Effective software defect prediction with deep neural networks

• By leveraging learning machines, SDP models facilitate the quicker identification of defects, leading to faster resolution and better software quality. Redundant or irrelevant features, smaller data-size and Class Imbalance in defect datasets detriments the prediction power of SDP models. • This paper proposes a Deep Network based SDP model (DN_SDP) to investigate quantitatively how much Deep Learning (DL) alleviates the issue of close-knit between the concepts (i.e., feature learning, data-size and class imbalance) and classifier performance that existed in traditional machine learning SDP model (or Shallow Network based SDP model aka SN_SDP). • An empirical comparison between the traditional model and deep learning model is made considering the impact of Dataset Size (DS) and Class Imbalance (CI) on the predictors’ performance over 11 datasets from PROMISE repository and 8 evaluation metrics. A variant of DN_SDP (i.e., GAN+DN_SDP) is presented to empirically handle CI for deep model and compared with selected baselines. • Execution of a series of experiments and Friedman Test with Bonferroni post-hoc Test confirms that DN_SDP alleviates DS and CI issue occurring in SN-SDP and improves the AUC by 8 %, F-measure by 12 %, MCC by 42 %. The alliance of DN-SDP with GAN oversampling (i.e., GAN+DN_SDP) results in improvement of 10 %, 14 %, 25 % in AUC, F-measure and MCC respectively over DN-SDP. GAN_DN_SDP outperforms baselines and achieves 17.9 % rise in F-measure. Software Defect Prediction (SDP) allows the early identification for potentially defective modules by uncovering the patterns in the historical defect data that enables targeted and effective testing. It is crucial for large and complex software, where testing is expensive and time-consuming. By leveraging deep neural networks, SDP models facilitate the quicker identification of defects, leading to faster resolution and better software quality. This paper proposes a Deep Network based SDP model (DN_SDP) to investigate quantitatively how Deep Neural Network (DNN) performance differs from that existed in traditional machine learning SDP model (or Shallow Network based SDP model aka SN_SDP). A variant of DN_SDP (i.e., GAN+DN_SDP) is presented to empirically handle class imbalance for deep model and compared with selected baselines. An empirical comparison between the traditional model and deep learning model is made considering the impact of Dataset Size (DS) and Class Imbalance (CI) on the predictors’ performance over 11 datasets from PROMISE repository and 8 evaluation metrics. Execution of a series of experiments and Friedman Test with Bonferroni post-hoc Test confirms that DN_SDP alleviates DS and CI issue occurring in SN-SDP and improves the AUC by 8 %, F-measure by 12 %, MCC by 42 %. The alliance of DN-SDP with GAN oversampling (i.e., GAN+DN_SDP) results in improvement of 10 %, 14 %, 25 % in AUC, F-measure and MCC respectively over DN-SDP. GAN_DN_SDP outperforms baselines and achieves 17.9 % rise in F-measure.