Employee Turnover Prediction Model Based on Feature Selection and Imbalanced Data Handling

Employee turnover prediction is a critical task for organizations aiming to retain talent and reduce costs. In this study, we utilized the Kaggle HR dataset to predict employee turnover by employing a comprehensive machine learning pipeline. The dataset underwent rigorous preprocessing and exploratory data analysis (EDA) to identify key patterns and relationships. Feature selection was performed using correlation matrix analysis, Chi-Square tests, and Recursive Feature Elimination (RFE) to identify the most relevant features. To address the class imbalance issue, we applied both Synthetic Minority Over-sampling Technique (SMOTE) and Generative Adversarial Network (GAN-based) oversampling techniques. Eight models—Logistic Regression (LR), Naive Bayes (NB), Decision Tree (DT), Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Deep Neural Network (DNN), and a hybrid model (RF+DNN) —were evaluated under different scenarios, including before and after feature selection and imbalance treatment. Model performance was assessed using recall, F1-score, Cohen’s Kappa, ROC-AUC, PR-AUC, and confusion matrix. To evaluate potential overfitting, we compared performance across training and validation sets. Model generalization was further validated using cross-validation techniques. Results showed that the RF model, trained on GAN-balanced data and using RFE-selected only five features, achieved the best performance with an accuracy of 0.98, precision of 0.98, recall of 0.92, F1-score of 0.95, and Cohen’s Kappa of 0.94. XGBoost model also showed outstanding performance, with evaluation metrics closely matching those of RF. To validate the robustness and generalizability of our approach, the same methodology was applied to an external dataset—the IBM HR dataset. The application of RFE and GAN-based oversampling significantly improved minority class detection in this independent dataset. Notably, XGBoost and the hybrid RF+DNN model performed best, further demonstrating the effectiveness and generalization of the proposed framework across different datasets.