Bayesian Elastic Net for Multi-Class Classification and Survival Analysis

• Activity Number: 424
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Lingling Zheng*+
• Companies: Duke University
• Keywords: Variational Bayesian ; Bayesian Elastic Net ; Survival Analysis ; Multinomial Probit Regression ; Variable Selection ; Gibbs Sampling

Abstract

Bayesian Elastic Net is an advanced technique for addressing the problem of grouped variable selection and sparseness.I am interested in adopting this strategy for clinical application. In this paper, I develop a multinomial probit regression model of Bayesian Elastic Net for classification problem, i.e. identifying classifiers to infer a set of important and highly correlated predictors, e.g. genes or peptides. Additionally, in order to study association between survival and gene expression signature, I present censored exponential regression model of Bayesian Elastic Net. Furthermore, missing data imputation is also considered for both cases. Inference of these approaches is conducted through both Gibbs sampling and variational Bayesian (VB) approximation. The two models are validated by first performing simulation on toy datasets; then I obtain biochemical measurements data from sepsis patients to classify and predict their disease status. Finally, I assess Bayesian survival model on lung cancer patients' microarray data and failure time. The methods show that Gibbs sampling has better accuracy in classification, while VB tends to achieve better sparseness and efficiency.