Optimal ROC Curve for a Combination of Classifiers
Marco Barreno, Alvaro A. Cárdenas, J. D. Tygar Poster M33


Main contributions:
­

ROC curve

A new method for combining classifiers optimally Generalization of previous work (Provost & Fawcett 2001, Flach & Wu 2005) Experiments showing effectiveness and feasibility
Figure 1: Our meta-classifier (lines) outperforms base classifiers (points).

­

­



Theorems:

1. We prove our method's ROC curve is everywhere optimal 2. In the independent case, we show AND and OR are optimal points 3. We give sufficient conditions for optimality of Provost & Fawcett's method