A novel deep neural network structure for software fault prediction

Software fault prediction is crucial to compute the potential occurrence of faults within the software components, before code testing or execution. Machine learning, especially deep learning, has been applied to predict faults, but both encounter challenges such as insufficient accuracy, imbalanced data, overfitting, and complex structure. Moreover, deep learning yields superior predictions when working with large datasets. The most common fault dataset is NASA MDP from the PROMISE repository. The BugHunter dataset, however, offers a larger number of instances compare to other fault datasets, leaving a gap in the literature for exploring the application of machine learning and deep learning. In this study, we present a novel structure of deep neural network (DNN), which utilizes the convolutional layers to extract valuable knowledge from the BugHunter data projects. Our proposed model addresses class imbalance and overfitting issues while accurately predicting fault-prone methods. To evaluate the effectiveness of our predictive model, we conduct extensive empirical studies comparing it with seven traditional machine learning, three ensemble learning, and three state-of-the-art deep learning baseline models. Our findings revealed that the proposed DNN structure significantly improved the average F1-score in 15 projects of the BugHunter datasets by 20.01%, indicating that DNN is a practical approach for predicting faulty methods. Leveraging these results could lead to the preservation of software development resources and the production of more reliable software.