source: Google Research: How mobility gives language models a deeper understanding of place
level: research
Google Research introduced Mobility-Embedded POIs (ME-POIs), a framework that enriches language model representations of places with aggregated, anonymized mobility patterns. The approach combines static text metadata with dynamic visit data such as arrival times, stay durations, and surrounding movement. This creates embeddings that capture both a place's identity and its functional rhythm, enabling better predictions about real-world attributes like opening hours, price levels, and busyness.
In experiments across Los Angeles and Houston, ME-POIs delivered up to an 81.9% relative gain in visit intent prediction, 75.1% improvement in price level classification, and 24.7% increase in busyness estimation accuracy on unseen places. The framework uses a three-step pipeline: visit alignment, spatial multiscale visit propagation to address data sparsity, and text-mobility synergy. Notably, mobility-only models sometimes outperformed text-only models, showing that aggregate behavior can be more descriptive than formal labels.
The work is part of Google's broader Earth AI effort to create geospatial models that turn planetary data into actionable intelligence. ME-POIs focuses on aggregate place understanding and cannot be used for individual personalization. By shifting mobility from an output prediction task to an input feature, the framework reduces computational burden on downstream systems. This could improve services like local search, business analytics, and urban planning by providing richer, more dynamic place representations.
why it matters: For AI and data science, ME-POIs shows how combining text with mobility data can significantly improve geospatial predictions, enabling more accurate models for local search, business intelligence, and urban analysis.
source: Google Research: How mobility gives language models a deeper understanding of place