Human errors during software development lead to many defects, which emphasizes the importance of early detection and minimization. However, existing approaches often fall short in delivering accurate, scalable, and generalizable predictions due to challenges such as class imbalance, feature extraction limitations, and computational inefficiencies. This study proposes a hybrid method using a Convolutional Neural Network (CNNs) + Long Short-Term Memory (LSTM) for feature extraction, addressing class imbalance with Adaptive Synthetic Sampling (ADASYN) and subsequent training using Extreme Gradient Boosting (XGboost), to predict software defects. The proposed approach was evaluated on five publicly available datasets (CM1, MC1, KC1, PC1, and PC4) and compared with state-of-the-art (SOTA) models. Experimental results demonstrated that the hybrid model significantly outperforms traditional XGBoost-based models in terms of recall, F1-score, and area under the receiver operating characteristic curve (AUC), addressing the shortcomings of existing methods. Results demonstrate the effectiveness of the proposed method, with notable performance metrics achieved across all datasets. For example, on the MC1 dataset, the model attained an accuracy of 0.9980, a precision of 0.9971, a recall of 0.9988, an F1-score of 0.9980, and an AUC-ROC of 0.9999. On the KC1 dataset, it achieved an accuracy of 0.9344, a precision of 0.9265, a recall of 0.9375, an F1-score of 0.9320, and an AUC-ROC of 0.9839. The model achieves better performance than traditional machine learning methods and separate deep learning models, especially in the areas of recall and AUC-ROC. This research presents a robust solution through hybrid approaches that address class imbalance and maintain high predictive accuracy for software development process tasks, offering insights into the trade-offs between machine learning and deep learning methods.
Haruna et al. (Wed,) studied this question.