Choosing an Appropriate Ensemble Classifier for Balanced Bioinformatics Data

Abstract

Bioinformatics datasets contain a number of characteristics, such as noisy data and difficult to learn class boundaries, which make it challenge to build effective predictive models. One option for improving results is the use of ensemble learning methods, which involve combining the results of multiple predictive models into a single decision. Since we do not rely on a single model, we reduce the effect of any hidden bias which may reside in a single model. In this study, we investigate two ensemble learning methods, Select-Bagging and Random Forest, to find which one is better suited for classification of bioinformatics data. In addition, we examine how the choice of learning algorithms affects the classification results of the Bagging method. We conduct an empirical study using six ensemble classifiers (Random Forest and Select-Bagging utilizing five different classifiers) applied to 12 balanced datasets using three feature rankers along with four feature subset sizes. Based on our results, including statistical analysis, we recommend Random Forest as it is competitive with the best of the Select-Bagging classifiers, and does not require an additional choice of classifier, which can significantly affect classification performance. To our knowledge, this work is unique in terms of investigating the effectiveness of these two ensemble learning methods in the domain of bioinformatics as well as examining how the choice of classifier impacts classification results when using a Bagging-based ensemble learning method.