Google Research Introduces ME-POIs: A Mobility-Informed Framework that Adds “How a Place Is Used” to Text-Based POI Embeddings
A workforce from Google Research and USC has launched Mobility-Embedded POIs (ME-POIs), a framework that folds combination human motion into text-based place embeddings. The premise is that language fashions describe what a place is, however not how it’s used. Two espresso retailers can share a class, an handle block, and a textual content vector, whereas one runs commuter turnover and the opposite holds clients for ninety minutes. ME-POIs encodes every go to as a contextualized vector, then makes use of contrastive studying to align these visits with one learnable prototype per POI. Across 5 map-enrichment duties on Los Angeles and Houston mobility knowledge, including ME-POIs to robust textual content encoders improved 34 of 35 model-task pairings in Los Angeles, with relative positive factors up to 81.9% F1 on go to intent and a 24.7% MAE discount on busyness. Notably, a variant educated on mobility alone beat Gemini embeddings on price-level classification.
Is it deployable?
Partially, it’s a framework you rebuild, not a checkpoint you obtain. As of publication, Google Research has launched the paper however no public code or weights. The compute bar is low: the mannequin is ~53.7M parameters and was pretrained on a single NVIDIA Tesla V100 16GB. The actual gate is knowledge — you want licensed foot-traffic or first-party go to logs plus POI polygons.
How the framework works
Each go to is a triple: coordinates, arrival time and departure time. Three factorized encoders deal with them: Space2Vec for multi-scale location, and two Time2Vec encoders for arrival and departure individually, so begin time and dwell length keep distinguishable. The concatenated vectors get sinusoidal positional encoding and move by a 4-layer, 8-head Transformer (d_h = 512) to produce contextualized go to embeddings.
The core goal is contrastive. Every POI owns a learnable prototype, and an InfoNCE loss pulls every go to embedding towards its personal POI’s prototype whereas pushing away the opposite POIs within the minibatch. The prototype turns into a useful centroid that averages out particular person consumer schedules.
Sparsity is the arduous half. Only 9.07% of Los Angeles POIs and seven.04% of Houston POIs cleared the anchor threshold (100 and 50 complete visits respectively). For the lengthy tail, ME-POIs computes normalized Gaussian kernels at three bandwidths — 0.3 km, 1.0 km, 3.0 km — and transfers anchor go to histograms to sparse POIs, then provides a KL time period forcing the sparse embedding to predict that prior. A second KL time period supervises anchors towards their very own empirical distributions. A fourth loss maximizes cosine similarity with projected textual content embeddings, whose prompts observe the GeoLLM recipe: coordinates, class, handle, and the ten nearest POIs with distance and path.
What the numbers say
Evaluation covers 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) — throughout 5 map-enrichment duties with frozen-embedding probing. Labels come from SafeGraph for opening hours and closures, and Google Maps for go to intent, busyness, and worth degree.
Adding ME-POIs improved 34 of 35 model-task pairings in Los Angeles. Peak relative positive factors: 16.2% F1 on weekly opening hours (OpenAI-large), 81.9% F1 on go to intent (Gemini), 6.5% F1 on everlasting closure (E5), and a 24.7% MAE discount on busyness (Gemini). In Houston, price-level F1 rose 75.1% for GTR-T5. The single regression was Gemini on everlasting closure, down 0.4%.
The extra fascinating result’s the mobility-only variant. Trained with no textual content alignment in any respect, it reaches 0.600 accuracy on Los Angeles worth degree towards Gemini’s 0.559 — collective habits outperforming the phrases used to label the place. It additionally beats each trajectory-based baseline on each job.