Multi-source domain adaptation with knowledge transfer for credit risk classification
Document Type
Research-Article
Journal Name
Engineering Applications of Artificial Intelligence
Keywords
Correlation alignment, Credit risk classification, Dempster–shafer evidence theory, Multi-source domain adaptation, Three-way decisions, Transfer learning
Abstract
In the field of financial risk management, credit risk classification faces significant challenges due to labeled data scarcity, data distribution discrepancy, and class imbalance. To address these critical issues, this paper proposes a novel multi-source domain adaptation approach that synergistically integrates weighted correlation alignment (CORAL), three-way decisions, and an improved Dempster–Shafer (DS) evidence theory. In the proposed approach, a class-weighted CORAL method is designed by adjusting the alignment strength on the basis of the inverse class frequency, enhancing feature alignment particularly for high-risk samples. Three-way decisions divide samples into positive, negative, and boundary regions via entropy-driven uncertainty quantification, deferring decisions to high-uncertainty cases. The DS evidence theory is improved by incorporating conflict reallocation weights to increase sensitivity to minority default samples. Experiments on listed company datasets outperform baseline methods across multiple evaluation metrics. This approach provides a multi-source domain adaptation solution for credit risk classification. © 2025 Elsevier Ltd.