A research team from Google Research and USC has released Mobility-Embedded POIs (ME-POIs), a new framework that integrates aggregate human movement patterns into text-based place embeddings. The core insight is that language models can describe what a place is, but they fail to capture how it is used. For instance, two coffee shops might share the same category, sit on the same block, and have similar text descriptions—yet one sees a steady flow of commuters grabbing quick cups, while the other hosts customers who linger for over an hour. ME-POIs bridges this gap by encoding each visit as a contextualized vector and then applying contrastive learning to align those visits with a single learnable prototype per place.
Across five map-enrichment tasks using mobility data from Los Angeles and Houston, adding ME-POIs to strong text encoders yielded significant improvements. In Los Angeles, 34 out of 35 model-task pairings showed gains, with relative improvements reaching 81.9% F1 on visit intent classification and a 24.7% reduction in mean absolute error (MAE) for busyness prediction. Notably, a variant trained solely on mobility data outperformed Gemini embeddings on price-level classification.
Is it deployable?
Partially. This is a framework you rebuild rather than a ready-to-use checkpoint. As of the publication date (August 24, 2026), Google Research has released the paper but no public code or model weights. The computational requirements are modest: the model has approximately 53.7 million parameters and was pretrained on a single NVIDIA Tesla V100 with 16GB memory. The main challenge is data. You need access to licensed foot-traffic data or first-party visit logs, along with accurate POI polygons. This makes the framework most practical for organizations that already possess substantial mobility datasets or have the resources to acquire them.
via MarkTechPost