Improving defect prediction by code clustering

Recently developed defect prediction models can help to find optimal testing strategies. Such models  work very well on projects that have a strong focus on single application functionalities (like InternetExplorer, NetMeeting or DirectX).  But it tends to drop the prediction accuracy while being applied on software systems which  functionality are much wider (e.g. customer relation or operating systems). In this thesis we try to automatically decompose software projects into clusters whereas each cluster contains software entities belonging to the same application layer.  To determine if such decomposition techniques will improve the accuracy of defect prediction models, we will implement and compare four decomposition strategies and compare their prediction results against each other and against an overall benchmark model.