Empirical evaluation of semi-supervised naïve Bayes for active learning

• Main Author: Saeed, Awat Abdulla
• Published: University of East Anglia 2018
• Subjects: 004
• Online Access: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.753892

This thesis describes an empirical evaluation of semi-supervised and active learning individually, and in combination for the naïve Bayes classifier. Active learning aims to minimise the amount of labelled data required to train the classifier by using the model to direct the labelling of the most informative unlabelled examples. The key difficulty with active learning is that the initial model often gives a poor direction for labelling the unlabelled data in the early stages. However, using both labelled and unlabelled data with semi-supervised learning might be achieve a better initial model because the limited labelled data are augmented by the information in the unlabelled data. In this thesis, a suite of benchmark datasets is used to evaluate the benefit of semi-supervised learning and presents the learning curves for experiments to compare the performance of each approach. First, we will show that the semi-supervised naïve Bayes does not significantly improve the performance of the naïve Bayes classifier. Subsequently, a down-weighting technique is used to control the influence of the unlabelled data, but again this does not improve performance. In the next experiment, a novel algorithm is proposed by using a sigmoid transformation to recalibrate the overly confident naïve Bayes classifier. This algorithm does not significantly improve on the naïve Bayes classifier, but at least does improve the semi-supervised naïve Bayes classifier. In the final experiment we investigate the effectiveness of the combination of active and semi-supervised learning and empirically illustrate when the combination does work, and when does not.