AUC Maximizing Support Vector Learning

The area under the ROC curve (AUC) is a natural performance measure when the goal is to find a discriminative decision function. We present a rigorous derivation of an AUC maximizing Support Vector Machine; its optimization criterion is composed of a convex bound on the AUC and a margin term. The number of constraints in the optimization problem grows quadratically in the number of examples. We discuss an approximation for large data sets that clusters the constraints. Our experiments show that the AUC maximizing Support Vector Machine does in fact lead to higher AUC values.