B0657
Title: Learning of classifiers from partially classified training data
Authors: Geoffrey McLachlan - University of Queensland (Australia) [presenting]
Abstract: A decision rule classifier in supervised learning tasks where labelled data is abundant can have excellent performance. However, labelling large amounts of data is often prohibitive due to time, financial, and expertise constraints. The goal of semi-supervised learning (SSL) is to leverage large amounts of unlabelled data to improve the performance of supervised learning over small datasets. Using a generative model approach rule can be constructed via SSL learning with Bayes' error smaller than that of the rule produced by full supervision. It applies to situations where the probability that a feature vector has a missing label depends solely on its entropy; that is, where the unlabelled data are those that are difficult to classify.