Feature Selection and Software Defect Prediction by Different Ensemble Classifiers.

Software defect prediction can improve its quality and is actively studied during the last decade. This paper focuses on the improvement of software defect prediction accuracy by proper feature selection techniques and using ensemble classifier. The software code metrics were used to predict the defective modules. JM1 public NASA dataset from PROMISE Software Engineering Repository was used in this study. Boruta, ACE, regsubsets and simple correlation are used for feature selection. The results of selection are formed based on hard voting of all features selectors. A new stacking classifier for software defects prediction is presented in this paper. The stacking classifier for defects prediction algorithm is based on combination of 5 weak classifiers. Random forest algorithm is used to combine the predictions. The obtained prediction accuracy was up to 96.26%.