Comparing BDD and TDD: Machine Learning Analysis of Software Quality with SHAP Interpretability

This study evaluates the impact of Behavior-Driven Development (BDD) and Test-Driven Development (TDD) on software quality using machine learning models, including Random Forest, XGBoost, and LightGBM. Key metrics such as bug detection, test coverage, and development time were analyzed using a dataset from multiple software projects. Polynomial feature expansion captured non-linear interactions, while SHapley Additive exPlanations (SHAP) enhanced interpretability. Results indicate that Random Forest achieved the best predictive accuracy, with an average RMSE of 7.64 and MAE of 6.39, outperforming XGBoost (average RMSE: 8.63, MAE: 7.37) and LightGBM (average RMSE: 6.89, MAE: 5.38). However, negative values across all models reveal challenges in generalization. SHAP analysis highlights the critical influence of higher-order interactions, particularly between test coverage and development time. These findings underscore the complexity of predicting software quality and suggest the need for additional features and advanced techniques to enhance model performance. This study provides a comprehensive, interpretable framework for assessing the comparative effectiveness of BDD and TDD in improving software quality.