Penalized AUC Maximization with Kernel Property

• Activity Number: 511
• Type: Contributed
• Date/Time: Wednesday, August 1, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Shuxin Yin*+ and Youyi Fong and Ying Huang
• Companies: Fred Hutchinson Cancer Research Center and Fred Hutchinson Cancer Research Center and Fred Hutchinson Cancer Research Center
• Keywords: ROC Curve ; DC Algorithm ; Classification ; Nonlinear Kernel ; Combination of Biomarker

Abstract

The area under ROC curve is better performance measure than the likelihood of a logistic model when the goal of study is to derive a combination of predictors for classification. However, optimization of AUC is difficult since AUC objective function is neither smooth nor convex. To overcome this difficulty, several smoothed AUC functions had been proposed to approximate such AUC function. In this talk we introduce a new approach to approximate the AUC function by a non-smooth function which can be optimized by DC algorithm (difference of convex functions). The optimization procedure naturally leads to a kernel property, ie, the predictors can be mapped to an infinitely dimensional space while the optimal linear combination of the predictors has a finitely dimensional representation. We shall present some simulation study results showing our new method performs well. Real data from HIV vaccine studies will be used to illustrate our method.