This study examines how traditional ensemble techniques and recent deep learning approaches perform in predicting employee attrition, particularly when working with small and imbalanced tabular datasets. In addition, it proposes an interpretable decision-support framework designed not only for technical evaluation but also for practical HR use. The analysis was conducted on the IBM HR Analytics dataset using a specially structured validation design to prevent data leakage. To address class imbalance, SMOTE was applied strictly within the training folds, without touching the validation data. The models compared include Random Forest, XGBoost, CatBoost, Artificial Neural Networks (ANN), and TabNet. Rather than relying on default probability thresholds, dynamic threshold adjustment was introduced to improve sensitivity to the minority class. Model performance was evaluated using Recall, Precision, F1-score, and confusion matrices, while SHAP analysis was employed to enhance interpretability and support transparent decision-making in HR contexts. To strengthen the reliability of the evaluation, a Stratified 5-Fold Cross-Validation scheme was adopted. The findings show that CatBoost produced the most balanced and consistent results, achieving a mean F1-score of 0.468 ± 0.053 together with a Brier score of 0.097. After dynamic threshold adjustment, TabNet demonstrated the highest sensitivity (Recall: 0.573 ± 0.064), making it particularly effective for early risk detection. According to the SHAP-based interpretation, OverTime and Stock Option Level emerged as the most influential predictors.
Cevher Özden (Mon,) studied this question.