Covariance-Insured Screening Methods for Ultrahigh-Dimensional Variable Selection

• Activity Number: 579
• Type: Invited
• Date/Time: Wednesday, August 3, 2016 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Yi Li* and Ji Zhu and Jiashun Jin and Kevin He and Yanming Li
• Companies: University of Michigan and University of Michigan and Carnegie Mellon University and University of Michigan and University of Michigan
• Keywords: big data ; classification ; generalized linear regressio ; screening ; survival analysis ; variable selection

Abstract

Effective screening methods are crucial to the analysis of big biomedical data. The popular marginal screening relies on restricted assumptions such as the partial faithfulness condition, e.g, the partial correlation between outcome and covariates can be inferred from their marginal correlation. However, such a restrictive assumption is often violated, as the marginal effects of predictors may be quite different from their joint effects, especially when the covariates are correlated. We propose a covariance-insured screening (CIS) framework that utilizes the dependence among covariates and identify important features that are likely to be missed by marginal screening procedures such as sure independence screening. The proposed framework encompasses linear regression models, generalized linear regression models, survival models, and classification of multi-level outcomes.