Explainable Ensemble Deep Learning for Credit Risk Classification

Abstract

Although eXplainable Artificial Intelligence (XAI) has significantly increased interest in recent years, its application in credit risk classification (CRC) still requires further exploration to evaluate its effectiveness in classifying loan applicants. In credit risk classification, the application of deep learning models requires an explanation of their outputs to make the decision-making process transparent. However, existing explainable credit risk classification models often use a single classifier, which can degrade the prediction accuracy and reduce the quality of explanations. In order to improve both explainability and accuracy, we design a new explainable ensemble deep learning model for credit risk classification (EED-CRC). The proposed model involves two main phases: model building and model interpretation. In the first phase, we propose an ensemble CRC model that uses three Convolutional Neural Network (CNN) architectures to classify applicants. The outputs of the three CNN architectures are then integrated using a machine learning model, yielding more accurate predictions than those obtained from any individual classifier. In the second phase, we aim to interpret the obtained classification model through three XAI methods, such as Local Interpretable Model-agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), and Permutation Feature Importance (PFI), providing valuable insights to decision makers for classifying applicants. Experimental results on three real credit risk datasets show that the EED-CRC approach achieves superior performance compared to traditional CRC methods, both in accuracy and in explainability.