T-food-ELA: A Time-Based Ensemble Learning Approach for Link Prediction in the Global Food Trade Network

The global food trade network is a complex system whose resilience is paramount to food security. Forecasting its evolution is critical for anticipating disruptions, yet accurate link prediction remains challenging due to the multi-dimensional nature of trade relationships. We present T-food-ELA, a time-based ensemble learning algorithm that integrates a suite of network similarity metrics within an XGBoost framework to predict link formation. Leveraging UN Comtrade data (2013–2023) for ten major food categories across 50 core nations, we demonstrate that our model significantly outperforms single-metric benchmarks, achieving an out-of-sample AUC of 0.84. Beyond predictive performance, we use model interpretation tools to uncover the fundamental principles governing the network's topology. We identify and characterize three structural effects: a preferential attachment effect reinforcing hub dominance; a saturation effect where an excess of common neighbors inhibits direct trade; and an intermediary effect whereby efficient indirect connections suppress direct links. Finally, we derive a resilience-based hierarchy of trade links, revealing distinct "core-core," "core-periphery," and "marginal" patterns. Our data-driven framework provides a computational foundation for enhancing the robustness of global food systems.