A Feature Extraction Method for Multivariate Time Series Classification Using Temporal Patterns

Multiple variables and high dimensions are two main challenges for classification of Multivariate Time Series (MTS) data. In order to overcome these challenges, feature extraction should be performed before performing classification. However, the existing feature extraction methods lose the important correlations among the variables while reducing high dimensions of MTS. Hence, in this paper, we propose a new feature extraction method combined with different classifiers to provide a general classification strategy for MTS data which can be applied for different area problems of MTS data. The proposed algorithm can handle data of high feature dimensions efficiently with unequal length and discover the relationship within the same and between different component univariate time series for MTS data. Hence, the proposed feature extraction method is application-independent and therefore does not depend on domain knowledge of relevant features or assumption about underling data models. We evaluate the algorithm on one synthetic dataset and two real-world datasets. The comparison experimental result shows that the proposed algorithm can achieve higher classification accuracy and F-measure value.