<![CDATA[The paradigm of human resource management in the digital era demands an objective and data-driven employee promotion process. However, the extreme class imbalance (ratio 10.74:1) has the potential to introduce bias against minority groups that deserve promotion. This study proposes a stacking ensemble classification framework consisting of Random Forest, XGBoost, and LightGBM as base learners and Logistic Regression as a meta-learner, with the integration of SMOTE-ENN and two-level SHAP interpretability. This study shows that the application of SMOTE-ENN before cross-validation can result in a biassed performance estimate of up to +110% on the F1-Score; thus, the use of imblearn.Pipeline is proposed, which restricts resampling only to the training fold. Based on the evaluation using 10-fold stratified cross-validation free from data leakage, the stacking ensemble model achieved an accuracy of 0.9022, precision of 0.4342, recall of 0.4889, F1-score of 0.4598, and AUC-ROC of 0.8053. Although it did not achieve the highest F1 score, this model attained the best recall value among competitive models, making it relevant for contexts sensitive to false negative errors. SHAP analysis identifies avg_training_score, age, and performance_index as the main determinants of promotion decisions. The proposed framework provides a methodological contribution to model evaluation on imbalanced data while offering a more transparent and accountable decision support system to support the implementation of meritocracy in both government and corporate organisations.]]>
