A Universal Model for Defective Classes Prediction Using Different Object-Oriented Metrics Suites

Abstract

Recently, research studies were directed to the construction of a universal defect prediction model. Such models are trained using different projects to have enough training data and be generic. One of the main challenges in the construction of a universal model is the different distributions of metrics in various projects. In this study, we aim to build a universal defect prediction model to predict software defective classes. We also aim to validate the Object-Oriented Cognitive Complexity metrics suite (CC metrics) for its association with fault-proneness. Finally, this study aims to compare the prediction performances of the CC metrics and the Chidamber and Kemerer metrics suite (CK metrics), taking into account the effect of preprocessing techniques. A neural network model is constructed using these 2 metrics suites (CK & CC metrics suites). We apply different preprocessing techniques on these metrics to overcome variations in their distributions. The results show that the CK metrics perform well whether a preprocessing is applied or not, while CC metrics’ performance is significantly affected by different preprocessing techniques. The CC metrics always outperform in the recall, while the CK metrics usually outperform in other performance metrics. Normalization preprocessing results in the highest recall values using either of the two metrics suites.