A0959
Title: Time series of weakly dependent tensors
Authors: Dorcas Ofori-Boateng - Portland State University (United States) [presenting]
Abstract: There are several areas of application for time series of weakly dependent tensors across disciplines involving meteorological, brain imaging, power grid, electricity trading, and bitcoin data. Of course, for many of these datasets, we could consider full spatiotemporal covariances. For example, for meteorological data, the spacing in many cases is approximately a rectangular grid, which gives rise to a latitude against longitude or a latitude against longitude against the time tensor model. By considering the lower dimensional tensor, we can properly account for the dependence within the time component, allowing us to observe breaks/anomalies within the spatial structure of the tensor-related data. We extend tensor graphical models to weakly dependent data, that allows for the modelling of data with arbitrary tensor degree K. For example, in a functional magnetic resonance imaging (fMRI) data set that is collected over space, time, subjects, and multiple scans (or other modalities such as electroencephalography), our new estimator's three-way decomposition cannot only account for temporal dependence within a scan but also over longitudinal time. Next, we provide rates for statistical convergence for the new estimator. We demonstrate the usefulness of our developed algorithm on simulated datasets and against benchmark method results.
