关键词: 多变量时间序列, 分类, 奇异值分解, 判别局部保持投影, 最大间距准则

Abstract: The existing multivariate time series classification algorithms require sequences of equal length and neglect categories information. In order to solve these defects, a multivariate time series classification algorithm was proposed based on Singular Value Decomposition (SVD) and discriminant locality preserving projection. Based on the idea of dimension reduction, the first right singular vector of samples by SVD was used as feature vector to transform unequal length sequence into a sequence of identical size. Then the feature vector was projected by utilizing discriminant locality preserving projection based on maximum margin criterion, which made full use of categories information to ensure samples of the same class as close as possible and heterogeneous samples as dispersed as possible. Finally, it achieved the classification in a low dimension subspace by using 1 Nearest Neighbor (1NN), Parzen windows, Support Vector Machine (SVM) and Naive Bayes classifier. Experiments were carried out on Australian Sign Language (ASL), Japanese Vowels (JV) and Wafer, the three public multivariate time series datasets. The results show that the proposed algorithm achieves lower classification error rate under the condition of the same time complexity basically.

Key words: multivariate time series, classification, Singular Value Decomposition (SVD), Discriminant Locality Preserving Projection (DLPP), maximum margin criterion