Evaluating directed and multiplex network representations for prediction of emerging contamination pathways in food safety surveillance

Food safety surveillance in globalised supply chains remains largely reactive, and it is unclear how much network-based machine learning contributes to making it proactive. We ask which components of a network-analytic pipeline genuinely improve prediction of emerging contamination pathways. We represent 28,724 European Union Rapid Alert System for Food and Feed notifications from 2019 to 2025 as a directed multiplex network, with edges from a product’s origin country to the notifying country across twelve hazard-specific layers. Because only 31 of 163 countries ever issue a notification, most country pairs are structurally impossible edges; sampling negatives uniformly therefore yields an AUC of 0.97 that a degree-preserving null reproduces exactly ( \\(p = 0.632\\) ). Under corrected sampling, directional modelling improves on an undirected treatment by 0.023 AUC ( \\(p = 10^{-5}\\) , ten seeds), giving 0.804 ± 0.011; only 2 of 200 rewired networks reach this value ( \\(p = 0.015\\) ). Multiplex layering adds nothing further ( \\(p = 0.487\\) ), nor does a graph neural network ( \\(p = 0.059\\) ). Ranked pairs identify pathways materialising in 2024–2025 with 45% precision at rank 100 against a 9.4% base rate, but an out-degree by in-degree product does equally well. Hazard layers differ sharply in severity, from 10.4% to 95.2% serious.