Νευρωνικά δίκτυα γράφων για την πρόβλεψη διμερών μεταναστευτικών ροών : μια ανάλυση με επίκεντρο την Ευρώπη
Graph neural networks for bilateral migration prediction : a European-focused analysis
View/
Abstract
Migration is among the most consequential and least predictable processes shaping contemporary societies, with direct implications for policy and humanitarian planning. This thesis investigates whether Graph Neural Networks (GNNs) improve on classical baselines for predicting bilateral migration flows, and analyses the European migration system through network-analytic tools. Three GNN architectures (GCN, GraphSAGE, Temporal-GNN) are benchmarked against three classical baselines (naive persistence, log-linear gravity model, Random Forest) on twenty-one years of bilateral flow data (2000–2020) covering 217 countries. The results show that the naive persistence baseline outperforms every learned model by a factor of approximately 4.5 in mean absolute error on non-zero cells, attributable to a short temporal axis (T = 21), severe target sparsity (90.4% zero cells), and strong year-on-year inertia. The network analysis identifies three structural patterns in European migration: the post-2007 Romanian enlargement spike, the gradual fade of the United Kingdom's network centrality across the post-Brexit window, and the eastward redirection of Ukrainian emigration toward Poland after 2014. A complementary sub-national analysis of internal displacement within Ukraine (2022–2026) using IOM DTM data is also reported.