Seminars

Optimality of ROC Surfaces

74
reads

Yun-Jhong Wu

2010-12-10
12:45:00 - 14:45:00

R440 , Astronomy and Mathematics Building

Receiver operating characteristic (ROC) analysis is used to evaluate the performance of classification procedures. Although relevant techniques are now fairly well-established in binary classification, some theoretical gaps need to be filled before extending ROC analysis to ternary or more complicated classification problems. Indeed, the optimality of ROC surfaces and its connection with linear classifiers would become of importance for classification with several populations; striking distinction of properties between the true-positive and false-positive ROC surfaces also arise from this. From a different angle, utility and related classification procedures are worthwhile being mentioned as a tool to characterize ROC surfaces. It is additionally illustrated that the well-definition of the volume under the ROC surface (VUS), used traditionally as an index to figure out the accuracy of classifiers, might be thrown into doubt. This discussion will enable us to construct models of ROC surfaces on a formalistic basis.