Revolutionizing Software Defect Prediction Through Deep Learning

This study aims to revolutionize software defect prediction by leveraging deep learning (DL) techniques, specifically focusing on Convolutional Neural Networks (CNN) and Stack Sparse Autoencoders (SSAE). The research involves training these models on datasets from the NASA Metrics Data Program, using metrics such as accuracy, precision, detection rate, and True Negative Rate (TNR) to evaluate their performance. The proposed methodology includes normalization of data, application of neural network architectures, and extensive experimentation with varying parameters. Results demonstrate that CNN outperforms SSAE, achieving a higher accuracy range of 0.84 to 0.93 compared to SSAE’s 0.80 to 0.90, particularly excelling on the PC1 dataset with an accuracy of 0.93. Both models, however, show strong capabilities in predicting software defects, with CNN consistently delivering better performance across diverse datasets. The study concludes that DL models, especially CNN, significantly enhance the efficiency and accuracy of software defect prediction, suggesting future research to explore additional DL techniques and larger datasets for further advancements in software quality assessment.