Learning Unlabeled-Unlabeled Classifier Using Nearest Neighbor Distance Ensemble

Abstract

Most of instances in big-data are unlabeled, and the number of the available labeled instances are so limited that semi-supervised learning approaches are not effectively applied in many cases. This fact is one of the main obstacles for effective use of the big-data. In this study, we propose a novel approach to highly efficiently learn an accurate binary classifier from two given unlabeled data sets only. The approach classifies a given instance based on ensemble difference between its nearest-neighbor distances in the two unlabeled data sets. It provides consistent classification results within a constant computation time based on its mathematical background nature. Numerical experiments show high accuracies of the approach close to their upper bounds provided by Bayes error rates.