LAUSR

Software defect prediction is one of the most important quality assurance activities during software development. This paper contributes empirical insights into the effectiveness of three resampling ensemble methods (bagging, boosting, and dagging) of Deep Learning Neural Networks (DLNN) for cross-project software defect prediction, compared to individual DLNN. An empirical study was conducted using five datasets. The results indicate that the bagging ensembles of DLNN offer only 0.24% increase in accuracy, on average, compared to the individual DLNN models, whereas the boosting and dagging ensembles degrade the accuracy. Furthermore, the results show that the three resampling ensembles of DLNN outperform the individual DLNN models in precision; with a maximum improved precision by 25.15% on average using the boosting ensembles. The results however indicate that none of the resampling ensembles improve the recall. Lastly, for a balanced performance in terms of both precision and recall, the results indicate improvements ranging from 0.98% on average by applying the bagging ensembles to 11.67% on average by applying the boosting ensembles.