Abstract

This paper examines the problem of selecting a suitable subset of data to be labelled when building pattern classifiers from labelled and unlabelled data. The selection of representative set is guided by a clustering information and various options of allocating a number of samples within clusters and their distributions are investigated. The experimental results show that hybrid methods like Semi-supervised clustering with selective sampling can result in building a classifier which requires much less labelled data in order to achieve a comparable classification performance to classifiers built only on the basis of labelled data.