Testing is considered as one of the expensive activities in software development process. Fixing the defects during testing process can increase the cost as well as the completion time of the project. Cost of testing process can be reduced by identifying the defective modules during the development (before testing) stage. This process is known as “Software Defect Prediction”, which has been widely focused by many researchers in the last two decades. This research proposes a classification framework for the prediction of defective modules using variant based ensemble learning and feature selection techniques. Variant selection activity identifies the best optimized versions of classification techniques so that their ensemble can achieve high performance whereas feature selection is performed to get rid of such features which do not participate in classification and become the cause of lower performance. The proposed framework is implemented on four cleaned NASA datasets from MDP repository and evaluated by using three performance measures, including: F-measure, Accuracy, and MCC. According to results, the proposed framework outperformed 10 widely used supervised classification techniques, including: “Naïve Bayes (NB), Multi-Layer Perceptron (MLP), Radial Basis Function (RBF), Support Vector Machine (SVM), K Nearest Neighbor (KNN), kStar (K*), One Rule (OneR), PART, Decision Tree (DT), and Random Forest (RF)”.

Software defect prediction is one of the emerging research areas of software engineering. The prediction of defects at early stage of development process can produce high quality software at lower cost. This research contributes by presenting a feature selection based ensemble classification framework which consists of four stages: 1) Dataset selection, 2) Feature Selection, 3) Classification, and 4) Results. The proposed framework is implemented from two dimensions, one with feature selection and second without feature selection. The performance is evaluated through various measures including: Precision, Recall, F-measure, Accuracy, MCC and ROC. 12 Cleaned publically available NASA datasets are used for experiments. The results of both the dimensions of proposed framework are compared with the other widely used classification techniques such as: “Naïve Bayes (NB), Multi-Layer Perceptron (MLP). Radial Basis Function (RBF), Support Vector Machine (SVM), K Nearest Neighbor (KNN), kStar (K*), One Rule (OneR), PART, Decision Tree (DT), and Random Forest (RF)”. Results reflect that the proposed framework outperformed other classification techniques in some of the used datasets however class imbalance issue could not be fully resolved.