Study of Classification Threshold Correction in PU-AUC Optimization Learning

Abstract

It is difficult to prepare enough labeled data. Positive and Unlabeled(PU) classification is a method for learning a binary classifier from only positive and unlabeled data and is expected to solve the problem of data collection cost. The performance of Classifiers trained to minimize classification errors is affected by imbalanced data. There is a way to deal with this problem by optimizing the AUC, and there is a study on AUC optimization for PU classification as well. In this study, when AUC optimization was performed for PU classification, the threshold for determining the classification class was not determined, and a problem occurred in which the predicted labels were biased toward one class. To solve this problem, we devised a learning method in which a correction term for the class fraction of predicted labels is added to the loss formula.