时间序列挖掘中一种新的相似性度量

针对时间序列的全序列聚类展开,提出一种新的相似性度量——全局特征,即从时间序列的统计分布特征、非线性和Fourier频谱转换等3个方面提取11个全局特征构建特征向量。利用特征向量来描述原时间序列,不仅保留了大部分原有的信息,还能加快聚类计算的速度。经过大量的实验验证表明,基于全局特征提取的相似性度量能得到合理的聚类结果,特别是对经济领域的时间序列效果更为明显。例举了2个数据进行实验,并从主观和客观两个角度对聚类结果进行评估。

New similarity measure for mining time series

Proposes a new similarity measure-global characters for whole clustering of time series,that replaces the raw data with 11 global characteristics,from the aspects of statistical distribution,non-linear and Fourier transformation,thus can get a characteristic vector,which can hold most information of the original time seiries and reduce the calculating complexity.Experimentally compares the four similarity measures on three database under group-ward hierarchical clustering,evaluates the results objectively and subjecttively respectively,and is shown to yield useful and reasonable clustering,especially for economic time series.