Research on automatic grouping of multivariate time series based on the information granularity

在这项研究中，我们发展了一种多变量时间序列的多尺度比较方法。我们的方法首先从时间序列构造多维轨迹，并用多尺度表示来表示它们。然后，根据曲率极大值的位置将轨迹分割成数据粒。然后跟踪数据粒的层次结构，并跨尺度执行逐个粒的匹配，以找到轨迹之间的最佳对应。在医学数据集上的实验结果表明，我们的方法可以生成具有相似时间过程的轨迹组，并且一些簇显示出关于纤维化阶段分布的有趣特征。

In this research we have developed a multiscale comparison method for multivariate time series. Our method firstly constructs multidimensional trajectories from the time series, and represent them using multiscale representation. Next, it splits the trajectories into data granules according to the positions of curvature maxima. Then it traces the hierarchical structure of data granules and performs granule-by-granule matching across the scales to find the best correspondences between the trajectories. Experimental results on a medical dataset showed that our method could generate groups of trajectories that exhibited similar temporal courses, and some of the clusters showed interesting characteristics about the distribution of fibrotic stages.