Credit Scoring Enhancement via Ensemble Learning and Self-Organizing Map-Based Feature Transformation

Credit scoring is a fundamental component of financial risk assessment, yet conventional modeling techniques often exhibit limited effectiveness in addressing critical challenges such as class imbalance and complex, non-linear feature interactions. To overcome these constraints, this study introduces a novel ensemble-based credit scoring framework designed to enhance both predictive performance and model robustness across heterogeneous financial datasets. Central to the proposed methodology is the application of self-organizing maps (SOMs) for unsupervised feature transformation. This technique systematically restructures the input space to preserve topological and clustering properties, thereby capturing latent structural relationships within the data. SOM hyperparameters are meticulously tuned using grid search to ensure optimal representational fidelity. The transformed feature set is subsequently integrated into a stacked ensemble architecture, wherein diverse base learners are combined via a meta-learner to exploit model complementarity. To mitigate the effects of class imbalance and optimize decision boundaries, an F1-score-driven threshold calibration strategy is employed. The framework is empirically validated on five benchmark credit scoring datasets, achieving classification accuracies of 84.00% (German), 92.03% (Australian), 87.68% (Japanese), 99.00% (Polish~1), and 96.77% (Taiwan), consistently outperforming baseline models. These results underscore the robustness and efficacy of the proposed approach in complex credit risk environments.