Partial AUC Maximization via Nonlinear Scoring Functions.

摘要: We propose a method for maximizing partial area under receiver operating characteristic (ROC) curve (pAUC) binary classification tasks. In tasks, accuracy is the most commonly used as measure of classifier performance. some applications such anomaly detection and diagnostic testing, not an appropriate since prior probabilties are often greatly biased. Although in cases pAUC has been utilized performance measure, few methods have proposed directly pAUC. This optimization achieved by using scoring function. The conventional approach utilizes linear function contrast we newly introduce nonlinear functions this purpose. Specifically, present two types based on generative models deep neural networks. show experimentally that fucntions improve through application real bogus objects obtained with Hyper Suprime-Cam Subaru telescope.