Explainable Deep Learning Approaches to Credit Risk Evaluation

Abstract

Credit scoring is a critical tool for risk management in financial lending. Although traditional statistical and machine learning models have been extensively researched over the years, the trend has shifted towards deep learning models due to their enhanced performance across various fields. Despite the superior performance of deep learning models, their opaque nature poses challenges for their adoption in credit scoring. The inability to understand how these models arrive at their decisions is problematic, especially in fields where decisions are critical. Leveraging the credit dataset from Germany on the UCI repository, our approach integrates cutting-edge explanation techniques to ensure that predictions are accurate but also interpretable and justifiable. Our study implemented the following models: Convolutional Neural Network (CNN), Long Short-term Memory (LSTM), Restricted Boltzmann Machine (RBM), Autoencoder, Graph Neural Network (GNN), Multilayer Perceptron (MLP), and introduced two hybrid models one that combines MLP & RBM models and another that combines the LSTM & CNN models, achieving ROC-AUC scores of 76.3%, 78.7%, 69.1%, 60.3%, 62.3%, 75.7%, 75.5%, and 76.2% respectively. We also applied the following Explainable Artificial Intelligence (XAI) techniques: Layer-wise Relevance Propagation (LRP), SHapley Additive exPlanations (SHAP), Local Interpretable Model-agnostic Explanations (Lime), and Integrated Gradients. The application of LRP revealed significant feature influences across different layers, SHAP highlighted the positive and negative impacts of specific features on model predictions, while LIME offered detailed local interpretations, and Integrated Gradients quantified the contribution of each feature to the prediction decisions. Overall, these XAI methods have set a new benchmark for transparency and reliability, crucial for ethical AI deployments in financial decision-making.