Defect count prediction via metric-based convolutional neural network

With the increasing complexity and volume of the software, the number of defects in software modules is also increasing consistently, which affects the quality and delivery of software in time and budget. To improve the software quality and timely allocation of resources, defects should be detected at the initial phases of the software development life cycle. However, the existing defect prediction methodology based on high-dimensional and limited data only focuses on predicting defective modules. In contrast, the number of defects present in the software module has not been explored so far, especially using deep neural network. Also, whether deep learning could enhance the performance of defect count prediction is still uninvestigated. To fill this gap, we proposed an improved Convolutional Neural Network model, called metrics-based convolutional neural network (MB-CNN), which combines the advantages of appropriate metrics and an improved CNN method by introducing dropout for regularization between convolutions and dense layer. The proposed method predicts the presented defect count in the software module for homogeneous scenarios as within-version and cross-version. The experimental results show that, on average, across the fourteen real-world defect datasets, the proposed approach improves Li’s CNN architecture by 31% in within-version prediction and 28% in cross-version prediction. Moreover, the Friedman ranking test and Wilcoxon nonparametric test reveal the usefulness of our proposed approach over ten benchmark learning algorithms to predict defect count.