<![CDATA[Japan's demographic crisis has increased demand for the Technical Intern Training Program (TITP). However, for Sending Organizations (SOs) in Indonesia, this process carries high financial risk due to an upfront talent funding scheme, where significant costs (up to IDR 35,000,000) are paid in advance. Trainee failure (dropouts or runaways) leads to substantial bad debt. This research aims to develop and validate a robust machine learning model for risk mitigation. We compare XGBoost and Random Forest on a dataset of 784 historical trainee records, characterized by extreme class imbalance (75.5% majority class). To address prior methodological weaknesses and prevent data leakage, we implement a 10-fold stratified cross-validation pipeline incorporating StandardScaler and SMOTE. The results show XGBoost (mean macro F1-Score: 0.5470 ± 0.15) significantly outperforms Random Forest (mean macro F1: 0.5098 ± 0.15), which is confirmed as statistically significant (p=0.0384) by a paired t-test. Furthermore, SMOTE is validated as a superior imbalance strategy compared to class_weight (p=0.0076). SHAP analysis identified 'contract duration' and lifestyle factors (e.g., 'alcohol consumption') as key predictors. The final model effectively predicts 'Runaway' cases (F1=0.533) but struggles with 'Training Dropouts' (F1=0.170), indicating a key limitation and a need for temporal features in future work.]]>
Copyrights © 2025
