Recursive Clustering and SHAP-Based Interpretability for Rural Bank Credit Prediction in Indonesian Provinces

This study uses a data-driven approach to investigate the clustering of Indonesian provinces to predict total rural bank credit. Thirteen financial variables, sourced from rural and regional bank data, were employed to identify regional patterns and form clusters of provinces with similar characteristics. Recursive clustering techniques were utilized to prevent imbalance in cluster formation. Four machine learning models—LightGBM, XGBoost, Linear Regression, and their stacking ensembles—were applied to predict rural bank credit within each cluster. SHAP (SHapley Additive exPlanations) values were utilized to interpret the best-performing models, offering insights into the relative importance of predictive features. The findings from SHAP revealed relevant results consistent with real-world observations, highlighting key factors such as third-party funds and rural bank assets that influence rural bank credit. Integrating recursive clustering techniques with interpretable machine learning models provides practical implications for policymakers and stakeholders in financial and regional planning.