Google Research、移動データ活用で場所理解を深める言語モデル枠組みを発表

We introduce a dynamic, mobility-informed framework that allows AI models to understand the temporal activity rhythms of places over time, and, in doing so, significantly improve predictions about real-world attributes like opening hours, price levels, and busyness.

Artificial intelligence has made incredible progress in understanding the world through text. However, to build AI models that truly understand the physical world, they must comprehend more than just words: they need to capture the dynamic, real-world functionality of the built environment. Every place has two distinct signatures: its identity on paper, and its actual functional rhythm.

Traditional language models typically build representations of places (commonly referred to as “points of interest” or POIs), whether it’s a business or a place like a park or landmark, by relying heavily on this static metadata. They successfully analyze addresses, business categories, and text descriptions. While world-class language models like Gemini are incredibly proficient at processing text data, their geospatial representations can be significantly enriched by incorporating the real-world functional dynamics of the urban environment. Complementing semantic labels with mobility data can enable these models to effectively capture the unique temporal activity rhythms of POIs in a city.

To demonstrate this complementary capability, we introduce Mobility-Embedded POIs (ME-POIs), a novel framework that improves text-based place representations derived by language models. Using publicly available benchmark datasets, ME-POIs incorporates aggregated and anonymized mobility patterns, such as arrival times, stay durations, and surrounding movement patterns. Rather than treating a place as a frozen set of words, ME-POIs use a self-supervised approach to blend text descriptions with large-scale, anonymized mobility patterns from public benchmarks (capturing the aggregate spatial activity footprints of the environment throughout the day). In doing so, the model constructs a numerical vector representation (a mathematical "signature", technically called an embedding) that encodes both the identity of a place and its dynamic functionality. Integrating ME-POIs with advanced text models delivered a context advantage that yielded 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 across unseen places.

How the ME-POIs framework works

By providing a pre-enriched representation of a place, the ME-POIs framework makes it vastly easier for AI models to draw accurate inferences about many distinct attributes — such as operating hours, target price levels, and current business status — without needing to calculate those attributes from scratch every single time.

To build a model with a deeper understanding of places, we must distinguish between a place's *identity* (its name and category) and its *function* (its aggregated visit footprint). In prior geospatial AI research works, mobility patterns were almost exclusively applied to predicting the next POI a user will visit. In contrast, the ME-POIs framework shifts mobility from an output prediction task to an input feature that defines the place itself.

But how do you transform raw geographic points into a clean and numerically useful mathematical signature (or embedding)? We accomplish this through a three-step pipeline: visit alignment, spatial multiscale visit propagation, and text-mobility synergy.

Visit alignment

The model treats aggregate visits to a specific POI as fundamental data points. It analyzes the temporal arrival windows, departure trends, and typical stay durations. Rather than calculating simple averages, a temporal encoder maps these temporal sequences into a dense vector space. This process establishes a "functional centroid" — a unique, multidimensional signature that maps the aggregate anonymized mobility patterns associated with that specific place over a one-year cycle and across different days of the week.

Solving the "long tail" of data sparsity

A persistent challenge in geospatial data science is the "long tail" problem. While famous landmarks, massive shopping malls, and popular downtown chains generate an abundance of visit data, the vast majority of local businesses — small neighborhood boutiques, specialized repair shops, or newly opened cafes — suffer from severe data sparsity. Previously, when a model encountered a place with few or no recorded visits, it would incorrectly assume the place had zero activity, leading to broken predictions.

ME-POIs solves this through a novel spatial multiscale visit propagation mechanism. The architecture recognizes that visits are usually regionally constrained; a small boutique on a high-end shopping street shares systemic behavioral traits with its neighbors. The framework looks at adjacent places across multiple spatial scales: the immediate street, the block, and the wider neighborhood. It then statistically transfers the aggregated visit patterns of busy, data-rich neighbors to nearby sparse places. By learning the regional "rhythm" from active areas, the model applies an intelligent geographical prior knowledge base to smaller shops, enabling it to learn something about them even when they have few to no appearances in the data.

Text-mobility synergy

Rather than discarding textual descriptions, the ME-POIs framework enriches them. It does so by aligning high-level language embeddings (the standard vector representations extracted from advanced models like Gemini) with the newly generated mobility vectors by maximizing their cosine similarity. We layer the mobility signal directly on top of the language representation, creating a more holistic view of the place or business.

This hybrid approach ensures that the model preserves structural semantics (e.g., knowing that a place sells food from its text description) while absorbing its operational context (e.g., knowing whether it functions as a lunch spot or a late-night diner from its mobility signature).

Experiments

To evaluate whether ME-POIs truly achieves a generalized, attribute-agnostic understanding of physical places, we conducted extensive testing across two large, culturally distinct metropolitan areas: Los Angeles and Houston.

We evaluated the framework on five distinct tasks. Crucially, to prove the model develops general intelligence rather than just memorizing local patterns, we trained the framework on a set of observed places and then asked it to predict attributes on entirely unseen places.

The five downstream tasks are as follows:

• Opening/closing hours prediction: Can our place embeddings help us infer the exact schedule of a business?
• Price-level classification: Can our place embeddings help us distinguish between a luxury high-end boutique and a thrift store based only on mobility context?
• Permanent closure detection: Businesses often "go dark" and close permanently long before an owner updates an online profile or a crowd-sourced report is filed. Can our embeddings help us flag such closed businesses?
• Visit intent classification: Can our embeddings help us estimate aggregate search and navigation interest for a place? This task serves as a proxy for the overall popularity of a place.
• Busyness forecasting: Predicting future crowd densities and peak hour dynamics of businesses.

To establish a baseline, we compared the ME-POIs framework to standard text-only embedding models (like Gemini embeddings), existing trajectory-based geospatial models (like TrajGPT), and hybrid variations to isolate exactly how much value the mobility patterns added to the equation.

結果

The experimental results were striking, confirming that adding real-world mobility data to existing text models provides a noticeable "context advantage”. When evaluating model performance across unseen test places, the integration of ME-POIs led to substantial accuracy gains, consistently outperforming both purely text-based and mobility-based baselines across all predictive tasks.

Beyond outperforming standard baselines, one of the most notable findings emerged when comparing a model trained *exclusively on mobility data* against those that had access *only to text metadata*. In several cases, like in price-level classification, the mobility-only model surpassed the text-only language models. This reveals a fascinating and often underused truth about urban dynamics: our collective actions at a physical place are often far more descriptive than the formal words used to label it.

まとめ

By successfully moving beyond static digital labels, the ME-POIs framework demonstrates a new way to enable AI to model and understand the physical world. It’s important to emphasize that this framework focuses on understanding things about the world in aggregate — it cannot draw conclusions about individual users or anything personalized. That is, the ME-POIs framework can provide a holistic representation of a place or business, capturing how it’s visited across broad populations and time frames; it can’t be used for individual personalization. Instead, its power lies in creating rich aggregate numerical signatures of places that reduce the computational burden on downstream systems when they’re required to make inferences about those places.

Ultimately, ME-POIs lay the groundwork for AI models that truly understand the rhythm of our cities. It is also part of our broader Google Earth AI effort to create geospatial models and datasets that turn planetary data into actionable intelligence.

Acknowledgements

*We thank our co-authors on this paper: Neha Arora (Google Research), alongside Prof. Cyrus Shahabi and Ph.D. student Shang Ling Hsu from the University of Southern California (USC).*