Comparison of Distance Measures in Evolutionary Time Series Segmentation

Time series segmentation is a fundamental component in the process of analyzing and mining time series data. Given a set of pattern templates, evolutionary computation is an appropriate tool to segment time series flexibly and effectively. However, the choice of distance measure in fitness function is very important to evolutionary time series segmentation, for it will affect the convergence of the algorithm greatly. As a simple and easy method, direct point-to-point distance (DPPD) is always used as similarity measure. However, it is brittle to time phase. In this paper, we present three other distance measures for fitness evaluation, which are based on enclosed area, time warping and trend similarity respectively. Moreover, experiments are conducted to compare the performances of new distance measures with the DPPD approach. Results show that new distance measures outperform the DPPD approach in correct match, accurate segmentation.