Predictive Analytics for Supply Chain Resilience: A Graph-Based Model for Identifying and Mitigating Disruption Risks

Members: Suwanmanon Ruthakarn

Supervisor: Wu Zhenqin

Modern supply chains are highly interconnected networks vulnerable to cascading disruptions. While traditional risk management platforms and standard machine learning (ML) models excel at isolated forecasting, they fundamentally fail to capture the complex topological propagation of failures across supply chain tiers. To address this critical gap, this study proposes an advanced Graph Neural Network (GNN) framework designed to predict node-level resilience and model disruption cascades. Utilizing a reality-grounded synthetic dataset comprising 2,000 distinct disruption scenarios across a multi-tier supply chain, we benchmarked six state-of-the-art GNN architectures against five traditional ML baselines. The empirical results demonstrate that GNNs strictly outperform flat-feature ML models. Specifically, the edge-aware Graph Isomorphism Network (GINE) achieved the highest predictive performance (F1 = 0.7678), outperforming the strongest ML baseline, XGBoost, by a statistically significant 7.89 percentage points. This performance highlights the necessity of explicitly integrating edge attributes—such as lead times and capacities—with structural message-passing. To operationalize these theoretical advancements, we developed an interactive, web-based Decision Support System (DSS) that enables real-time topological vulnerability assessment and “what-if” scenario simulation. Ultimately, this research validates that graph-structured reasoning is essential for capturing relational dependencies in supply chains, providing practitioners with a robust, data-driven toolset for proactive resilience management.