Analisis Komparatif Algoritma Tuned Random Forest dan Hist Gradient Boosting untuk Prediksi Niat Pindah Kerja Berdasarkan Indikator Beban Mental Karyawan

Abstract

Poorly managed mental workload in the digital transformation era contributes significantly to the employee turnover phenomenon, which compromises organizational stability. This study aims to conduct a comparative performance analysis of predictive models for job change intention using two ensemble learning algorithms: Random Forest (RF) and Hist Gradient Boosting (Hist GB). An empirical dataset of employee mental workload was processed through a series of data mining stages, including feature engineering to construct a work-hours-to-sleep ratio and a composite total mental load. The severe class imbalance challenge where the target class proportion stands at a 78% (retain) to 22% (turnover) ratio was addressed using the Synthetic Minority Over sampling Technique (SMOTE). The models were optimized via hyperparameter tuning and validated using 5-Fold Stratified Cross-Validation. Experimental results based on real program data demonstrate that the Base scenario yields a pseudo accuracy (~69%) with a very low F1-Score (~27%). The SMOTE intervention successfully rescued the models from majority class bias, with the SMOTE-optimized Random Forest outperforming in generalization stability, achieving a Mean CV Accuracy of 74.47%. On the other hand, the Optimized Hist Gradient Boosting produced the best minority class detection performance, reaching the highest Test F1-Score of 52.60% while maintaining an AUC Score of 0.68. Feature Rank analysis identified burnout score, deadline pressure score, and stress score as the three most crucial indicators in triggering job change intention