Leakage-Free Evaluation for Employee Attrition Prediction on Tabular Data

In the context of employee attrition prediction using imbalanced tabular data, we propose a reproducible, leakage-aware evaluation protocol and validate it on the IBM HR Attrition dataset. We perform the train/test split prior to any rebalancing; SMOTE (Synthetic Minority Over-sampling Technique) is applied exclusively within the training portion of each fold in stratified 5-fold cross-validation, while the test set remains untouched. One-Hot Encoding is performed consistently using pd.get_dummies. We benchmark Logistic Regression, Random Forest, ExtraTrees, LightGBM, and XGBoost using imbalance-aware metrics: F1 for the minority class, PR-AUC reported as Average Precision (AP), and ROC-AUC reported both in cross-validation and on the held-out test set. XGBoost attains the best mean AP in cross-validation (0.556 ± 0.056). Logistic Regression achieves the highest mean F1 (0.439 ± 0.048), while LightGBM yields the best mean ROC-AUC (0.791 ± 0.026). On the test set, XGBoost achieves a precision value of 0.65 and a recall value of 0.45 at a fixed threshold of 0.5. Overall, the results highlight a trade-off between stable minority-class detection (Logistic Regression) and stronger risk ranking performance (boosting models) under class imbalance.