Abstract: The task of multivariate time series (MTS) classification aims to determine the label of MTS samples. MTS data have rich relationship information such as temporal relationship and sample similarity relationship. However, the existing methods fail to make full use of these relationship information, which makes it difficult to improve the classification performance. For this reason, this paper proposes an MTS classification method based on graph convolutional network (GCN), which improves the classification performance by mining the potential relationship between samples. Firstly, in order to effectively represent the sample relationship, rules of building graph based on sample similarity are designed to model the samples, which can map the potential relationship information of samples into a graph space. Then, a classification model based on graph convolution is proposed, which captures the potential sample relationship conducive to classification by aggregating sample features, and updates them to the sample's own feature vector to improve the classification accuracy. Extensive experiments on eleven public datasets show that the proposed method is superior to twelve comparison methods, which shows that the proposed method provides a new approach for dealing with the problem of time series classification. It really has an important influence on the classification results by mining the potential relationship between time series data for classification.
Keywords: multivariate time series classification; sample similarity; graph convolutional network(GCN); potential relationship; feature aggregation
合肥工业大学学报自科版