B0275
Title: Spatiotemporal high-dimensional matrix autoregressive models via tensor decomposition
Authors: S Yaser Samadi - Southern Illinois University Carbondale (United States) [presenting]
Rukayya Ibrahim - Penn State Harrisburg (United States)
Tharindu P De Alwis - Worcester Polytechnic Institute - Worcester (United States)
Abstract: With the rapid increase in massive, interactive datasets, including time-dependent big data and spatiotemporal data, various domains such as econometrics, geospatial technologies, and medicine face the challenge of efficiently handling their high dimensionality. To address this complexity, tensor decomposition techniques offer valuable advantages, such as latent structure identification, information extraction, data imputation, and complexity control, making them popular for analyzing, predicting, and forecasting these datasets. A novel approach is introduced for modelling and analyzing matrix-valued spatiotemporal data by formulating it as a tensor regression model based on matrix autoregression. The method capitalizes on the matrix structure of both the response and predictors while achieving dimension reduction through a low-rank tensor structure. Comparative analyses demonstrate the superior efficiency of the model compared to existing approaches for high-dimensional data. Furthermore, two estimation methods are proposed to estimate the transition tensor in both low and high-dimensional scenarios. The asymptotic and non-asymptotic properties of the proposed estimators are also derived, providing a solid theoretical foundation. Simulation studies and real data analysis are conducted to illustrate the advantages of the model over current methodologies.
