<![CDATA[The banking sector faces significant challenges in accurately classifying loan applications, particularly for datasets dominated by categorical features with imbalanced class distribution (71%:29%). This study proposes the application of CatBoost (Categorical Boosting) with Optuna-based hyperparameter optimization a Bayesian optimization framework using Tree-structured Parzen Estimator (TPE) for bank loan approval prediction. Two class imbalance handling scenarios are comparatively evaluated: SMOTE and CatBoost built-in class_weight. Experiments are conducted on a dataset of 381 samples with 15 active features (12 original features and 3 engineered features) using 5-fold stratified cross-validation. Results show LightGBM achieves the best overall performance with Accuracy 93.42%, Precision 94.83%, Recall 96.49%, F1-Score 95.65%, ROC-AUC 0.9215, and MCC 0.8220. CatBoost (SMOTE) achieves competitive performance with AUC-CV 0.9030 and F1 94.02%. SHAP (SHapley Additive exPlanations) analysis identifies Credit_History as the dominant feature (mean|SHAP|=3.2225), followed by ApplicantIncome (0.6896) and Property_Area_Semiurban (0.5222). This study contributes as the first investigation integrating CatBoost+Optuna+XAI-SHAP in retail bank loan approval prediction with dominant categorical features, while providing systematic comparison against LightGBM, XGBoost, and Random Forest.]]>
Copyrights © 2026
