Abstract: We consider a setting in which we construct a binary classifier from a panel of features in order to optimize either the sensitivity at a fixed specificity level or the area under the partial receiver operating characteristic (ROC) curve. To this end, we propose an efficient iterative numerical algorithm to solve a simple constrained optimization problem that mimics the original target. We also present the associated asymptotic statistical inference procedures, including the construction of the credible intervals for the realized sensitivity/specificity or the area under the partial ROC curve of the estimated risk scores. We apply the method to simulated data sets and show that the proposed method outperforms the classifiers based on the generic logistic regression, without considering the specific criterion we want to optimize. We also apply the new proposed method to two real-data examples.

Key words and phrases: Feature ensemble, ROC curve, sensitivity, specificity.