Analysis of the Threshold Variation of the FlexCon-C Algorithm for Semi-supervised Learning
Arthur C. Gorgônio, Cainan T. Alves, Amarildo J. F. Lucena, Flavius L. Gorgônio, Karliane Medeiros Ovidio Vale, Anne M. P. Canuto · 2018
Semi-supervised learning algorithms are able to train classifiers from a small portion of initially labeled objects. The reliability of the classification process depends on several factors that include the type of classifier used and a set of parameters that customize them. One of the most important factors is a threshold that determines which instances are included per iteration, allowing to label only instances with high confidence values. This article analyzes different values for the variation factor of the FlexCon-C algorithm and measures the impact of this change on its accuracy. The results consider thirty different databases, four classifiers and five different percentages of pre-labeled data.