ME-POIs solves a core limitation of current place representations: language models capture what a place is—its name, category, address—but ignore how it is actually used. Two coffee shops can share identical text embeddings yet experience radically different visitation patterns, one serving commuters with short stops, the other hosting lingering customers. This gap hurts downstream tasks such as predicting opening hours, visit intent, price level, busyness, and closure detection.
The Mobility‑Embedded POIs (ME‑POIs) framework bridges this gap by turning raw foot‑traffic logs into context‑independent place vectors. Each visit is encoded as a triple of coordinates, arrival time, and departure time. Space2Vec captures multi‑scale location, while two Time2Vec modules separately model arrival and dwell duration. A lightweight Transformer (4 layers, 8 heads, 512‑dim hidden) then contextualizes the visit within a user’s sequence.
Contrastive learning aligns each visit embedding to a single learnable prototype per POI using an InfoNCE loss, turning the prototype into a functional centroid that averages out individual schedules. For the long tail of sparsely visited POIs (over 90 % in Los Angeles), ME‑POIs transfers visit histograms from nearby anchors via multi‑scale Gaussian kernels and enforces consistency with KL‑divergence terms. An optional loss also aligns the embedding with text‑based place descriptions.
Empirically, adding ME‑POIs to strong text encoders improved 34 of 35 model‑task pairings on Los Angeles data, delivering up to 81.9 % relative F1 gain on visit intent and a 24.7 % MAE reduction on busyness. A mobility‑only variant even surpassed Gemini text embeddings on price‑level classification. The model is modest (~53.7 M parameters) and trains on a single V100 16 GB GPU; the main barrier is obtaining licensed visit logs and POI polygons, not compute.
For teams looking to enrich maps with behavioral semantics, ME‑POIs offers a practical, data‑driven plug‑in that captures real‑world usage patterns without requiring massive infrastructure.
#AI #MachineLearning #NLP #Geospatial #DataScience #POI