一种时间序列快速分段及符号化方法

作为一类重要的复杂类型数据,时间序列已成为数据挖掘领域的热点研究对象之一.针对时间序列的挖掘通常首先需要将时间序列分段并转变为种类有限的符号序列,以利于进一步进行时间序列模式挖掘.针对当前的时间序列分段方法复杂度较大,效率不高等问题,本文提出了一种简单高效的基于拐点检测的时间序列分段方法,并且采用动态时间弯曲度量计算不等长子序列的相异度,最后运用层次化聚类算法实现子序列的分类及符号化.实验表明,本文所提出的方法切实可行,实验结果具有较为明显的物理意义.

A Fast Time Series Segmentation and Symbolization Method

As one of the important forms of complex data, time series is a hotspot in data mining area. Sequence pat- tern mining is based on time series symbolization, which segments the time series into sub-series and labels them. But most current time series segmentation algorithms are with large computation complexity, so the paper introduces a sim- ple but high efficiency time series segmentation method based on change point detection. And dynamic time warping (DTW) method is used to compute the distance of the sub-series, later the hierarchical clustering is used to group the sub-series and label them. The experiments show the proposed method is feasible and the results are meaningful.