Supply Chain Credit Risk Propagation Modeling Through Heterogeneous Graph Learning and Enterprise Relation Inference

Abstract

Supply chain finance involves complex credit relationships among core enterprises, upstream suppliers, downstream distributors, logistics providers, and financial institutions. Traditional risk assessment models often rely on isolated financial indicators and fail to capture risk propagation caused by delayed payments, guarantee chains, ownership ties, and transaction dependency. This study proposes a heterogeneous graph learning method for supply chain credit risk propagation modeling. The method constructs an enterprise relation graph using procurement records, invoice flows, payment delays, equity links, guarantee relationships, and litigation events. A graph neural network is used to encode structural dependencies among enterprises, while a knowledge inference module identifies hidden risk paths between financially connected firms. Experiments are conducted on a supply chain finance dataset containing 86,000 enterprises, 1.42 million transaction edges, 318,000 invoice records, 74,000 guarantee links, and 21,600 risk event labels collected over 36 months. The proposed method reduces median early-warning delay from 63 days to 24 days compared with a financial-ratio scoring baseline. The top-decile risk queue captures 3,940 confirmed delayed-payment and credit-deterioration cases during the test period. Graph inference identifies 12,700 indirect risk propagation paths, including supplier concentration risk, cross-guarantee exposure, and delayed receivable chains. The model also reduces manual review volume from 18,400 enterprise alerts to 5,260 relation-level investigation cases through risk-path aggregation. The results show that heterogeneous graph learning and relation inference can improve supply chain credit risk prediction by incorporating both enterprise attributes and network-level risk transmission.