Google and USC’s ME-POIs Adds Mobility Data to Place AI, Lifting Visit-Intent F1 by 81.9%

Google Research and the University of Southern California have introduced ME-POIs, a framework that combines text-based place embeddings with aggregate human-movement signals. In tests across Los Angeles and Houston, the researchers reported that the added mobility representation improved 34 of 35 model-task pairings in Los Angeles, including an 81.9% relative F1 gain on visit-intent classification using Gemini embeddings.

The work tackles a narrow but important gap in mapping and local-search AI. A text embedding can capture a business category, address and description; ME-POIs is designed to add evidence about how a place is used, based on anonymized, aggregated visits. The framework is intended for place-level understanding, not conclusions about individual users or personalization.

A representation built from visits, not routes

ME-POIs treats each visit as a record of coordinates, arrival time and departure time. It uses contrastive learning to turn those records into one prototype vector for each point of interest, creating a representation of the place rather than a prediction of where a particular person will go next.

That design has a second component for the places with little recorded activity. ME-POIs transfers aggregated visit patterns from data-rich neighboring locations to nearby sparse ones, then aligns the mobility vector with a text embedding through cosine similarity. Sparse coverage is not an edge case in the evaluation: only 9.07% of Los Angeles POIs and 7.04% of Houston POIs met the study’s anchor-visit thresholds.

The benchmark is broad, but local

The evaluation covered five map-enrichment tasks: opening and closing hours, price level, permanent closure, visit intent and busyness forecasting. Its data spanned 39,557 Los Angeles POIs and 6.9 million visits during 2019, plus 28,419 Houston POIs and 715,604 visits recorded over 20 days in March 2020.

One result helps clarify what the mobility signal contributes. A mobility-only variant reached 0.600 accuracy on Los Angeles price-level classification, compared with 0.559 for Gemini embeddings. That does not establish that mobility is generally superior to text; it shows that, on this task and dataset, observed aggregate behavior supplied information that a text embedding did not.

The deployment constraint is data

• The published system is a framework to rebuild, rather than a downloadable model: no public code or weights had been released at publication.
• The reported model has about 53.7 million parameters and was pretrained on a single NVIDIA Tesla V100 with 16GB of memory.
• The practical input requirement is licensed foot-traffic data or first-party visit logs, together with point-of-interest polygons.

That leaves the central open question outside the benchmark: which organizations can obtain the required mobility data and apply the method beyond the two tested cities. The reported compute footprint suggests training need not be the main barrier. Data rights, coverage and geographic transferability are likely to determine whether the approach becomes a broadly usable mapping layer or remains a research recipe for operators with privileged location data.