AUC Maximization in Deep Neural Network Learning for Imbalanced Classification Problems

摘要

As shown in Table 1, due to the AUC maximizationduring the optimization, the proposedmethod performed higher AUC than the conventionalmethods on CIFAR10, MNIST and SVHNexcept USPS, but the gap is close. Also our AUCloss has only one scaling parameter to tune comparewith focal loss which has two hyperparameters.This strategy makes our loss function easierto be applied.