Research And System Construction On Software Defect Prediction

Software products have been closely related to our daily work and life,once there are defects in the software,it may bring serious consequences.Because the software itself is a logical entity,and the software product development process often involves personnel,technology,management,cost and other aspects,defects are often difficult to avoid.At present,the mainstream method to deal with defects is software testing technology,but with the vigorous development of software and information technology service industry,the scale of software is getting larger and larger and the complexity of software is getting higher and higher,which leads to the shortage of testing resources and the slow speed of testing.The cost of repairing defects increases exponentially with the extension of the time point in which defects are found,so it is particularly important to find defects as early as possible.Software defect prediction is a feasible method to deal with the above problems,mainly based on historical data to predict whether the software module has defects,this technology can improve the efficiency of test resource allocation.The supervised models need the labeled modules of historical data of the current project or external projects as the training set.According to the different sources of the training set,supervised defect prediction can be divided into wpdp(within-project defect prediction)and cpdp(cross-project defect prediction).In the two kinds of scenarios,the training set comes from the same project and other external projects respectively.This thesis mainly studies the new technologies based on machine learning to solve the different problems of the two kinds of defect prediction scenarios,such as the class imbalance issue and distribution difference between source project and target project,aiming to further improve the performance of defect prediction.The main contributions of this thesis are described as follows:(1)For the class imbalance issue,this thesis explores the impact of class imbalance on the performance of software defect prediction models.Additionally,we evaluate the performance of cost-sensitive models and ensemble models with class imbalance.Furthermore,we propose a software defect prediction method based on unbalanced rate by combining sampling and ensemble learning,which determines whether sampling is needed before constructing model according to imbalanced rate value.Finally,experiments are carried out on 42 datasets in PROMISE,and the results show that the method is effective.(2)In order to solve the problem of distribution difference between source project and target project,this thesis proposes a multi-source cross-project defect prediction method for defect severity marking.This method first builds a multi-source model to fully take advantage of knowledge recorded in all candidate source projects and preprocesses metric value to make feature distributions in source and target projects similar.Defect severity marking quantitatively describes the severity of the defect for each instance and subdividing the severity of defects can improve the testing efficiency.The experimental results show that this method can effectively improve the performance of cross-project defect prediction.(3)In the process of software defect prediction,datasets,model evaluation and model application are the focus of attention.A software defect prediction system is designed and implemented,hoping to analyze the datasets,evaluate the model and predict the instances to be tested.The application of machine learning technologies in the field of software engineering are expanded and new solutions to software defect prediction is provided,which is of great significance for software quality assurance.