This paper highlights the importance of pre-training geospatial models for urban representation learning, driven by the increasing volume and continuous updating of geospatial data from diverse sources, to advance data-driven urban planning. Specifically, to address the inability of existing models to infer places across multiple spatial granularities and context-rich regions, we propose PlaceFM, a clustering-based approach that captures place representations without requiring training. PlaceFM encapsulates a complete point-of-interest (POI) graph built on U.S. Foursquare data, generating general-purpose region embeddings and automatically identifying POIs. These embeddings can be directly integrated into geolocation data pipelines to support a variety of urban-related follow-up tasks. PlaceFM provides a scalable and efficient solution for multi-granular geospatial analysis on large-scale POI graphs without the need for expensive pre-training. Experimental results on two real-world prediction tasks, such as ZIP code-level population density and housing price prediction, demonstrate that PlaceFM outperforms most state-of-the-art graph-based geospatial models and is up to 100 times faster at generating place-level representations from large POI graphs.
