ME-POIs: Google's 53.7M-Parameter Model Adds Visit Rhythms to Place Embeddings

Google Research and USC have published Mobility-Embedded POIs (ME-POIs), a framework that augments text-based place embeddings with aggregate human movement data. The core argument: language models encode what a location is — name, category, address, nearby landmarks — but not how it is used. Two coffee shops on the same block can share identical text vectors while one runs eight-minute commuter turnover and the other anchors ninety-minute afternoon stays. ME-POIs resolves this by encoding each visit as a contextualized vector and aligning those vectors with per-POI learnable prototypes via contrastive learning.

Across five map-enrichment tasks on Los Angeles and Houston mobility data, adding ME-POIs to existing text encoders improved 34 of 35 model-task pairings in Los Angeles, with peak relative gains of 81.9% F1 on visit intent and a 24.7% MAE reduction on busyness prediction. For practitioners building location intelligence pipelines, this represents the kind of pipeline architecture improvement driving the biggest AI gains in 2026 — incremental at the component level, material at the system level.

Architecture: Three Encoders, One Transformer, One Prototype Per POI

Each visit is a triple: coordinates, arrival time, and departure time. Three factorized encoders handle these independently before concatenation. Space2Vec processes location at multiple scales. Two separate Time2Vec encoders handle arrival and departure, keeping start time and dwell duration distinguishable. The concatenated representation receives sinusoidal positional encoding and passes through a 4-layer, 8-head Transformer with hidden dimension d_h = 512, producing a contextualized visit embedding.

The contrastive objective is InfoNCE-based. Each POI owns exactly one learnable prototype vector; training pulls every visit embedding toward its POI's prototype while pushing embeddings from other POIs in the minibatch away. The prototype converges to a functional centroid that absorbs individual user schedules, making the representation context-independent rather than trajectory-conditioned — which matters for downstream frozen-embedding probing. A fourth loss term maximizes cosine similarity with projected text embeddings, where prompts follow the GeoLLM recipe: coordinates, category, address, and the ten nearest POIs with distance and bearing. The full model is approximately 53.7M parameters, pretrained on a single NVIDIA Tesla V100 16GB GPU.

Long-Tail Transfer: Multi-Scale Gaussian Kernels and KL Supervision

Sparsity is the structural problem. Only 9.07% of Los Angeles POIs and 7.04% of Houston POIs cleared the anchor threshold — at least 100 total visits in Los Angeles and 50 in Houston. The remaining 91% of LA POIs and 93% of Houston POIs lack sufficient direct visit data.

ME-POIs addresses this with normalized Gaussian kernels at three bandwidths: 0.3 km (street-level), 1.0 km (block-level), and 3.0 km (district-level). Anchor visit histograms are transferred to sparse POIs using these kernels, and a KL divergence term forces each sparse embedding to predict the resulting blended prior. A second KL term supervises anchor embeddings against their own empirical distributions. The three-bandwidth design preserves both local specificity and regional rhythm simultaneously.

Benchmark Results Across Two Cities and Five Tasks

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) — with frozen-embedding probing across five map-enrichment tasks. Labels came from SafeGraph for opening hours and permanent closures, and Google Maps for visit intent, busyness, and price level.

The result worth isolating is the mobility-only variant. Trained without any text-alignment loss, it achieves 0.600 accuracy on Los Angeles price-level classification against Gemini embeddings' 0.559 — collective behavioral signal outperforming the semantic description of the place. It also outperforms every trajectory-based baseline across every task in the evaluation.

Deployability: The Data Barrier, Not the Compute Barrier

The compute profile is accessible: ~53.7M parameters trained on a single V100 16GB. The paper targets ICML 2026. What is not yet available is public code or model weights, meaning ME-POIs is currently a framework to rebuild rather than a checkpoint to deploy. The harder constraint is upstream: the framework requires licensed foot-traffic data or first-party visit logs along with POI polygon boundaries. Proprietary data access is the real adoption gate, not hardware.

ME-POIs formalizes something the geospatial ML community has intuited but lacked a clean mechanism for: behavioral context is a first-class feature type, not an annotation layer. The contrastive prototype design keeps the learned representation context-independent and composable with any frozen text encoder — a property that matters for production systems where encoder substitution is routine. Whether the framework generalizes beyond North American urban datasets with dense SafeGraph coverage remains the open empirical question.