Efficient AUC learning curve calculation

Abstract

A learning curve of a performance measure provides a graphical method with many benefits for judging classifier properties. The area under the ROC curve (AUC) is a useful and increasingly popular performance measure. In this paper, we consider the computational aspects of calculating AUC learning curves. A new method is provided for incrementally updating exact AUC curves and for calculating approximate AUC curves for datasets with millions of instances. Both theoretical and empirical justifications are given for the approximation. Variants for incremental exact and approximate AUC curves are provided as well. © Springer-Verlag Berlin Heidelberg 2006.