EcoSta 2023: Start Registration
View Submission - EcoSta2023
A0522
Title: Variable selection for AUC-optimizing classification in diverging dimensions Authors:  Hyungwoo Kim - Pukyong National University (Korea, South) [presenting]
Abstract: The purpose is to investigate the asymptotic behaviours of the estimator of the AUC-optimizing classification penalized by the smoothly clipped absolute deviation (SCAD) penalty. First, the AUC consistency over the linear function class is studied. Then it is proven that the SCAD-penalized estimator possesses the oracle property under both cases where the predictor dimension is fixed and diverges to infinity. For choosing the regularization parameter in SCAD-penalized AUC-optimizing classification, a BIC-type information criterion is proposed,, shown to capture the true model consistently. The technical proofs are based on the theory of U-processes. In addition, an applicable computation algorithm has been developed to estimate the SCAD-penalized estimator. Both simulated and real data analysis results demonstrate the promising performance of the proposed method in terms of variable selection and prediction.