CAREER: Statistical Methodology in Multi-view Learning with Large Data

West Virginia University Research Corporation, Morgantown WV

Investigators

Abstract

In complex scientific research classification and regression problems, it is common for several different data sets to be used to describe the response. Each of these data sets provides a unique view of the response, but typically none of these views describes the response perfectly. The investigator develops computationally efficient statistical methodology to model multi-view data within a framework for statistical analysis, variable and view selection, and the interpretation of results. The methodology is based on both regularization approaches involving sparse penalties in additive models and algorithmic intensive iterative-based approaches. In addition, the investigator provides a solid foundation for analysis of residuals in this context, establishes the consistency of the model, and addresses the practical issue of concurvity. The investigator is committed to raising awareness in scientific research communities and industries of the advantages of using these modeling techniques in the analysis of complex research problems involving data from multiple sources and to training future statisticians and related professionals through hands-on experiences with these data sets. To support this effort, the investigator develops and maintains a powerful, user-friendly statistical software package to implement this methodology. In the modern world our ability to collect data from many different sources has expanded dramatically due in part to computer innovations over the past few decades. What has not kept pace is the ability to analyze data from many different sources simultaneously. As a result, scientific researchers in academia and industry are not fully harnessing the information that can be found by appropriately combining multiple, diverse sources of data in a way that can provide interpretable results. This is a challenging problem involving advances in statistics, computer science, mathematics, and database management. The investigator addresses this problem from a statistical analysis viewpoint. Many applications of this research involve new statistical methods to help with cancer research, pharmacology, genetics, proteomics, text data processing, and homeland security. The need for analyses of large, complex, multi-view data sets is substantial in scientific research today, and this need is currently unmet. The results of this work are transforming how researchers in many fields analyze and interpret data.