Google's ME-POIs Framework Combines Human Movement with Place Embeddings

Google Research has developed a new framework called ME-POIs (Mobility-Informed POIs) that combines aggregate human movement with text-based place embeddings. The goal is to better understand how places are used, rather than just what they are.

The ME-POIs framework uses contrastive learning and multi-scale KL transfer to align visit embeddings with learnable prototypes for each POI. This allows it to capture the nuances of how a place is used, even if the text-based embedding doesn't explicitly mention it.

In experiments using mobility data from Los Angeles and Houston, ME-POIs improved performance on five map-enrichment tasks by up to 81.9% F1 on visit intent and reduced mean absolute error (MAE) by 24.7% on busyness. A variant trained only on mobility data even outperformed Gemini text embeddings on price-level classification.