Markers selection via the partialarea under ROC curve


Yuan-chin Chang
2009-12-25  10:40 - 11:30
Room 308, Mathematics Research Center Building (ori. New Math. Bldg.)

Classification is a popular task in statistics and machine learning. Among many performance measures for binary classification problems, the receiver operation characteristic (ROC) curve and some related criteria derived from it, such as the area under the ROC curve (AUC) and the partial AUC, are the most commonly used measures. In this talk, we first review the traditional approaches, which are usually based on the normality assumption. Then we discuss some nonparametric-based algorithm for finding the best linear combination of linear markers such that the final classifier can have the maximum AUC or the partial AUC. For illustration purpose, we apply the proposed methods to some simulated and real data sets.