Enhancing Explainability and Transparency in Machine Learning-Based Credit Scoring Models.

Abstract

This research addresses the critical challenge of opacity and potential bias in machine learning (ML) models used for credit scoring. While advanced models like Gradient Boosting Machines and Deep Neural Networks offer superior predictive accuracy, their "black box" nature hinders regulatory compliance, erodes stakeholder trust, and risks perpetuating discriminatory outcomes. This study employs a Design Science Research (DSR) methodology to design, develop, and evaluate a hybrid framework aimed at enhancing explainability and transparency. The framework integrates a suite of ML models, from interpretable baselines like Logistic Regression to complex ensembles, with leading Explainable AI (XAI) techniques, primarily SHAP (Shapley Additive Explanations) and LIME (Local Interpretable Model-agnostic Explanations). Using public credit datasets, the research demonstrates that simplistic performance metrics like accuracy are misleading in imbalanced credit data, and more robust metrics like F1-Score and ROC AUC are essential. The findings reveal that XAI techniques are highly effective not only for generating local and global explanations for loan decisions but also for debugging model behavior and identifying biases. The primary artifact is a functional, interactive prototype that translates complex model outputs into stakeholder-centric dashboards for data scientists, loan officers, and applicants. This work contributes a practical, integrated solution that bridges the gap between technical ML implementation and the pressing need for fair, transparent, and accountable AI in the financial services industry.