Temporal Urban Region Evolution Forecasting with Graph Neural Networks and Adaptive Quantile Labeling

Understanding long-term urban functional evolution is challenging because cities change dynamically and reliable large-scale trajectory labels are rarely available. This paper proposes a graph-based framework for multi-horizon urban trajectory classification using Point-of-Interest (POI) distributions and Graph Neural Networks (GNNs). Urban graphs at the District-level are constructed from OpenStreetMap data collected in five cities, covering 41 administrative districts. Each graph represents spatial relationships among POIs, while temporal evolution is modeled through a fusion-score strategy that combines normalized POI distributions with category rarity information. Growth, Stable, and Decline labels are generated from temporal score variations using adaptive quantile thresholds. The framework is evaluated across multiple forecasting horizons and threshold settings using GraphSAGE, GAT, and GIN architectures with graph augmentation and multi-seed grid search. The experimental design focuses on graph-native models to preserve relational spatial dependencies within district-level urban structures. Results show that the Q30-Q70 threshold setting provides the most reliable overall trade-off, and GraphSAGE achieves the most consistent final multi-horizon performance. The additional comparison against a non-graph Random Forest baseline further demonstrates the importance of preserving spatial relational structures for urban trajectory forecasting across heterogeneous urban environments and forecasting horizons.