Multi-Layer Perceptron with Dropout Regularization for Credit Scoring on Imbalanced Financial Datasets

This paper presents an enhanced Multi-Layer Perceptron (MLP) model with Dropout regularization for credit scoring on imbalanced financial datasets. The proposed model introduces three key innovations: an adaptive Dropout rate mechanism that adjusts regularization based on layer-wise feature importance, a balanced mini-batch sampling strategy for handling class imbalance ($12.3 \%$ default rate), and an optimized network architecture with skip connections. Experimental results on a comprehensive dataset of $\mathbf{8 5, 0 0 0}$ records demonstrate superior performance with an AUC-ROC of 0.892 and F1-score of 0.836, representing a 4.7% improvement over baseline methods. The model achieves $\mathbf{9 2 \%}$ precision in high-risk identification while reducing computational requirements by $35 \%$. The architecture incorporates SHAP values and LIME for model interpretability, revealing payment history and debt utilization ratio as the strongest predictors. Our comprehensive evaluation across different economic cycles validates the model’s effectiveness in real-world financial applications, particularly in identifying minority class samples while maintaining interpretability and computational efficiency. The implementation demonstrates remarkable resilience to input noise and feature corruption, with performance variations remaining within acceptable bounds under stress testing.