A Systematic Review of Software Fault Prediction Using Deep Learning: Challenges and Future Perspectives

The accurate prediction of software faults is crucial in guaranteeing the quality and reliability of software systems. Recent research has highlighted the vast potential of various deep learning models, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTM) networks, autoencoders, deep belief networks (DBNs), and generative adversarial networks (GANs), in achieving this objective. In this extensive review, we have thoroughly examined the performance metrics and scalability of these models for software fault prediction, as well as their respective architectures. Additionally, we have evaluated frequently used datasets, such as NASA MDP and PROMISE, and addressed challenges such as insufficient data and model interpretability. Our analysis has revealed that the ANN-PSO and LSTM models are particularly effective in predicting software faults. Therefore, we recommend that future research prioritize enhancing software fault prediction using deep learning techniques.