基于分量属性近邻传播的多元时间序列数据聚类方法  

Multivariate time series clustering based on affinity propagation of component attributes

作　　者：李海林[1] 王成[2] 邓晓懿[1] LI Hai-lin, WANG Cheng2, DENG Xiao-yi1(h College of Business Administration, Huaqiao University, Quanzhou 362021, China; 2. College of Computer Sciences and Technology, Huaqiao University, Xiamen 361021, Chin)

机构地区：[1]华侨大学工商管理学院,福建泉州362021 [2]华侨大学计算机科学与技术学院,福建厦门361021

出　　处：《控制与决策》2018年第4期649-656,共8页Control and Decision

基　　金：国家自然科学基金项目(71771094;61300139);福建省社会科学规划项目(FJ2017B065);福建省科技计划引导性项目(2017H01010065);福建省高等学校新世纪优秀人才支持计划项目(Z1625112)

摘　　要：鉴于传统方法不能直接有效地对多元时间序列数据进行聚类分析,提出一种基于分量属性近邻传播的多元时间序列数据聚类方法.通过动态时间弯曲方法度量多元时间序列数据之间的总体距离,利用近邻传播聚类算法分别对数据之间的总体距离矩阵和分量近似距离矩阵进行聚类分析,综合考虑这两种视角下序列数据之间的关联关系,使用近邻传播方法对反映原始多元时间序列数据的综合关系矩阵实现较高质量的聚类.数值实验结果表明,与传统聚类方法相比,所提出方法不仅能够有效地反映总体数据特征之间的关系,而且通过重要分量属性序列之间的关联关系分析能够提高原始时间序列数据的聚类效果.In view of the problem that the traditional methods can not be directly effective on such data clustering analysis, a clustering method of multivariate time series data based on component attributes affinity propagation is proposed. The overall distance between multivariate time series data can be measured by dynamic time warping. The clustering analysis of the overall distance matrix and component approximate distance matrix is processed by using the affinity propagation clustering algorithm, considering the relationship between two sequence data from the two kinds of perspectives. The synthetical relationship matrix of the raw multivariate time series data is used for clustering by using the affinity propagation method. The numerical experiment results how that, compared with the traditional clustering methods, the proposed method not only can effectively reflect the relationship of the overall data characteristics, but also improve the clustering effect of the original time series data through the analysis of the relationship between the important component attributes.

关 键 词：多元时间序列 聚类分析 近邻传播 动态时间弯曲 分量属性 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]