tSNE-Spec: A new classification method for multivariate time series data

Abstract

Classification of multivariate time series (MTS) data has applications in various domains, for example, medical sciences, finance, sports analytics, etc. In this work, we propose a new technique that uses the advantages of dimension reduction through the t-distributed stochastic neighbor embedding (t-SNE) method, coupled with the attractive properties of the spectral density estimates of a time series, and k-nearest neighbor algorithm. We transform each MTS to a lower dimensional time series using t-SNE, making it useful for visualizing and retaining the temporal patterns, and subsequently use that in classification. Then, we extend the standard univariate spectral density-based classification in the multivariate setting and prove its theoretical consistency. Empirically, at first, we establish that the pairwise structure of the multivariate spectral density based distance matrix is retained in the t-SNE transformed spectral density-based distance calculation method, thus indicating that the consistency derived based on multivariate spectral density is transferable to our proposed method. The performance of our proposed method is shown by comparing it against other widely used methods and we find that the proposed algorithm achieves superior classification accuracy across various settings. We also demonstrate the superiority of our method in a real-life health dataset where the task is to classify epilepsy seizures from other activities like walking and running based on accelerometer data.