A Strategy for Cross-Project Defect Prediction Models in Industry

Abstract

Defects are a costly concern within software development. A way to mitigate this is the use of cross-project defect prediction models that aid quality assurance by pinpointing defect-prone parts of the software. We have studied the ability to incorporate this kind of prediction models at an automotive company by developing such models. During this development, we applied metrics and modelling techniques from research to create cross-project models that would fill the needs of the company. We succeed in creating cross-project defect prediction models using data gathered from local data repositories. Our findings suggest that change-level metrics are of interest in the automotive sector since they allow for quick feedback. Random forest seem to result in the best predictive performance when training models, though these findings have to be considered in relation to a low performance for our models in comparison to similar studies. At the same time, we gathered interesting insight into which performance measurement is most valued by practitioners. These findings pose that high precision is requested by practitioners while a low recall is acceptable. This is in contrast to the view held by research within the field and should be considered in future studies.