Turning the hyperparameter of an AUC-optimized classifier

The Area under the ROC curve (AUC) is a good alternative to the standard empirical risk (classification error) as a performance criterion for classifiers. While most classifier formulations aim at minimizing the classification error, few methods exist that directly optimize the AUC. Moreover, the reported methods that optimize the AUC are often not efficient even for moderately sized datasets. In this paper, we discuss a classifier that optimizes the AUC using a linear programming formulation such that classification constraints can easily be subsampled. Furthermore, this approach enables to use the non-used constraints to for the optimization of hyperparameters. In a performance evaluation, we compare the AUC-optimized Linear Programming Classifier (AUC-LPC) to other classifiers on several real world datasets.