Effective fault prediction model developed using Least Square Support Vector Machine (LSSVM)

Software developers and project teams spend considerable amount of time in identifying and fixing faults reported by testers and users. Predicting defects and identifying regions in the source code containing faults before it is discovered or invoked by users can be valuable in terms of saving maintenance resources, user satisfaction and preventing major system failures post deployment. Fault prediction can also improve the effectiveness of software quality assurance activities by guiding the test team to focus efforts on fault prone components. The work presented in this paper involves building an effective fault prediction tool by identifying and investigating the predictive power of several well-known and widely used software metrics for fault prediction. We apply ten different feature selection techniques to choose the best set of metrics from a set of twenty source code metrics. We build the fault prediction model using Least Squares Support Vector Machine (LSSVM) learning method associated with linear, polynomial and radial basis function kernel functions. We perform experiments on 30 Open Source Java projects. Experimental results reveals that our prediction model is best suitable for projects with faulty classes less than the threshold value depending on fault identification efficiency (low- 52.139%, median- 46.206%, and high- 32.080%).