Time Series Classification using Covariance Descriptors
Winsala, Oluwaseun Ebenezer
MetadataShow full item record
This thesis presents a novel framework for time series classification that leverages the geometric structure of covariance matrices when labeling signals. Our method maps each signal to a new multivariate localized feature signal (MLFS) representation, from which we compute a covariance descriptor. This robust MLFS covarieance representation handles classification tasks where the sampling rates of the signals vary within a class or classes. We demonstrate that by simply using the k-nearest neighbor classification rule and multiclass kernel support vector machine with the Riemannian metric between the MLFS convariance matrices, which produces state-of-the-art results on a number of standard datasets. Moreover, for the first time, we showcase results on the full library of typical infrasonic signals dataset, which contains four categories of infrasound observations.