Google's ME-POIs Captures How Places Are Actually Used

Two coffee shops can sit on the same block, carry the same category tag, and produce nearly identical text embeddings, while one turns over commuters in five minutes and the other holds customers for an hour and a half. Text-based place embeddings, the kind built from reviews, menus and descriptions, capture what a location claims to be but miss what actually happens inside it. Google Research and USC set out to close that gap with Mobility-Embedded POIs, or ME-POIs, a framework that folds aggregate movement data, arrival times, departure times, coordinates, into the same vector space as text descriptions of a place.

The approach encodes each visit as its own contextualized vector, then uses contrastive learning to pull those visit vectors toward a single learnable prototype per point of interest. Tested across five map-enrichment tasks using mobility data from Los Angeles and Houston, the method was checked against strong existing text encoders to see whether adding movement signal actually changes what a model can predict about a place, from how busy it gets to what kind of visit it attracts.