Low-carbon supply chain logistics risk prediction using meta-learning-based graph convolutional network on prototype space

Predicting rare yet high-impact risks in low-carbon supply chains (LCSCs) is critical for sustainable operations but challenged by data imbalance and complex interdependencies. This study proposes a Meta-Learning-based Graph Convolutional Network on Prototype Space (ML-GCNPS) to address these issues. Using a real-world, multi-modal dataset of firm-level carbon emissions and supply-chain relationships, we model the LCSC as a graph. The model first extracts node features and embeds them into a prototype space for enhanced class discrimination. An adaptive Vertex-to-Edge (V2E) network then constructs the graph topology, enabling risk propagation via GCN. Extensive experiments demonstrate that ML-GCNPS significantly outperforms state-of-the-art baselines, achieving an AUPRC of 0.850 and reducing the False Negative Rate (FNR) to 0.080. These results, coupled with a low Weighted Average Cost (WAC) of 45, confirm its practical value for accurate and cost-effective risk early-warning.