An interpretable credit risk assessment model with boundary sample identification.

Area of Science:

• Credit Risk Modeling
• Machine Learning
• Data Science

Background:

• Interpretability is vital for trustworthy and compliant credit risk assessment models.
• Distinguishing noise from boundary samples is crucial for enhancing prediction accuracy.

Purpose of the Study:

• Introduce a novel, interpretable credit risk assessment model (IAIBS).
• Improve credit risk prediction accuracy by effectively handling noise and boundary samples.

Main Methods:

• Utilize a logistic regression sub-model for interpretability.
• Employ the ARPD algorithm to identify and filter noisy/boundary samples.
• Train a deep learning sub-model on boundary samples and use clustering for final prediction.

Main Results:

• The IAIBS model significantly outperformed 11 baseline models across four public datasets.
• Achieved high Area Under the Curve (AUC) scores, demonstrating strong predictive performance.
• Exhibited strong generalization ability and positive contributions from each model module.

Conclusions:

• The IAIBS model offers superior performance and interpretability in credit risk assessment.
• Effective identification of boundary samples enhances prediction accuracy.
• The model provides clear interpretations of key predictors and outcomes.