Comparing three lower bounding methods for DTW in time series classification

Abstract

In comparison to Euclidean distance, Dynamic Time Warping (DTW) is a much more robust distance measure for time series data. For speeding up DTW computation, a few lower bounding techniques have been proposed in literature to guarantee no false dismissals in time series similarity search. In this work, we apply DTW lower bounding method in time series classification and empirically compare three different typical lower bounding techniques for DTW: LB_Keogh, FTW and LB_Improved in this time series data mining task. Our experimental results show that LB_Keogh and LB_Improved perform well with small warping window widths while FTW is only suitable with large warping window widths or without any constraint on warping windows. Besides, runtime efficiency of LB_Improved is quite poor due to its high complexity in lower bound computation despite of its better pruning power.