Impact of attribute selection on defect proneness prediction in OO software

Defect proneness prediction of software modules always attracts the developers because it can reduce the testing efforts as well as software development time. In the current context, with the piling up of constraints like requirement ambiguity and complex development process, developing fault free reliable software is a daunting task. To deliver reliable software, software engineers are required to execute exhaustive test cases which become tedious and costly for software enterprises. To ameliorate the testing process one can use a defect prediction model so that testers can focus their efforts on defect prone modules. Building a defect prediction model becomes very complex task when the number of attributes is very large and the attributes are correlated. It is not easy even for a simple classifier to cope with this problem. Therefore, while developing a defect proneness prediction model, one should always be careful about feature selection. This research analyzes the impact of attribute selection on Naive Bayes (NB) based prediction model. Our results are based on Eclipse and KC1 bug database. On the basis of experimental results, we show that careful combination of attribute selection and machine learning apparently useful and, on the Eclipse data set, yield reasonable good performance with 88% probability of detection and 49% false alarm rate.