A Comparative Analysis of Techniques, Datasets, Feature Selection Methods, and Evaluation Metrics in Software Fault Prediction

This study presents a systematic literature review (SLR) that investigates recent advancements in Software Fault Prediction (SFP) methodologies. The review focuses on key dimensions including techniques, datasets, feature selection methods, software metrics, and evaluation criteria. By analyzing significant studies from renowned digital libraries such as ACM, IEEE, Springer Link, and Science Direct, five research questions were defined to guide the assessment of current trends in SFP research. Findings reveal that machine learning approaches— particularly neural networks, deep learning, and ensemble methods—are increasingly employed due to their capability to manage the complexity of software fault data. Public datasets, notably those from the PROMISE and NASA MDP repositories, are widely utilized, underlining the importance of dataset diversity for enhancing model performance. Feature selection methods, particularly wrapper techniques, are often employed to improve predictive accuracy. Evaluation of models predominantly relies on confusion matrix-based metrics such as Accuracy, Precision, Recall, and F1-Score. Despite these advances, challenges remain in addressing class imbalance, adapting to rapidly evolving software environments, and achieving real-time fault prediction. The study highlights the need for greater classifier diversity and ongoing methodological improvements to enhance the robustness and generalizability of SFP models.