Baseline framework for inferring POI floors to enhance 3D urban vertical analytics: A comparative study in Shanghai

Accurate floor-level information of Points of Interest (POIs) is essential for extending urban analytics from planar to three-dimensional geospatial environments. Many studies of earth observation adopt POIs and focus on planar functional representations as well as urban morphology. However, most open POI datasets lack vertical attributes, causing incomplete semantic representations. This study proposes a baseline framework for urban POI floor inference by integrating GraphSAGE and XGBoost approaches. Using the city-wide POI dataset of Shanghai city extracted from digital map providers, 573,023 valid POIs were identified within building footprints, of which 31.1 % contained recoverable floor information. A K-nearest-neighbor spatial graph was constructed to extract localized spatial-contextual embeddings with GraphSAGE, while XGBoost served as a stand-alone model as well as a downstream regressor/classifier using GraphSAGE embeddings. Accuracy, tolerance accuracy (±1 to ±3 floors) and quadratic weighted kappa (QWK) are used for evaluating performance. Results show that XGBoost models provide strong and stable baselines, and GraphSAGE embeddings bring complementary gains. On the test dataset, a GraphSAGE-embedding–XGBoost hybrid model achieves the best ±2 tolerance accuracy (about 0.73) and QWK (about 0.75). The results demonstrate the feasibility of inferring POI floor levels from partially labeled open data using lightweight features, which can support vertical semantic enrichment for 3D urban analytics and serve as an updateable semantic enrichment module for urban digital twins, when new POI and building data become available.