Maximum Realisable Performance: a Principled Method for Enhancing Performance by Using Multiple Classiiers Maximum Realisable Performance: a Principled Method for Enhancing Performance by Using Multiple Classiiers in Variable Cost Problem Domains

A novel method is described for obtaining superior classiication performance over a variable range of classiication costs. By analysis of a set of existing classiiers using a receiver operating characteristic (ROC) curve, a set of new realisable classiiers may be obtained by a principled random combination of two of the existing classiiers. These classiiers lie on the convex hull that contains the original ROC points for the existing classiiers. This hull is the maximum realisable ROC (MRROC). A theorem for this method is derived and proved from an observation about ROC data, and experimental results verify that a superior classii-cation system may be constructed using only the existing classiiers and the information of the original ROC data. This new system is shown to produce the MRROC, and as such provides a powerful technique for improving classiication systems in problem domains within which classii-cation costs may not be known a priori Lovell et al., 1997b, Lovell et al., 1997a].