Explainable Neural Network Models for Big Data- Driven Credit Risk Evaluation

The rapid digitalization of financial services has generated vast amounts of customer and transactional data, enabling more advanced approaches to credit risk evaluation. However, while deep learning models offer superior predictive accuracy compared to traditional statistical methods, their opaque "black-box" nature poses challenges for usage in regulated financial environments where transparency and fairness are essential. This study proposes an explainable deep learning framework for credit scoring that integrates multiple neural architectures-including MLP, CNN, LSTM, RBM, Autoencoders, GNN, and two hybrid models (MLP-RBM and LSTM-CNN). Using the German Credit dataset and extensive preprocessing through normalization and SMOTE balancing, the models were evaluated using accuracy, ROC-AUC, Precision-Recall AUC, and confusion matrices. Results show that the LSTM and hybrid LSTM-CNN models outperform other architectures, demonstrating strong classification ability across all metrics. To address the need for interpretability, several Explainable AI (XAI) techniques- Layer-wise Relevance Propagation (LRP), SHAP, Integrated Gradients, and LIME-were applied to reveal feature contributions and enhance model transparency. The XAI analyses highlight the importance of features such as account balance, credit duration, repayment patterns, and employment stability. Overall, this research confirms that combining big data analytics with explainable deep leaming provides a powerful, transparent, and ethically aligned solution for modern credit risk assessment.

کلیدواژه ها:

Credit Risk Evaluation ، Big Data Analytics ، Explainable Artificial Intelligence (XAI) ، Deep Learning