Author : Linchao Chen
Publisher :
ISBN 13 :
Total Pages : pages
Book Rating : 4.:/5 (913 download)
Book Synopsis Predictive Modeling of Spatio-temporal Datasets in High Dimensions by : Linchao Chen
Download or read book Predictive Modeling of Spatio-temporal Datasets in High Dimensions written by Linchao Chen and published by . This book was released on 2015 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: Spatio-temporal datasets in many research areas may be very high-dimensional and have complicated dependence structures in both space and time. Various dimension reduction techniques have been introduced to extract key information from single or multiple datasets. Standard dimension-reduced Bayesian hierarchical modeling approaches use a few spatial basis functions to capture the spatial structure and model temporal dynamics through their coefficients. However, these approaches are not always satisfactory in predictive modeling of complicated systems. It is challenging to build statistical models that are computationally efficient for very large datasets and able to adequately model the dependence structure. In this dissertation, we focus on developing modeling strategies for high-dimensional spatio-temporal datasets in the context of predictive modeling, especially when standard dimension reduction methods do not perform well. We discuss modeling strategies to synthesize multiple high-dimensional information sources and produce accurate predictions. We consider the sea surface temperature (SST) prediction problem of Berliner et al. [2000] and demonstrate challenges and advantages of incorporating high-dimensional climate model output along with temperature observations to improve predictive accuracy. Furthermore, we develop a two-scale modeling strategy for dimension-reduced modeling. This modeling strategy is able to capture both large-scale and local dynamics using a large-scale component derived from standard dimension reduction techniques and a local component customized for each pre-partitioned subregion. Empirical experiments show that this strategy is more efficient in capturing variability when standard dimension reduction methods do not perform well. It delivers more balanced results in the presence of outliers. We also discuss its potential as a spatially varying modeling approach. Finally, we gather examples of how one can use coarse scale information in dimension reduction, EOF analysis, predictive modeling, generation of partitions, and provide discussion on potential strategies to choose appropriate aggregation scales.
