Enhancing Enterprise Credit Risk Assessment with Cascaded Multi-level Graph Representation Learning

Area of Science:

• Computational Finance
• Graph Machine Learning
• Risk Management

Background:

• Enterprise Credit Risk (ECR) assessment is vital for investment and regulation.
• Traditional ECR methods rely on accessible credit indicators, often unavailable for SMEs.
• Existing graph learning methods for ECR overlook complex, high-order enterprise relationships.

Purpose of the Study:

• To address the limitations of traditional ECR assessment and existing graph methods.
• To propose a novel framework, MS-CGNN, for enhanced enterprise representation learning.
• To incorporate high-order enterprise relationships using multi-structure graph learning.

Main Methods:

• Developed a Multi-Structure Cascaded Graph Neural Network (MS-CGNN) framework.
• Utilized knowledge graphs for pairwise relationships and hypergraphs for high-order relationships (homogeneous and heterogeneous).
• Introduced type-dependent hyperedge weight matrices for heterogeneous hypergraph convolutions.

Main Results:

• MS-CGNN effectively enhances enterprise representation learning.
• The framework leverages diverse graph structures (pairwise, homogeneous, heterogeneous hypergraphs).
• Achieved state-of-the-art performance on real-world ECR datasets.

Conclusions:

• MS-CGNN provides a robust solution for ECR assessment, overcoming data deficiency.
• Incorporating high-order relationships significantly improves risk assessment accuracy.
• The proposed method offers a more comprehensive approach to understanding enterprise interdependencies.