CAREER: Methodology for Statistical Computing in Massive Datasets: Parallel Approaches to Cluster and MCMC Estimation

University Of Maryland Baltimore County, Baltimore MD

Investigators

Abstract

CAREER: Methodology for Statistical Computing in Massive Datasets: Parallel Approaches to Cluster and MCMC Estimation DMS 0239734 PI: Ranjan Maitra This project is aimed at developing practical methodology for statistical analysis and estimation in massively sized databases. Because of automated data collection methods, there is now a surfeit of severely multi-dimensional records. Grouping them into homogeneous clusters to better understand them is a desirable goal in a variety of applications, yet classical statistical methods are computationally infeasible in many cases. I propose to develop parallel methodology for this context. Although the methodology and theory developed will be quite general and for potential use in applications ranging from business, medicine, the environment, software quality assessment, I will conduct the research in the context of three scientific collaborations. The first pertains to the US Environmental Protection Agency's (EPA) self-reported Toxic Releases Inventory (TRI) databases, where profiling the different facilities in terms of their product, demographic and business information can improve the accuracy of records, as well as better characterize them vis-a-vis their emissions mix. The second project is to assess the reliability of functional Magnetic Resonance Imaging (fMRI) scans, with a view to understanding the cognitive processes of the brain, as a first step to patient care and therapy. The third application is in bioinformatics where the goal is to cluster microarray data and also to analyze two-dimensional proteomic gel images. This will help in isolating genes and understanding their relationship with different disorders. Clustering is in general a very difficult problem, with empirical solutions even for very moderately sized datasets. I propose to develop multi-pass methodologies in several different scenarios. I also propose multi-scale simulation approaches to estimation in a high-dimensional context. One of the biggest challenges faced by simulation methods due to high dimensionality is the low mobility around the space to be traversed because of its vastness. I propose to address this issue by connecting these high-dimensional spaces to lower-dimensional ones (which are significantly smaller) and by using these lower scales to traverse from one corner of the higher-dimensional space to another. A final goal of this five-year plan is to investigate the development and estimation in more complex models for proteomic gel data. Most of the plans proposed will be possible only with a parallel computing interface. This is increasingly critical in a large number of scientific applications, and I propose to simultaneously provide statistics students with the necessary expertise by designing suitably tailored graduate and undergraduate classes.