High-Dimensional Universal Dependence Variable Selection

• Activity Number: 296
• Type: Contributed
• Date/Time: Tuesday, July 31, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Hesen Peng*+ and Tianwei Yu and Yun Bai
• Companies: Amazon.com and Emory University and Philadelphia College of Osteopathic Medicine
• Keywords: High-dimensional data ; machine learning ; variable selection ; universal dependence

Abstract

The emergence of high-throughput data in biological science and computer networks has generated novel challenges for statistical methods. Nonlinear relationships and multivariate interactions are abundant. The sheer volume of high-throughput data have limited the application for traditional case-by-case analysis methods, whose model assumptions, like linearity, are not supported in high-throughput scenarios. To meet these challenges, we developed Mira score, a novel probabilistic dependency measure that accounts for high-dimensional nonlinear dependence. Mira score is defined as a function of observation graph, and thus circumvents the curse of dimensionality in high-dimensional data. The superior statistical property enjoyed by Mira score has lead to our development of efficient network reverse-engineering procedure for multivariate dependence. As an example, the procedure has been applied to celiac disease and lung cancer pathway interaction analysis, and has achieved interesting findings. Further more, in the supervised-machine learning scenario, we proposed SeMira procedure, an efficient variable selection procedure that accounts for high-dimensional nonlinear dependence.