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 present Mobility‑Embedded POIs (ME‑POIs), a method that augments standard text‑based place embeddings with aggregated human‑movement signals. Each visit to a POI is represented as a triple of latitude/longitude, arrival timestamp, and departure timestamp. These triples are encoded by a Space2Vec module for location and two independent Time2Vec modules for arrival and dwell duration, concatenated, positionally encoded, and fed into a 4‑layer, 8‑head Transformer (d_model=512) to yield contextualized visit vectors.

The framework learns a single prototype per POI via an InfoNCE contrastive loss that pulls each visit embedding toward its POI’s centroid while pushing away others. Sparse POIs are handled with multi‑bandwidth Gaussian kernel smoothing and KL‑divergence regularizers that transfer visit histograms from dense anchors and align them with empirical distributions. A fourth loss aligns the POI prototype with text embeddings constructed from coordinates, category, address, and the ten nearest neighbours (GeoLLM prompt). Across five map‑enrichment tasks on Los Angeles and Houston mobility data, adding ME‑POIs to strong text encoders improved 34 of 35 model‑task pairings in Los Angeles, with relative gains up to 81.9% F1 on visit intent and a 24.7% MAE reduction on busyness. Notably, a variant trained on mobility alone beat Gemini embeddings on price‑level classification, reaching 0.600 accuracy versus Gemini’s 0.559 on Los Angeles price level.

• ~53.7M parameters; pretrained on one NVIDIA Tesla V100 16 GB GPU.
• Encoders: Space2Vec (multi‑scale location) + two Time2Vec (arrival, departure).
• Visit encoder: 4‑layer, 8‑head Transformer, d_h=512.
• Requires licensed foot‑traffic or first‑party visit logs plus POI polygons (Los Angeles 6.9M visits, Houston 715k visits).
• No public code or weights released; only the paper is available.

Why this matters

The results indicate that collective mobility patterns can capture functional place attributes—such as price level or visit intent—more effectively than textual descriptors alone, suggesting that hybrid representations fusing language with behavior‑derived signals may improve downstream POI understanding. However, the method’s broader adoption hinges on access to licensed, high‑resolution foot‑traffic data and the inherent sparsity of visits for many POIs, which necessitates kernel‑based smoothing and additional regularization. Future research could explore privacy‑preserving aggregation or synthetic visit generation to reduce the data barrier while retaining the observed performance gains.