Project · Research
Fuzzy-Monotonic LightGBM for Explainable Credit Default Prediction
An explainable credit-risk framework combining behavioral features, fuzzy linguistic features, monotonic gradient boosting constraints, and SHAP.
Research problem
Credit-risk modeling has a persistent trade-off: stronger nonlinear models can be harder to interpret, while highly interpretable systems can lose predictive power. This research explores a hybrid that keeps competitive performance while embedding domain assumptions and explainability.
Approach
The framework combines behavioral feature engineering, overlapping fuzzy membership regions, LightGBM, monotonic constraints, and SHAP. Domain priors can enforce economically sensible directions, such as delinquency and utilization not reducing predicted risk and stronger repayment behavior not increasing it.
Experiments use the Taiwan Credit Card Default dataset for primary modeling and ablations, with the German Credit dataset for an additional interpretability demonstration.
Evaluation and explainability
The recorded comparison uses ROC-AUC, PR-AUC, Brier Score, and KS Statistic. The fuzzy-monotonic model retained similar discrimination, slightly improved PR-AUC in that comparison, and kept competitive calibration.
Explainability comes from two layers: structural constraints and fuzzy features, followed by SHAP feature attribution. This is research and manuscript work, not a confirmed conference publication or a claim of regulatory compliance.