基于因子模型和动态规划的多元时间序列分段方法

针对经典动态规划分段算法只适用于低维时间序列的问题,提出一种基于因子模型和动态规划的多元时间序列分段方法.首先利用增量聚类自动对变化趋势相似的变量序列进行聚类,然后引入动态因子模型使降维后的低维多元时间序列能够最大限度反映原始多元时间序列的整体变化趋势,最后利用动态规划在低维多元时间序列的架构上实现高维多元时间序列的分段.实验结果表明,所提方法对变量个数较多的多元时间序列数据具有良好的分段效果.

Segmentation of multivariate time series with factor model and dynamic programming

The classical dynamic programming based segmentation algorithm is only suitable for low dimensional time series. To solve this problem, a segmentation method of multivariate time series with factor model and dynamic programming is proposed. Firstly, incremental clustering is used to automatically cluster variable sequences with similar trend. Then, a dynamic factor model is introduced to make the low-dimension multivariate time series obtained after dimension reduction reflect the overall trend of the original multivariate time series. Finally, the segmentation of high-dimension multivariate time series in the framework of low-dimension time series is realized by using dynamic programming. The experimental studies show that the proposed method has a good segmentation effect on multivariate time series data with a large number of variables.