Testing is one of the crucial activities of software development life cycle which ensures the delivery of high quality product. As software testing consumes significant amount of resources so, if, instead of all software modules, only those are thoroughly tested which are likely to be defective then a high quality software can be delivered at lower cost. Software defect prediction, which has now become an essential part of software testing, can achieve this goal. This research presents a framework for software defect prediction by using feature selection and ensemble learning techniques. The framework consists of four stages: 1) Dataset Selection, 2) Pre Processing, 3) Classification, and 4) Reflection of Results. The framework is implemented on six publically available Cleaned NASA MDP datasets and performance is reflected by using various measures including: F-measure, Accuracy, MCC and ROC. First the performance of all search methods within the framework on each dataset is compared with each other and the method with highest score in each performance measure is identified. Secondly, the results of proposed framework with all search methods are compared with the results of 10 well-known supervised classification techniques. The results reflect that the proposed framework outperformed all of other classification techniques.

Predicting the defects at early stage of software development life cycle can improve the quality of end product at lower cost. Machine learning techniques have been proved to be an effective way for software defect prediction however an imbalance dataset of software defects is the main issue of lower and biased performance of classifiers. This issue can be resolved by applying the re-sampling methods on software defect dataset before the classification process. This research analyzes the performance of three widely used resampling techniques on class imbalance issue for software defect prediction. The resampling techniques include: “Random Under Sampling”, “Random Over Sampling” and “Synthetic Minority Oversampling Technique (SMOTE)”. For experiments, 12 publically available cleaned NASA MDP datasets are used with 10 widely used supervised machine learning classifiers. The performance is evaluated through various measures including: F-measure, Accuracy, MCC and ROC. According to results, most of the classifiers performed better with “Random Over Sampling” technique in many datasets.