Mobility-Embedded POIs: Learning What a Place Is and How It’s Used from Human Movement

ICLR 2026 Conference Submission21364 Authors

Keywords: poi embeddings, mobility-augmented embeddings, map enrichment, geospatial applications

Abstract: Recent progress in geospatial foundation models has highlighted the importance of learning general-purpose representations for real-world locations, particularly Points of Interest (POIs) where human activity concentrates. Yet, existing POI representations remain largely static, evolving from simple coordinates and metadata to visual features and, most recently, LLM-derived textual prompts, all of which describe what a place is, but not how it is actually used. We argue that human mobility provides a complementary and dynamic signal, capturing real-world visitation patterns that reveal how places function in practice. To this end, we introduce Mobility Embedded POIs (ME-POIs), a pretraining framework that augments static text-embedding representations with mobility-derived signals from visit sequences, capturing dynamic usage patterns. Each visit is represented as a contextualized embedding that integrates the POI’s static attributes with its temporal and sequential context, including when the visit occurs and which visits precede or follow it. To address the long tail of sparsely visited POIs, we transfer visit distributions from data-rich locations to sparse ones, leveraging multi-scale spatial proximity to capture local and regional patterns. We evaluate ME-POIs on large-scale human mobility datasets across a set of map enrichment tasks. We find that augmenting strong text embedding baselines with ME-POIs leads to consistent and substantial improvements across all tasks, confirming that mobility-informed embeddings offer complementary information that enhances static representations and enables a richer understanding of how places are used. Notably, even mobility embeddings alone, without any POI semantics, outperformed text-based embeddings on certain tasks, underscoring a key novelty of our approach.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 21364