How Google’s ME-POIs Framework Enhances Location Embeddings with Human Mobility Data

The Premise

The stated objective: language models describe what a place is, not how it is used. Two adjacent coffee shops can share category, address block, and text vector, yet diverge sharply in dwell-time distribution and visit cadence. ME-POIs treats this gap as an embedding-level deficiency.

Architecture and Training

Each visit is encoded as a triple — coordinates, arrival time, departure time — and routed through three factorized encoders: Space2Vec for multi-scale location and two Time2Vec encoders for arrival and departure separately, preserving dwell duration as a distinguishable signal. The concatenated vectors receive sinusoidal positional encoding and pass through a 4-layer, 8-head Transformer (d_h = 512), producing contextualized visit embeddings. The objective is contrastive: every POI owns a learnable prototype, and an InfoNCE loss pulls each visit embedding toward its POI's prototype while pushing away the other POIs in the minibatch. The prototype functions as a usage centroid, averaging out individual schedules.

Sparsity is the binding constraint. Only 9.07% of Los Angeles POIs and 7.04% of Houston POIs cleared the anchor threshold — 100 total visits in LA, 50 in Houston. For the long tail, ME-POIs computes normalized Gaussian kernels at three bandwidths (0.3 km, 1.0 km, 3.0 km), transfers anchor visit histograms to sparse POIs, and adds a KL term forcing the sparse embedding to predict that prior. A second KL term supervises anchors against their own empirical distributions. A fourth loss maximizes cosine similarity with projected text embeddings, with prompts following the GeoLLM recipe — coordinates, category, address, and the ten nearest POIs with distance and direction.

Benchmarks

Evaluation used two anonymized mobility datasets: Los Angeles (39,557 POIs, 6.9M visits, full-year 2019) and Houston (28,419 POIs, 715,604 visits, 20 days in March 2020), across five map-enrichment tasks with frozen-embedding probing. Labels came from SafeGraph for opening hours and closures, and Google Maps for visit intent, busyness, and price level. In Los Angeles, adding ME-POIs improved 34 of 35 model-task pairings. Peak relative gains: 16.2% F1 on weekly opening hours (OpenAI-large), 81.9% F1 on visit intent (Gemini), 6.5% F1 on permanent closure (E5), and a 24.7% MAE reduction on busyness (Gemini). In Houston, price-level F1 rose 75.1% for GTR-T5. The sole regression was Gemini on permanent closure, down 0.4%.

The more discriminating result: a variant trained on mobility alone outperformed Gemini embeddings on price-level classification. Text-only encoders do not recover usage signals from metadata; visit trajectories do.

What to Verify

The compute footprint is modest — ~53.7M parameters, pretrained on a single NVIDIA Tesla V100 16GB. As of publication, Google Research has released the paper but no public code or weights. The binding gate is data access: licensed foot-traffic or first-party visit logs plus POI polygons are required. Researchers should treat the paper as a methodology specification rather than a deployable checkpoint, and reproduce the InfoNCE prototype alignment and sparsity-handling ablation before drawing conclusions about the 81.9% F1 visit-intent gain in their own setting.