Gem

Abstract

The widespread use of digital sensor systems causes a tremendous demand for high-quality time series analysis tools. In this domain the majority of data mining algorithms relies on established distance measures like Dynamic Time Warping (DTW) or Euclidean distance (ED). However, the notion of similarity induced by ED and DTW may lead to unsatisfactory clusterings. In order to address this shortcoming we introduce the Gliding Elastic Match (GEM) algorithm. It determines an optimal local similarity measure of a query time series Q and a subject time series S. The measure is invariant under both local deformation on the measurementaxis and scaling in the time domain. GEM is compared to ED and (un)constrained DTW in terms of matching quality for several datasets and outperforms the competitors in most cases. In order to accelerate the application of GEM to large-scale datasets, we further present an efficient CUDA parallelization.