Novel defectiveness factors for change level software defect prediction

Abstract

Recently, the concept of software quality has been emphasized as the importance of software grows. Software defect prediction is the technique that predict the probability of defect exist in each software entity. Based on prediction result, the project managers put more the testing or code review effort i.e. time and developer at more likely to defect prone. There are two main stream of defect prediction research, which module level and change level defect prone. Module level defect prediction (MLDP) is traditional defect prediction area. In MLDP, software entity defined as software consists of source code instances such as package, file or method. Various factors such as SLOC of file, number of previous fault, and number of developer used as defect indicator. Among of them, defect distribution factors, which represent defect of file and developer shown good performance in MLDP. These factors denoted as defect distribution in this paper. Change level defect prediction (CLDP) has studied since early 2000s. In CLDP, software entity defined as software consists of sequence of code change. Change factors such as lines added lines deleted and number of file changed used in CLDP. However, defect distribution factors has not been used widely in CLDP. Furthermore, the prediction performance of CLDP is low (Recall - 67%, Precision - 34%) [18] To improve the prediction accuracy of CLDP, two defectiveness factors, which represent defectiveness of file and developer are introduced. 9 open source project repository selected for evaluate our prediction model. Precision, recall, f1 score, area under curve selected as performance measure, and cross validation used as validation technique. Student t-test and cliff delta effect size test used for statistical evidence. Our experiment result indicates that defectiveness factors can improve the performance in 6 of 9 open source project.