Thesis (M.Sc.) - Cairo University - Faculty of Computers and Artificial Intelligence - Department of Computer Sciences

Many software projects are shipped to customers containing defects. Defective software cost money, time, and lives. To reduce this harm, software companies allo- cate testing and quality assurance budgets. The enormous sizes of modern software pose challenges to traditional testing approaches due to the need for scalability. De- fect prediction models have been used to direct testing efforts to probable causes of defects in the software. Early approaches for software defect prediction relied on sta- tistical approaches to classify software modules and decide whether each module is a defect-prone module or not. Lately, many researchers used machine learning tech- niques to train a model that can classify software modules to defect-prone modules and not defect-prone modules. Starting from the new millennium, neural networks and deep learning won many competitions in machine learning applications. However, the use of deep learning to build a software defect prediction model was not inves- tigated thoroughly. In this study, a deep neural network is used to build a software defect prediction model and compared our proposed model with other machine learn- ing algorithms like random forests, decision trees, and Naive Bayesian networks. In addition, the usage of feature selection and class imbalance handling techniques are investigated to enhance the models{u2019} prediction quality. The result shows an improve- ment for our proposed over the other learning models by 3.55% on average if handling class imbalance issue is infeasible

Issued also as CD