Comparative Evaluation of Machine Learning Models with Class Imbalance Techniques for Employee Turnover Prediction
TL;DR
Overall, integrating cost-sensitive learning, threshold optimization, and explainability improved model interpretability and practical utility for evidence-based HR decision-making processes in employee retention management and planning.
Abstract
Employee turnover prediction remains challenging in Human Resource (HR) analytics because class imbalance can reduce the ability of machine learning models to identify employees at genuine risk of leaving. This study develops and evaluates a comprehensive machine learning framework that balances minority-class detection and false-positive control. A publicly available HR dataset containing demographic, organizational, performance, and training-related attributes was analyzed using seven algorithms: Logistic Regression, Support Vector Machine, Multilayer Perceptron, Random Forest, XGBoost, LightGBM, and CatBoost. Cost-sensitive learning and three resampling methods, SMOTEENN, ADASYN, and Tomek Links, were compared through stratified 10-fold cross-validation. Performance was evaluated using ROC-AUC, PR-AUC, Balanced Accuracy, Matthews Correlation Coefficient, G-Mean, Sensitivity, and Specificity, followed by threshold adjustment and SHAP analysis. Original LightGBM achieved the highest discrimination performance (ROC-AUC = 0.5975 ± 0.0546; PR-AUC = 0.2020 ± 0.0426), while cost-sensitive LightGBM produced the most balanced results (Balanced Accuracy = 0.5221 ± 0.0303; MCC = 0.0499 ± 0.0685). SHAP identified Department Type, Current Employee Rating, Training Cost, and Age as key predictors. Overall, integrating cost-sensitive learning, threshold optimization, and explainability improved model interpretability and practical utility for evidence-based HR decision-making processes in employee retention management and planning.