How mobility gives language models a deeper understanding of place
| Source: Google Research Blog
Tags: Google Research, place embeddings, geospatial AI, mobility data, POI, ME-POIs, language models
Google Research introduces ME-POIs, a framework enriching LLM place representations with anonymized mobility patterns — arrival times, stay durations, movement flows — achieving up to 81.9% relative gain in visit intent prediction, 75.1% improvement in price level classification, and 24.7% better busyness estimation.
Details
Google Research's Mobility-Embedded POIs (ME-POIs) framework targets a core limitation of language models: reliance on static metadata (addresses, categories, text descriptions) when representing places. Real-world locations have temporal rhythms — a coffee shop packed at 8am, quiet at 3pm — that text alone cannot capture. The framework uses a self-supervised approach to blend text descriptions with large-scale, anonymized mobility patterns from public benchmarks, capturing aggregate spatial activity footprints throughout the day: arrival times, stay durations, and surrounding movement flows. The result is a numerical embedding encoding both a place's identity and its dynamic functionality. On publicly available benchmarks, ME-POIs delivered an 81.9% relative gain in predicting visit intent, a 75.1% improvement in price level classification, and a 24.7% increase in busyness estimation accuracy across unseen places — without recalculating patterns at inference time. The approach matters for location-aware AI: navigation assistants, recommendation systems, and business analytics. Pre-enriched place representations let downstream models infer operating hours, price levels, and business status without complex real-time computation.