Predicting Employee Turnover Using Explainable Machine Learning Models: A Comparative Study of XGBoost, LightGBM, CatBoost, and Deep Neural Networks

Abstract

Objective: This study aimed to develop and compare explainable machine learning models for employee turnover prediction using XGBoost, LightGBM, CatBoost, and Deep Neural Networks and to identify the most influential predictors of turnover risk through explainable artificial intelligence techniques.

Methods and Materials: This quantitative predictive analytics study was conducted using a dataset of 4,862 employees from 37 organizations across multiple industries in Canada. Employee demographic, organizational, behavioral, performance-related, and psychological variables were collected from human resource information systems and validated survey instruments. The dataset was preprocessed through normalization, feature engineering, categorical encoding, and missing-value imputation procedures. Four machine learning algorithms, including XGBoost, LightGBM, CatBoost, and Deep Neural Networks, were developed and optimized using grid search and Bayesian hyperparameter tuning. The dataset was divided into training, validation, and testing subsets using stratified sampling. Model performance was evaluated using accuracy, precision, recall, F1-score, area under the receiver operating characteristic curve (AUC-ROC), area under the precision-recall curve (AUC-PR), Matthews correlation coefficient, and balanced accuracy. Explainability was achieved through SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and partial dependence analyses.

Findings: The comparative analysis revealed that all machine learning models achieved strong predictive performance; however, CatBoost outperformed the other algorithms, achieving an accuracy of 91.9%, an F1-score of 0.878, and an AUC-ROC value of 0.958. XGBoost demonstrated the second-highest performance with an accuracy of 91.3% and an AUC-ROC of 0.951, followed by LightGBM (accuracy = 90.6%, AUC-ROC = 0.944) and the Deep Neural Network (accuracy = 89.7%, AUC-ROC = 0.936). ROC curve analyses confirmed the superior discriminative ability of CatBoost. SHAP-based explainability analyses identified job satisfaction, employee engagement, organizational commitment, perceived organizational support, promotion frequency, salary growth, performance ratings, overtime hours, absenteeism, and training participation as the most influential predictors of employee turnover. The confusion matrix of the CatBoost model further demonstrated strong classification capability with high true positive and true negative rates and minimal classification errors.

Conclusion: The findings demonstrate that explainable machine learning models can accurately predict employee turnover and provide meaningful insights into the organizational, behavioral, and psychological factors associated with employee departure. Among the evaluated algorithms, CatBoost offered the optimal balance between predictive accuracy and interpretability. The integration of explainable artificial intelligence techniques enhanced transparency and practical applicability, enabling organizations to identify high-risk employees and implement proactive retention strategies. These results support the adoption of explainable machine learning as an effective decision-support tool for workforce analytics, talent management, and organizational sustainability.