An investigation into time-varying characteristics of multivariate time series in Grassmann classification

Since multivariate time series (MTS), which lie in a non-Euclidean space, exhibit temporal evolution and correlation characteristics, its classification is considered a non-trivial task. To mitigate the impact of the time-varying characteristics and thus enhance the classification accuracy, in this paper, we propose to model MTS data using a time-varying linear dynamical system followed by a neural network-based classification on the Grassmannian manifold. Our experiments on publicly available MTS datasets show promising classification results.