Robust cost-sensitive kernel method with Blinex loss and its applications in credit risk evaluation

Area of Science:

• Financial analysis
• Machine learning
• Computational finance

Background:

• Credit risk evaluation is vital for financial institutions and consumer fairness.
• Imbalanced learning, where non-default samples dominate, presents a significant challenge.
• Existing methods like cost-sensitive support vector machines (CSSVMs) have limitations in cost-sensitive generalization.

Purpose of the Study:

• To introduce a robust cost-sensitive kernel method with Blinex loss (CSKB) for credit risk evaluation.
• To address the limitations of traditional methods in handling imbalanced data and misclassification costs.
• To achieve a mutually beneficial outcome for financial institutions and consumers.

Main Methods:

• Developed a novel cost-sensitive kernel method (CSKB) utilizing the Blinex loss function.
• Implemented solutions for linear and nonlinear CSKB using Nesterov accelerated gradient and Pegasos algorithms.
• Provided theoretical analysis of the generalization capability of CSKB.

Main Results:

• CSKB demonstrates superior performance compared to benchmark methods across various datasets.
• The Blinex loss function provides asymmetry and boundedness, enhancing noise robustness and flexible cost control.
• CSKB achieves better cost-sensitive generalization than traditional CSSVMs.

Conclusions:

• CSKB offers a robust and effective data-driven approach for credit risk evaluation.
• The method successfully balances the needs of financial institutions and consumers.
• CSKB represents a significant advancement in applying AI to financial risk assessment.