SMoBaT: Improving Software Defect Prediction via Balanced Ensemble Learning and SHAP Interpretability

Software Defect Prediction (SDP) is crucial in improving software quality by identifying defect-prone modules early in the development lifecycle. However, class imbalance in most SDP datasets often biases machine learning models toward the majority class, reducing their ability to detect actual defects. To address this challenge, we propose SMoBaT. This hybrid framework integrates the Synthetic Minority Over-sampling Technique (SMOTE) for data balancing with Bagged Decision Trees for robust and stable prediction. SMoBaT is evaluated on two widely used benchmark datasets, KC1 and KC2, and demonstrates superior performance compared to traditional classifiers and recent state-of-the-art models. Notably, SMoBaT achieves a test accuracy of 96.77% on the KC1 dataset and 93.85% on KC2, outperforming prior methods in precision, recall, and F1-score. Additionally, we incorporate Shapley explanations (SHAP) analysis to interpret model predictions and identify the most influential software metrics contributing to defects. The results validate the effectiveness of SMoBaT in prediction accuracy and model transparency, highlighting its potential for practical deployment in software quality assurance workflows.