Software quality prediction models: A comparative investigation based on machine learning techniques for object-oriented systems

The purpose of this thesis study is investigating correlation between Chidamber and Kemerer (CK) Object-Oriented (OO) software metrics and determining the accuracy rate in software bug prediction. For this reason, eleven most frequently used Machine Learning (ML) techniques and two Support Vector Machine (SVM) libraries performance was analyzed in order to find the best technique for 33 latest version of open source projects. In this thesis study, the relation between CK metrics and reliability is also determined. Each technique was evaluated using RapidMiner and WEKA tools. Dataset was validated with a 10-fold cross-validation technique. Furthermore, Bayesian belief's networks form used for determining which CK metric are primary estimators. Receiver Operating characteristic (ROC), Precision, Accuracy, Area Under the Curve (AUC), Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) performance metrics used for evaluation of results. Results of this study show that Random Forrest, Bagging, AdaBoost ML techniques are the most effective for prediction models in terms of AUC values. In contrast, SVMs are the least effective models. This thesis study also revealed that Weighted Methods per class (WMC) is the most effective software metric. Then, the Number of Children (NOC), Depth of Inheritance Tree (DIT) metrics are good contributor for determining the quality of software.