FlexCon-CE: a semi-supervised method with an ensemble-based adaptive confidence

Abstract

Semi-supervised learning is characterized by a low number of labeled instances and a high number of unlabeled instances. FlexCon-C (Flexible Confidence Classifier) is a well-known semi-supervised method that uses the self-training learning algorithm as basis to generate prediction models. The main difference between self-training and FlexCon-C is that the former uses a fixed threshold to select the unlabeled instances, while the latter has a dynamically adjusted confidence. FlexCon-C applies a confidence adjustment equation based on the the classifier performance. In this sense, the classifier performance is used to select and to label unlabeled instances. In Machine Learning, it is well-known that the classifier performance can be further improved through the use of classifier ensembles. Therefore, this study proposes the use classifier ensembles in the FlexCon-C confidence adjustment equation, aiming to provide a more efficient measure to select and to label unlabeled instances. In order to assess the viability of the proposed method (FlexCon-CE), an empirical analysis will be conducted, using 20 datasets, three different classification algorithms and five different configurations of initially unlabeled data. The results indicate that the proposed method outperformed the traditional method, therewith proving itself promising for the task of automatic data selection and labeling in the semi-supervised context.