly worse than a full enumeration. Finally, three multivariate shapelet algorithms are developed and compared to benchmark results and multivariate dynamic time warping.