Accurate symbolization of time series

Abstract

Symbolization is a useful method for mining time series. As our experimental results demonstrated, the previous methods are not accurate enough due to their limitations in handling a prevalent kind of time series in which similar movements are often with different lengths. This paper considers the accuracy issue of symbolization of time series. We propose a novel approach that emphasizes the meaning of each movement in the time series, regardless of the length or shift of it. To make the proposed approach more practicable, we also provide a semiautomatic method for setting the parameters. The nature of the problem and the performance of our approach had been analyzed on both real data and synthetic data. Experimental results justified the superiority of our approach over the previous one and gave some useful empirical conclusions. © Springer-Verlag Berlin Heidelberg 2005.