Time Series Classification with Composite Shapelets

One approach to time series data classification is the use of partial data series that show characteristic variations referred to as shapelets.This method classifies data based on information regarding the presence of multiple shapelets.However it does not account for the order of occurrence among the shapelets.In this study, we propose a method that incorporates the occurrence order and intervals between shapelets as features by introducing composite shapelets, which are constructed from base shapelets extracted from the time series data.To reduce the number of shapelet combinations required to identify useful composite shapelets, the proposed method employs two strategies.First, it limits the analysis to base shapelets extracted from the same time series data.Second, it ranks combinations based on the average quality of each component and terminates search when the predefined number of consecutive failures to update quality is reached.Using 108 datasets from the UCR Time Series Classification Archive, experiments demonstrate that the introduction of composite shapelets enhances classification accuracy.