Using GNN to Identify Supply Chain Network Structure

In the context of globalization and frequent extreme events, the identification of supply chain network topology and the characterization of risk propagation are key points to enhance industry resilience.This study constructs a heterogeneous supply chain graph containing multiple types of nodes and multi-dimensional edge weights for enterprises, logistics, finance and regulation, and proposes explanatory indicators such as multi-type PageRank and heterogeneous redundancy coefficient.A heterogeneous graph neural network based on the meta-path attention mechanism is constructed, and time coding is introduced to capture the dynamic changes of the structure.The identification of key nodes and the estimation of the probability of supply interruption are jointly optimized by multi-task learning.In the context of Resilinc and BTS public data sets, the recall rate of the model for hidden key nodes is 14 percentage points higher than that of the traditional centrality method.In the simulation of the collapse of the Baltimore Bridge, a re-planned route plan can be recommended 36 hours in advance.This framework can provide accurate, interpretable, timely and effective network monitoring and decision support, and build a methodological foundation for the subsequent integration of federated learning and digital twins.