Performance Analysis of Unlabeled-Unlabeled Classifier Learning Using Class Prior Estimation

Abstract

The need to classify big data is increasing in various fields, whereas labeled big data sets are hardly obtained for supervised learning. An efficient remedy to overcome this problem is the use of Unlabeled-Unlabeled Classification (UUC) methods which belong to the category of weak learning. They construct a classifier using from only two unlabeled datasets having mutually different class prior probabilities. Particularly, an empirical loss minimization (ERM) based UUC approach enables to learn various types of binary classifiers in high accuracy. However, it requires class prior probabilities of the two unlabeled datasets, while they are usually unknown. This nature of the ERM-based UUC hinders its widespread application. In this study, we combined an approach to estimate the class prior probabilities of the two unlabeled datasets with the ERM-based UUC to address this limitation, and evaluated the performance of the combined method through numerical experiments.