AugLog-LightGBM: A Log-Based Feature AugmentationFramework for Class Imbalance in Credit RiskClassification

Abstract

Non-performing loan (NPL) detection is inherently a class-imbalance problem because defaulting borrowers represent a persistent minority. Standard gradient boosting often favors the majority class. This paper proposes AugLog-LightGBM, an extension of LightGBM that improves initialization through Log-Based Feature Augmentation (LBFA). Instead of using an uninformative constant, boosting starts from an informed prior combining a logistic-regression logit score and a kernel-density-estimation log-density ratio (LDR), which capture complementary global and local information. These representations are incorporated as augmented features and as the init_score, reformulating boosting as residual correction over an informed Bayesian prior. The proposed framework is evaluated on a dataset of 2,700 home-mortgage borrowers collected from partner banks in Malang, Indonesia (NPL rate = 16.11%), using repeated stratified cross-validation and comparison against four imbalance-aware baselines. AugLog-LightGBM achieves the highest ROC-AUC (0.815 ± 0.018), PR-AUC (0.679), and F1-score (0.631). DeLong tests show statistically significant ROC-AUC improvements over class-weighted Logistic Regression and Random Forest, while gains over XGBoost and SMOTE + LightGBM are positive but not statistically significant. Robustness analyses and SHAP interpretation further support the consistency and practical applicability of the proposed framework.