Mobility-Embedded POIs (ME-POIs) is a framework that enriches place representations used by language models by combining text-based metadata with aggregated, anonymized mobility patterns. The goal is to move beyond a place’s static identity on record (name, address, category) toward capturing its operational rhythm — for example, typical arrival windows, dwell times, and surrounding movement patterns over the course of days and seasons.
The approach is part of Google Earth AI efforts and stresses that mobility signals are used at an aggregate level: ME-POIs builds population-level functional signatures of places rather than any individual-level or personalized models.
The three-step pipeline: turning mobility into numerical signatures
ME-POIs converts raw spatiotemporal observations into dense vector embeddings through three key steps:
-
Visit alignment
- Aggregate visits to a POI are treated as primary observations. The system analyzes temporal arrival windows, departure trends and typical stay durations.
- Rather than simple averages, a temporal encoder maps these sequences into a dense vector space, producing a multi-dimensional “functional centroid” that summarizes anonymized mobility patterns across a one-year cycle and different weekdays.
-
Spatial multiscale visit propagation
- To address the common “long tail” data sparsity problem — most local businesses have few recorded visits — ME-POIs propagates visit patterns across multiple spatial scales: immediate street, block, and broader neighborhood.
- The mechanism transfers aggregated patterns from busy, data-rich neighbors to nearby sparse places, effectively learning regional rhythms that act as geographic priors for under-observed POIs.
-
Text–mobility synergy
- Textual descriptions are not discarded. High-level language embeddings (for example embeddings from models like Gemini) are aligned with the newly created mobility vectors by maximizing cosine similarity.
- The hybrid embedding preserves semantic structure (e.g., a place sells food) while injecting operational context (e.g., whether it functions primarily as a lunch spot or a late-night diner).
Evaluation: cities, tasks and baselines
The framework was evaluated using publicly available benchmark datasets in two large and culturally distinct metropolitan areas: Los Angeles and Houston. ME-POIs was tested on five downstream tasks, and crucially the model was trained on a set of observed places and evaluated on entirely unseen places to measure generalization.
The five tasks were:
- Opening/closing hours prediction
- Price-level classification
- Permanent closure detection
- Visit intent classification (a proxy for aggregate search and navigation interest)
- Busyness forecasting (future crowd-density and peak-hour dynamics)
ME-POIs was compared against standard text-only embedding models (such as Gemini embeddings), trajectory-based geospatial models (such as TrajGPT), and hybrid variants to isolate the contribution of mobility signals.
Results
In tests on unseen places, incorporating ME-POIs produced consistent and substantial accuracy improvements across all tasks, confirming that mobility adds a practical context advantage to text models. Representative gains include:
- Up to 81.9% relative improvement in predicting visit intent.
- 75.1% relative improvement in price-level classification.
- 24.7% increase in busyness estimation accuracy.
Notably, in several settings a mobility-only model outperformed text-only language models (for example in price-level classification), demonstrating that collective physical activity patterns can sometimes be more descriptive about a place’s function than formal textual labels.
Limitations and privacy considerations
ME-POIs is designed to operate on aggregated, anonymized mobility data and explicitly does not produce individual-level personalization. The spatial propagation used to mitigate sparsity is a statistical prior and may have limits on accuracy for very sparsely observed venues.
Conclusion
By integrating anonymized mobility patterns into place embeddings, ME-POIs produces richer, functionally informed representations that help downstream models infer attributes of places more easily and accurately. The framework reduces computational burden for inferential tasks and lays groundwork for geospatial models that better capture the rhythm of cities. The paper’s co-authors include Neha Arora (Google Research), Prof. Cyrus Shahabi and Ph.D. student Shang Ling Hsu from the University of Southern California.
