Regularization of unlabeled data for learning of classifiers based on mixture models

Abstract

In this paper we investigate the mixture models for classification tasks in the semi-supervised learning framework in which both labeled and unlabeled data are used for training. This framework is very important since in many domains the labeled data are very expensive while a large number of unlabeled data may be freely available. We present a regularization method, so-called the regularized weighting factor to adjust contribution of the unlabeled data during learning process in order to reduce the size of labeled data. Some experiments were performed using benchmark datasets to study this method using the generative classifiers based on Gaussian mixture models. The experiment results have shown that the proposed method can regularize contribution of labeled/unlabeled data during learning process and reduce the labeled data.