EVALUATION OF MACHINE LEARNING CLASSIFICATION TECHNIQUES IN PREDICTING SOFTWARE DEFECTS

Abstract

The advances in technology has brought about a significant rise in the number of software being developed and deployed on daily basis. This has brought about more dependencies on software and any defect in the software can lead to calamitous issues due to the type of data stored on the software or the type and importance of the functions the software perform. It is necessary to make sure that the all the defects are properly identified before deployment for use. The purpose of this study is to evaluate classifiers on software defects dataset and recommend appropriate classifier for defective software prediction. This will save the developer the stress and time of searching for the defects all through the program code, which will in turn lead to the software to be free from defects that can cause problems in the future of the use of the software. In this research, six categories of machine learning algorithms (two from each category) were tested in Waikato Environment for Knowledge Analysis (WEKA) which are; Bayes (Naives Bayes and Bayes Net), Functions (Multilayer Perceptron (MLP) and Sequential Minimal Optimization (SMO)), Lazy (IBK and KStar), Meta (Random Committee and Bagging), Rules (Decision Table and JRip) and Trees (J48 and Random Forest). The PROMISE dataset was used and the performance metric recorded were; accuracy, false positive rate, precision, recall, f-measure, Receiver Operating Curve (ROC), Kappa Statistics and Root Mean Square Error (RMSE). It was observed that Random Forest performs better under the 10 folds cross validation than the algorithms tested having an Accuracy of 0.818, a Recall of 0.818, a F-measure of 0.787, a ROC of 0.755 and a RMSE of 0.3669