Google Research Introduces ME-POIs: A Mobility-Informed Framework that Adds “How a Place Is Used” to Text-Based POI Embeddings
Google's ME-POIs adds movement data to place embeddings—beating text-only approaches on real-world map tasks by up to 24.7%.
Why it matters
A research contribution showing how multimodal embeddings (text + mobility patterns) outperform text-alone for location understanding. Relevant to practitioners building location services, maps, and recommendation systems; enthusiasts tracking embedding and multimodal learning progress.
The key facts
11 to knowME-POIs framework combines text embeddings with aggregate human mobility patterns
Improved 34 of 35 model-task pairings on Los Angeles data
81.9% relative F1 on visit intent classification
24.7% MAE reduction on busyness prediction
Mobility-only variant outperformed Gemini text embeddings on price-level classification
Tested on Los Angeles and Houston mobility data across five map-enrichment tasks
Uses contrastive learning and learnable prototypes per POI
ME-POIs framework encodes visit patterns as contextualized vectors with contrastive learning
Improved 34 of 35 model-task pairings in Los Angeles testing
Tested on Los Angeles and Houston mobility data across three spatial scales
Addresses limitation: LLM text embeddings miss how places are actually used
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: framework that folds aggregate human movement into text-based place embeddings. Language models describe what a place is; they miss how it is used. ME-POIs encodes each visit as a contextualized vector and aligns it with one learnable prototype per POI through contrastive learning, then transfers…