Relational Verification, Feature Sensitivity, and Cross-City Transfer in Heterogeneous Graph Learning for Urban Functional Zone Classification

Urban functional zone classification is a key component of data-driven urban analytics, supporting applications in planning, mobility, and digital governance. Although point-of-interest (POI) composition is widely used to infer regional functionality, it remains unclear whether predictive performance reflects meaningful relational structure or primarily direct semantic signals derived from the same data source. This study introduces a heterogeneous graph learning framework that models Region-POI, POI-Category, and Region-Region interactions for functional zone classification. Comparable graph representations are constructed for three cities with diverse urban characteristics - Baku, Berlin, and Montevideo - using a uniform grid partition and OpenStreetMap-derived POI data. To enable rigorous evaluation, the analysis incorporates structured relation ablation, cross-city transfer, and a feature sensitivity setting where direct semantic region features are removed. Experimental results show that conventional models and homogeneous graph baselines experience substantial performance degradation under feature removal, whereas the heterogeneous model maintains strong performance in both in-city and cross-city settings. Relation ablation further indicates that Region-POI interactions contribute most significantly to predictive performance, while spatial adjacency provides comparatively limited explanatory power. These findings suggest that heterogeneous relational modeling captures meaningful structural dependencies in urban data and provides a more reliable representation of functional organization under varying feature and transfer conditions.