Forecasting migration inflows through multi-view graph-temporal fusion

Lazaridis, Christos - Lazaros; Λαζαρίδης, Χρήστος - Λάζαρος

Πρόβλεψη μεταναστευτικών εισροών µέσω πολυπρισµατικής χωροχρονικής σύντηξης γράϕων

Bachelor Dissertation

Author

Lazaridis, Christos - Lazaros

Λαζαρίδης, Χρήστος - Λάζαρος

Date

2026-09

Advisor

Sotiropoulos, Dionisios
Σωτηρόπουλος, Διονύσιος

Keywords

Migration ; Forecasting ; Machine learning ; Graph neural networks ; Multi-view learning

Abstract

International migration inflows are difficult to forecast because they combine strong temporal persistence with highly unequal flow magnitudes, incomplete reporting, relational structure, and abrupt changes associated with economic, political, and humanitarian events. This thesis investigates whether these characteristics can be modelled through a multi-view graph-temporal forecasting framework. The empirical target is the annual inflow of foreign population reported in Table B.1 of the OECD International Migration Database, indexed by nationality or citizenship and reporting destination. These observations are therefore not interpreted as conventional origin-destination flows based on previous residence. The proposed Fusion architecture organizes its inputs into three representations: Tower A encodes the directed graph of historical inflows; Tower B encodes economic, conflict, and social or demographic country-level sequences for the nationality-labelled and destination entities; and Tower C combines static dyadic context with a recurrent encoding of dyadic inflow history. The fused representation predicts a residual around a causal five-year historical log-mean reference supplied separately to the decoder. A second training condition adds a Random Forest teacher signal through knowledge distillation. The framework is compared with deterministic temporal rules, Random Forest, gravity-style models, and graph-based alternatives. Evaluation follows an expanding rolling-origin design over 2019-2023, with models refitted using only information available before each forecast year. Performance is assessed using both regression metrics and event-detection metrics for abrupt increases and decreases. Under both training conditions, the full three-view A+B+C Fusion models performed worse than their corresponding controlled A+B (-C) variants, obtained by withholding Tower C. Relative to the corresponding full three-view Fusion models, the A+B (-C) variants reduced pooled RMSE by 63.4% under RF distillation and by 53.9% under rolling training. Nevertheless, Random Forest and EWMA-5 remained stronger on the principal regression metrics. The event results followed a different pattern: RF-distilled A+B (-C) achieved the highest pooled surge F1 among the forecasting methods evaluated, at 0.439. This advantage varied across forecast years and was not tested against dedicated classification models. The results therefore do not support the original three-view architecture. They instead identify a simpler configuration that warrants further investigation, while showing that additional representational complexity does not necessarily improve forecasting in this setting.