Software Change Proneness Prediction Using Machine Learning

Abstract

Software change-proneness is one of the vital quality metrics that represents the extent of change of a class across versions of the system. This change may occur due to evolving requirements, bug fixing, or code refactoring. Consequently, change-proneness may have a negative impact on software evolution. For instance, modules that are change-prone tend to produce more defects and accumulate more technical debt. This research work applies different Machine Learning (ML) techniques on a large dataset from a wide commercial software system to investigate the relationships between object-oriented (OO) metrics and change-proneness, and determine which OO metrics are necessary to predict change-prone classes. Moreover, several state-of-the-art combining methods were evaluated that were constructed by combining several heterogeneous single and ensemble classifiers with voting, Select-Best, and staking scheme. The result of the study indicates a high prediction performance of many of the ensemble classifiers as well as the combining methods selected and proved that ML methods are very beneficial for predicting change-prone classes in software. The study also proved that software metrics are significant indicators of class change-proneness and should be monitored regularly during software development and maintenance.