Analysis of Feature Ranking Techniques for Defect Prediction in Software Systems

Abstract

Software quality is an important parameter, and it plays a crucial role in software development. One of the most important software quality attributes is fault proneness. It evaluates the quality of the final product. Fault proneness prediction models must be built in order to enhance software quality. There are various software metrics which help in software modeling, but it is a cumbersome and time-consuming process to use all of them. So, there is always a need to select those set of software metrics which help in determining fault proneness. Careful selection of software metrics is a major concern, and it becomes crucial if the search space is too large. This study focuses on the ranking of software metrics for building defect prediction models. Hybrid approach is applied in which feature ranking techniques are used to reduce the search space along with the feature subset selection methods. Classification algorithms are used for training the defect prediction models. The area which is under the receiver operating characteristic curve is utilized for evaluating the performance of the classifiers. The experimental results indicate that most of the feature ranking techniques have almost similar results, and automatic hybrid search outperforms all other feature subset selection methods. Furthermore, the result helps us to focus only on those set of metrics which have almost the same impact on the end result as compared to the original set of metrics.