SCF credit risk assessment with limited labeled data using label propagation algorithm and complex network approaches

Assessing enterprise credit risk in supply chain finance (SCF) is critical for maintaining financial stability, but it is often hindered by the limited labeled data. We investigate the effectiveness of label propagation algorithm (LPA)-based approaches in credit risk assessment under condition of labeled data scarcity. We construct the feature-based, graph-based, and Dual Graph Convolutional Networks (DGCN)-optimized LPA models and compare their performances with those of the SL models and traditional SSL models. Based on the credit risk assessment results, we study the mechanism of credit risk transmission in SCF through investigating the influential enterprise and feature's effect in results of SEIRS epidemic model. We find that: (1) the LPA models outperform the SL and traditional SSL models when the labeled data is scarce; (2) the DGCN approach further enhances the LPA method's ability by capturing both the local and global network associations; and (3) the enterprises with high network centrality and operation efficiency have significant influence in credit risk transmission.