Defect cost flow model: a Bayesian network for predicting defect correction effort

• Authors: Thomas Schulz;Łukasz Radliński;Thomas Gorges;Wolfgang Rosenstiel
• Venue: Proceedings of the 6th International Conference on Predictive Models in Software Engineering
• Year: 2010

Background. Software defect prediction has been one of the central topics of software engineering. Predicted defect counts have been used mainly to assess software quality and estimate the defect correction effort (DCE). However, in many cases these defect counts are not good indicators for DCE. Therefore, in this study DCE has been modeled from a different perspective. Defects originating from various development phases have different impact on the overall DCE, especially defects shifting from one phase to another. To reduce the DCE of a software product it is important to assess every development phase along with its specific characteristics and focus on the shift of defects over phases. Aims. The aim of this paper is to build a model for effort prediction at different development stages. Our model is mainly focused on a dynamic DCE changing from one development phase to another. It reflects the increasing cost of correcting defects which are introduced in early, but found in later development phases. Research Method. The modeling technique used in this study is a Bayesian network which, among many others, has three important capabilities: reflecting causal relationships, combining expert knowledge with empirical data and incorporating uncertainty. The procedure of model development contains a set of iterations including the following steps: problem analysis, data analysis, model enhancement with simulation runs and model validation. Results. The developed Defect Cost Flow Model (DCFM) reflects the widely used V-model, an international standard for developing information technology systems. It has been pre-calibrated with empirical data from past projects developed at Robert Bosch GmbH. The analysis of evaluation scenarios confirms that DCFM correctly incorporates known qualitative and quantitative relationships. Because of its causal structure it can be used intuitively by end-users. Conclusion. Typical cost benefit optimization strategies regarding the optimal effort spent on quality measures tend to optimize locally, e.g. every development phase is optimized separately in its own domain. In contrast to that, the DCFM demonstrates that even cost intensive quality measures pay off when the overall DCE of specific features is considered.