A dimensionality reduction-based efficient software fault prediction using Fisher linear discriminant analysis (FLDA)

Software quality is an important factor in the success of software companies. Traditional software quality assurance techniques face some serious limitations especially in terms of time and budget. This leads to increase in the use of machine learning classification techniques to predict software faults. Software fault prediction can help developers to uncover software problems in early stages of software life cycle. The extent to which these techniques can be generalized to different sizes of software, class imbalance problem, and identification of discriminative software metrics are the most critical challenges. In this paper, we have analyzed the performance of nine widely used machine learning classifiers-Bayes Net, NB, artificial neural network, support vector machines, K nearest neighbors, AdaBoost, Bagging, Zero R, and Random Forest for software fault prediction. Two standard sampling techniques-SMOTE and Resample with substitution are used to handle the class imbalance problem. We further used FLDA-based feature selection approach in combination with SMOTE and Resample to select most discriminative metrics. Then the top four classifiers based on performance are used for software fault prediction. The experimentation is carried out over 15 publically available datasets (small, medium and large) which are collected from PROMISE repository. The proposed Resample-FLDA method gives better performance as compared to existing methods in terms of precision, recall, f-measure and area under the curve.