SOFTWARE DEFECT PREDICTION: HEURISTICS FOR WEIGHTED NAÏVE BAYES

Abstract

Defect prediction is an important topic in software quality research. Statistical models for defect prediction can be built on project repositories. Project repositories store software metrics and defect information. This information is then matched with software modules. Naive Bayes is a well known, simple statistical technique that assumes the ‘independence’ and ‘equal importance’ of features, which are not true in many problems. However, Naive Bayes achieves high performances on a wide spectrum of prediction problems. This paper addresses the ‘equal importance’ of features assumption of Naive Bayes. We propose that by means of heuristics we can assign weights to features according to their importance and improve defect prediction performance. We compare the weighted Naive Bayes and the standard Naive Bayes predictors’ performances on publicly available datasets. Our experimental results indicate that assigning weights to software metrics increases the prediction performance significantly.