An empirical study on predicting defect numbers.

Defect prediction is an important activity to make software testing processes more targeted and efficient. Many methods have been proposed to predict the defect-proneness of software components using supervised classification techniques in within- and cross-project scenarios. However, very few prior studies address the above issue from the perspective of predictive analytics. How to make an appropriate decision among different prediction approaches in a given scenario remains unclear. In this paper, we empirically investigate the feasibility of defect numbers prediction with typical regression models in different scenarios. The experiments on six open-source software projects in PROMISE repository show that the prediction model built with Decision Tree Regression seems to be the best estimator in both of the scenarios, and that for all the prediction models, the results yielded in the cross-project scenario can be comparable to (or sometimes better than) those in the within-project scenario when choosing suitable training data. Therefore, the findings provide a useful insight into defect numbers prediction for those new and inactive projects.