Google Research Unveils ME-POIs, Lifting Visit-Intent Prediction by 81.9%

2 articles · Updated · Google Research · Aug 21

ME-POIs blends text descriptions with aggregated, anonymized mobility signals to give AI models a dynamic representation of places rather than relying only on static metadata like category or address.
81.9% relative gains in visit-intent prediction, 75.1% better price-level classification and 24.7% higher busyness accuracy came when Google paired the framework with advanced text models on unseen places.
A three-step pipeline encodes arrival times, stay durations and surrounding movement patterns, then propagates neighborhood rhythms to sparse-data locations such as small shops and new cafes.
Tests in Los Angeles and Houston found the system outperformed text-only, mobility-only and hybrid baselines across five tasks, with mobility-only models sometimes beating text-only models on price classification.
Google said the framework is designed for aggregate place understanding rather than individual personalization, and positions it as part of its broader Earth AI geospatial modeling effort.