Multivariate Time-Series Analysis Via Manifold Learning

摘要

This paper presents a data-driven approach for analyzing multivariate time series. It relies on the hypothesis that highdimensional data often lie on a low-dimensional manifold whose geometry may be revealed using manifold learning techniques. We define a notion of distance between multivariate time series and use it to determine a low-dimensional embedding capable of describing the statistics of the signals at hand using just a few parameters. We illustrate our method on two simulated examples and two real datasets containing electroencephalographic recordings (EEG).