Introducing Mobility-Embedded POIs
The research team at Google Research has developed a new framework, Mobility-Embedded POIs (ME-POIs), designed to improve how language models understand places. The core idea is to combine textual descriptions of places with real-world mobility data – specifically, patterns of arrival times, stay durations, and surrounding movement – to create a more dynamic and accurate representation.
Traditional language models often rely solely on static metadata, such as addresses and business categories, to represent places. This approach can be limited because it doesn’t capture the actual functional rhythm of a place. ME-POIs addresses this by creating a numerical vector representation, called an embedding, that encodes both the identity of a place and its dynamic functionality. This embedding is constructed by blending text descriptions with large-scale, anonymized mobility patterns.
Performance Improvements
Experiments using publicly available benchmark datasets demonstrated significant improvements when integrating ME-POIs with advanced text models. The model achieved up to an 81.9% relative gain in predicting visit intent, a 75.1% improvement in price level classification, and a 24.7% increase in busyness estimation accuracy. These gains were observed across unseen places, indicating the framework’s generalizability.
Addressing Data Sparsity
Another key challenge in geospatial AI is the "long tail" problem – the scarcity of data for many local businesses. When a model encounters a place with limited visit data, it often makes inaccurate predictions, assuming zero activity. ME-POIs tackles this by utilizing a spatial multiscale visit propagation mechanism. This allows the model to learn from the behavior of nearby, more active places, effectively transferring aggregated visit patterns to sparse locations.
The ME-POIs Pipeline
The ME-POIs framework employs a three-step pipeline: visit alignment, spatial multiscale visit propagation, and text-mobility synergy. The model treats aggregate visits to a specific POI as fundamental data points, analyzing temporal arrival windows, departure trends, and stay durations. This process establishes a "functional centroid", a multidimensional signature that maps the aggregate anonymized mobility patterns associated with that specific place. The model then aligns high-level language embeddings with the newly generated mobility vectors by maximizing their cosine similarity.
