.png){kind=icon}
Return
Explainable multi-scale deep feature fusion for credit default prediction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.compeleceng.2026.111482.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate credit default prediction is crucial for sound financial risk management and informed lending decisions. Despite advances in existing machine learning (ML) and deep learning (DL) methods for credit scoring, their practical impact is constrained by limited hybrid ML-DL integration, inadequate multi-scale feature learning, class imbalance mitigation, and insufficient model explainability. Therefore, we propose an enhanced dilation convolutional neural network (ED-CNN) that unifies multiscale representation learning and interpretable feature refinement within a unified feature extraction framework. ED-CNN employs multi-branch convolutional pathways and dilated residual blocks to capture diverse non-linear dependencies among financial indicators. An attention-driven global context fusion mechanism dynamically prioritize behaviorally and structurally credit attributes. Furthermore, generative adversarial network is embedded to generate realistic minority-class (default) samples and alleviate class imbalance. Four feature selection techniques, including random forest importance, principal component analysis (PCA), least absolute shrinkage and selection operator (LASSO) regularization, and shapley additive explanations (SHAP), are utilized to refine the extracted ED-CNN embeddings. The selected deep features are then combined with the original financial variables and fed to multiple ML classifiers. The proposed framework achieved area under the curve scores of 0.9078 and 0.9053 on two benchmark credit default datasets. Notably, when original features are fused with PCA- or SHAP-refined ED-CNN embeddings, extreme gradient boosting yields the best overall performance. SHAP analysis further shows that the learned decision logic aligns with established financial risk indicators. The proposed framework advances toward regulator-aligned artificial intelligence solutions for credit risk management.
Keywords:
Decision support systems
Credit default classification
Enhanced dilation
Generative adversarial networks
Feature fusion
Journal
C
IF:
4.9
Papers:
250
Citations:
0
Organization
Cited Papers
No cited papers available