Evaluating Improvements to the Shapelet Transform

The Shapelet tree algorithm was proposed in 2009 as a novel way to find phase independent subsequences which could be used for time series classification. The shapelet discovery algorithm is O(n2m4), where n is the number of cases, and m is the length of the series. Several methods have sought to increase the speed of finding shapelets. The ShapeletTransform reduces the finding to a single pass, and FastShapelets smooths and reduces the series lengths through PAA and SAX. However neither of these techniques can enumerate all shapelets on the largest of the datasets present in the UCR repository. We first evaluate whether the FastShapelet algorithm is better as a transform, and secondly provide a contract classifier for the shapelet transform, by calculating the number of fundamental operations we can estimate the run time of the algorithm, and sample the data to fulfil this contract. We found that whilst the FastShapeletTransform does drastically reduce the operation count of finding shapelets it is not significantly better than FastShapelets, nor can it compete with the ShapeletTransform. The factory method for sampling the data is competitive with the ShapeletTransform and in some cases we see minor improvements despite being much faster.