Google introduces me-pois: a framework that enhances place understanding with mobility data

A team from Google Research and the University of Southern California has introduced ME-POIs, a novel framework that integrates human mobility patterns into text-based point-of-interest (POI) embeddings. The framework addresses the limitation of traditional language models, which describe what a place is but not how it is used. By encoding each visit as a contextualized vector and using contrastive learning to align these visits with learnable prototypes per POI, ME-POIs provides a more nuanced understanding of how places are utilized. The model was pretrained on a single NVIDIA Tesla V100 16GB GPU and has approximately 53.7 million parameters. ME-POIs processes visits as triples of coordinates, arrival time, and departure time, using three factorized encoders: Space2Vec for location and two Time2Vec encoders for arrival and departure times. These vectors are combined with sinusoidal positional encoding and passed through a 4-layer, 8-head Transformer to generate contextualized visit embeddings. The core objective is to minimize an InfoNCE loss, which pulls visit embeddings toward their respective POI prototypes while pushing them away from others in the minibatch. This results in a functional centroid that represents the average user schedule for each POI. The framework faces challenges with sparsity, as only 9.07% of Los Angeles and 7.04% of Houston POIs met the minimum visit threshold. To address this, ME-POIs uses normalized Gaussian kernels at three bandwidths and transfers anchor visit histograms to sparse POIs, with additional KL terms to ensure consistency. The model also incorporates text embeddings through prompts based on the GeoLLM recipe, including coordinates, category, address, and nearby POIs. Evaluation on anonymized mobility datasets from Los Angeles and Houston showed significant improvements. In Los Angeles, adding ME-POIs improved 34 out of 35 model-task pairings, with relative gains up to 81.9% F1 on visit intent and a 24.7% MAE reduction on busyness. In Houston, price-level F1 increased by 75.1% for GTR-T5. A mobility-only variant, trained without text alignment, achieved 0.600 accuracy on price level classification in Los Angeles, outperforming Gemini’s 0.559. It also surpassed all trajectory-based baselines across tasks. Despite the promising results, Google Research has not released public code or weights, and the model requires licensed foot-traffic data or first-party visit logs along with POI polygons. The paper and technical details are available for further exploration, along with tutorials, codes, and notebooks on the GitHub page. The research was first published on MarkTechPost.