A Comparison of Two Purity-Based Algorithms When Applied to Semi-supervised Streaming Data Classification

Abstract

Semi-supervised learning algorithms address the problem of learning from partially labeled data. However, most of the semi-supervised classification methods proposed in the literature considers a stationary distribution of data. Which means that future data patterns tend to conform to the data distribution presented in data set throughout the application lifetime. However, for plenty of new variety of applications, this expected scenario is not compatible to reality. Therefore, the research of semi-supervised methods which comprises nonstationary data classification is of a major concern nowadays. In this paper, the KAOGINCSSL algorithm, which copes with non-stationary semi-supervised learning, is analysed when using two different strategies to spread the labels to train the classifiers. The first consist of employing the inductive algorithm KAOGSS to build the classifier and the second relies on using the transductive algorithm PMTLA to spread the labels prior to build the classifier. Results regarding accuracy and processing time involving both algorithms when applied to non-stationary problems are presented.