多变量时间序列模式挖掘的研究  

Research of data mining method on multivariate time series

作　　者：张军[1] 吴绍春[1] 王炜[2] 

机构地区：[1]上海大学计算机科学和工程学院 [2]上海地震局

出　　处：《计算机工程与设计》2006年第18期3364-3366,3384,共4页Computer Engineering and Design

基　　金：国家地震科学联合基金项目(104090);上海市自然科学基金项目(7A05468)。

摘　　要：多变量时间序列数据集合在许多领域中存在,由于其观测变量之间的相互关联性,往往需要进行综合分析。使用基于时间序列相似性的多变量时间序列模式挖掘方法,从历史数据中寻找出相似的多变量时间序列。将多变量的数据集分段平均为连续矩阵,并采用基于主成分分析和奇异值分解的方法来对矩阵进行相似性比较,最后通过相邻片断的合并以组成更高层次的时序片断,以提高模式的匹配的范围。并在地震前兆数据进行了实现。The multivariate time series （MTS） dataset is a common data type in various scientific domains. An MTS is usually very high dimensional with its main distinguishing characteristic being the inter-correlations and/or interdependencies among its variables. Consequently, MTS may not be easily broken into multiple univariate time series and is treated as a whole. A time series pattern mining method based on similarity of MTS, finds some similar multivariate time series from history data, and then stored as time series pattern. Use extending SVD-based similarity measures for MTS datasets by representing an MTS as a matrix. Then get the longer time series segment through bottom-to-up merge between adjacent segments to improve the ability of real time forecast. Several experiments performed on the earthquake auspice datasets.

关 键 词：数据挖掘 多变量时间序列 相似性 数据预处理 频繁序列模式 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]