Some novel heuristics for finding the most unusual time series subsequences

Abstract

In this work, we introduce some novel heuristics which can enhance the efficiency of the Heuristic Discord Discovery (HDD) algorithm proposed by Keogh et al. for finding most unusual time series subsequences, called time series discords. Our new heuristics consist of a new discord measure function which helps to set up a range of alternative good orderings for the outer loop in the HDD algorithm and a branch-and-bound search mechanism that is carried out in the inner loop of the algorithm. Through extensive experiments on a variety of diverse datasets, our scheme is shown to have better performance than previous schemes, namely HOT SAX and WAT. © 2010 Springer-Verlag Berlin Heidelberg.