Software Defect Modeling And Prediction In Resource-constrained Scenarios

Software has become increasingly widely used and more complex.So the software quality assurance has become increasingly important.With the rise of the Internet,rapid development has been widely appreciated,in order to publish the product as soon as possible.Thus,it is particularly important to detect defects in a limited time.Software defect prediction can aid testers to emphasize on the defect-prone modules,and thus accelerates testing procedure.However,traditional defect prediction methods usually lead to the developers,testers and modeling waiting for each other,which is a waste of time.So the traditional defect prediction methods are not appropriate for rapid development.This thesis studies the problems mentioned above,and get three major results as follows:First,traditional software defect prediction methods are not appropriate for real-time prediction and modeling.Thus,this thesis proposed a framework for real-time defect prediction,which is called Depred-RT.Depred-RT enables the development,testing and modeling to be carried out simultaneously,and reduced the waiting time of the developers,testers and modeling,and thus speeds up the software development process.Second,when testing resources is limited,testers have to only test a minority part of modules,which means the labeled instances are a small portion of the training data.On the one hand,the lack of labeled instances may cause the low predictive performance of the prediction model.This thesis applied online semi-supervised learning algorithm OMike to handle this issue.On the other hand,the testers have to test effectively,in order to avoid the low predictive performance of the prediction model.Thus this thesis proposed an online active semi-supervised learning algorithm AcOMike,which can actively select highly informative modules by interacting with testers and achieve better performance than passive methods.Empirical studies show that AcOMike is effective and efficient for handling the problems mentioned above.Third,this thesis designs and implements a defect prediction prototype system called DPS-RT1.0,which implements the framework mentioned above into a real system.Furthermore,DPS-RT1.0 provides functions like defect prediction,model updating,test result feedback and so on.