Prediction of Refactoring-Prone Classes Using Ensemble Learning

Abstract

A considerable amount of software engineers’ efforts go into maintaining code repositories, which involves identifying code whose structure can be improved. This often involves the identification of classes whose code requires refactoring. The early detection of refactoring-prone classes has the potential to reduce the costs and efforts that go into maintaining source code repositories. The purpose of this research is to develop prediction models using source code metrics for detecting patterns in object oriented source code, which are indicators of classes that are likely to be refactored in future iterations. In this study, four different sets of source code metrics have been considered as an input for refactoring prediction to evaluate the impact of these source code metrics on model performance. The impact of these source code metrics are evaluated using eleven different classification technique, and two different ensemble classes on seven different open source projects. Ensemble learning techniques have been shown to incorporate the diversity of patterns learnt by different classifiers, resulting in an augmented classifier that is more robust than any individual classifier. Our work also creates distinction between various sets of features for the task of predicting refactoring-prone classes.