DeepDefectNet: Estimating software defects with SMOTE and mRMR

Software defect estimation is crucial for improving software quality and reducing development costs. In this work, we propose a deep learning-based framework with joint feature selection and data balancing to improve defect prediction on cross-software systems. In contrast to previous methods that run these techniques separately, the proposed technique integrates “Minimum Redundancy Maximum Relevance (mRMR)” to extract relevant features and “Synthetic Minority Over-sampling Technique (SMOTE)” to mitigate the class imbalance (up to around 1:10 defect-to-non-defect ratio). A predictive model named “Deep Neural Network (DNN)” is employed to model the complicated nonlinear relationship in software metrics. The framework is tested on ten benchmark datasets from the NASA PROMISE repository, which consists of safety-critical and varied software systems, showing good generalization and performance. The experimental results of the model are promising, obtaining accuracy above 90% on the multiple datasets with the highest accuracy of 96.2% on MC1 dataset. The model is further evaluated using multiple performance metrics, including Precision, Recall, F1-score, and AUC, to ensure robust and reliable prediction. The comparative analysis confirms improved predictive performance over existing methods. The study demonstrates that integrating feature selection, imbalance handling, and deep learning provides an effective and practical solution for software defect estimation.