Predicting Software Maintainability in Object-Oriented Systems Using Ensemble Techniques

Abstract

Prediction of the maintainability of classes in object-oriented systems is a significant factor for software success, however it is a challenging task to achieve. To date, several machine learning models have been applied with variable results and no clear indication of which techniques are more appropriate. With the goal of achieving more consistent results, this paper presents the first set of results in an extensive empirical study designed to evaluate the capability of bagging models to increase accuracy prediction over individual models. The study compares two major machine learning based approaches for predicting software maintainability: individual models (regression tree, multilayer perceptron, k-nearest neighbors and m5rules), and an ensemble model (bagging) that are applied to the QUES data set. The results obtained from this study indicate that k-nearest neighbors model outperformed all other individual models. The bagging ensemble model improved accuracy prediction significantly over almost all individual models, and the bagging ensemble models with k-nearest neighbors as a base model achieved superior accurate prediction. This paper also provides a description of the planned programme of research which aims to investigate the performance over various datasets of advanced (ensemble-based) machine learning models.