Software Defect Prediction Based On Cost-Sensitive Bayesian Network

Computer software system has become more and more huge and complex with its applications in various fields. Software defect prediction plays an important role in the software development life cycle. It is to detect and correct deficiencies in the development process, which will help avoid the problems increasing in a geometric scale at the latter part of the software development. A good prediction method can shorten the software development cycle, improve software quality and reduce development costs. Therefore, the software defect prediction is with great significance for technology development of the computer software.Both the traditional static metrics and machine learning models are considered to be the most popular tools for software defect prediction methods. However, there exists a large amount of the static metrics, and the characteristics of software described by each metric is often rather complex. As far, there has not been a uniform standard for metrics selection scheme, which makes it a major challenge to software defect prediction. In addition, the development of software technology makes the issues even more serious. As the new software defects emerge, many traditional metrics become obsolete. On the other hand, while the traditional machine learning models have many successful applications in the other pattern recognition problems, it is difficult to apply the machine learning models in software defect prediction for the uncertainly of software defects. Therefore, both the static metrics selection and the application of machine learning model are the rather difficult problems in the defect prediction field.This thesis proposes to employ the information gain ratio method for metrics selection, and designs two new efficient metrics based on the quality of test code and code design principles. Besides, a new dynamic cost-sensitive Bayesian Network is constructed for software defect prediction. The experiments are performed on the classic Promise data Repository data set. The experimental results show that the metric pool selected by using information gain ratio and with two new metrics can make full use of a small number of features to describe the critical information in the process of software development. Taking the famous criteria F value and AUC value as evaluation standard, the new dynamic cost-sensitive Bayesian Network is more efficient and accurate than the traditional machine learning model, such as support vector machine, random forest and C4.5 and so on.