Google Research Adds Mobility to Place Embeddings

Google Research, in collaboration with the University of Southern California (USC), has introduced Mobility-Embedded POIs (ME-POIs), a novel framework designed to enhance text-based place embeddings by incorporating aggregate human movement data. The core innovation of ME-POIs lies in its ability to capture the functional usage of a place, a dimension often overlooked by traditional language models that primarily describe what a place is. For instance, two coffee shops might share similar textual descriptions and geographical proximity but serve vastly different customer bases, with one catering to quick commuter turnover and the other to longer customer stays. ME-POIs addresses this by encoding each individual visit to a location as a contextualized vector. These visit vectors are then aligned with a learnable prototype for each Point of Interest (POI) using contrastive learning.

The framework's effectiveness was demonstrated across five map-enrichment tasks utilizing mobility data from Los Angeles and Houston. When ME-POIs was integrated with established text encoders, it led to significant improvements. Specifically, 34 out of 35 model-task pairings in Los Angeles showed enhanced performance. The framework achieved relative gains of up to 81.9% F1 score on visit intent prediction and a reduction of 24.7% in Mean Absolute Error (MAE) for busyness estimation. In a notable comparison, a variant of ME-POIs trained solely on mobility data outperformed Google's own Gemini embeddings in classifying the price level of locations.

ME-POIs is presented as a framework for researchers to rebuild rather than a ready-to-download model checkpoint. As of its publication, Google Research has released the academic paper detailing the framework but has not yet provided public code or pre-trained weights. The computational requirements for training ME-POIs are relatively modest, with the model comprising approximately 53.7 million parameters and having been pre-trained on a single NVIDIA Tesla V100 16GB GPU. The primary barrier to deployment is not computational power but access to necessary data, which includes licensed foot-traffic data or first-party visit logs, along with precise POI geographical boundaries.

The ME-POIs framework processes each visit as a data triple consisting of geographical coordinates, arrival time, and departure time. Three distinct factorized encoders are employed to process this information. Space2Vec is utilized for capturing multi-scale spatial information, while two Time2Vec encoders independently process arrival and departure times. This separation ensures that the start time of a visit and its duration remain distinguishable. The resulting vectors from these encoders are concatenated, augmented with sinusoidal positional encoding, and then passed through a four-layer Transformer with eight attention heads (hidden dimension size of 512) to generate the contextualized visit embeddings. The central learning objective is contrastive: each POI maintains a learnable prototype, and an InfoNCE loss function is applied to pull the visit embeddings closer to their corresponding POI's prototype while simultaneously pushing them away from the prototypes of other POIs within the same mini-batch. This process enables the prototype to evolve into a functional representation of how a place is typically used.