Google Research Introduces ME-POIs: A Mobility-Informed Framework that Adds “How a Place Is Used” to Text-Based POI Embeddings

Google Research and USC have put out ME-POIs, short for Mobility-Embedded POIs, a framework that tries to teach map models something text alone misses: how a place behaves in practice. A coffee shop can look nearly identical to another one in description and location, while one is all quick turnover and the other keeps people sitting for a long time. The model is built to catch that difference.

The setup is not a plug-and-play checkpoint. Google has published the paper, but not public code or weights, so anyone who wants it has to rebuild the system. The upside is that the compute needs are modest by current standards: about 53.7 million parameters, pretrained on a single NVIDIA Tesla V100 16GB. The real barrier is data. You need licensed foot-traffic or first-party visit logs, plus POI polygons.

Each visit is represented as coordinates, arrival time, and departure time. Those pieces go through separate encoders for space and for the two time fields, so when someone shows up and how long they stay remain distinct signals. The model then feeds the combined representation through a four-layer Transformer with eight attention heads, and trains it with contrastive learning so each visit moves toward a learnable prototype for its own place and away from other places in the batch.

That prototype idea matters because most places are sparse. In Los Angeles, only 9.07% of POIs cleared the anchor threshold of 100 total visits; in Houston, it was 7.04% at a threshold of 50. For the long tail, ME-POIs uses Gaussian kernels at 0.3 km, 1.0 km, and 3.0 km to borrow structure from nearby anchors, then adds KL losses to keep sparse places and anchors aligned with those priors. It also adds a text alignment term using prompts built from coordinates, category, address, and the ten nearest POIs.

The results are strong. Across five map-enrichment tasks on Los Angeles and Houston mobility data, adding ME-POIs improved 34 of 35 model-task pairings in Los Angeles. The biggest lift was 81.9% F1 on visit intent, and busyness saw a 24.7% MAE reduction. A mobility-only version, with no text alignment at all, even beat Gemini embeddings on price-level classification in Los Angeles. That is the part that should make people pause: sometimes collective behavior tells you more than the words on the storefront do.