Artificial intelligence has become remarkably good at understanding text, images, speech, and other forms of digital information. But understanding the physical world requires more than knowing what something is. It also requires understanding how people actually use it.
A coffee shop, for example, may be described online simply as a café. Yet in reality, it could function as a morning meeting spot, a remote-work destination during the afternoon, or a social gathering place in the evening. Similarly, a park may serve different purposes depending on the time of day, from jogging and dog walking to family activities and relaxation.
This gap between what a place is and how people use it is the problem Google Research is exploring with ME-POIs, short for Mobility-Informed Place Embeddings for Points of Interest.
The framework combines traditional text-based representations of Points of Interest (POIs) with aggregate human mobility patterns, giving AI a richer way to understand places and their real-world context.
Why Text Alone Cannot Fully Describe a Place
Modern mapping and recommendation systems rely heavily on information such as business names, categories, reviews, descriptions, websites, and other textual data.
This information can provide a strong semantic understanding of a location.
For example, an AI system can determine that a particular location is a:
- Restaurant
- Coffee shop
- Library
- Shopping centre
- Park
- Gym
- Hotel
It may also identify additional characteristics from descriptions and reviews, such as cuisine, atmosphere, price range, or facilities.
However, textual information doesn’t necessarily reveal how a location is used throughout the day.
A restaurant could become a popular breakfast destination in the morning, a business lunch venue at noon, and a takeaway hub during the evening. A public park could attract runners early in the morning, families during the afternoon, and visitors looking for relaxation later in the day.
These behavioral patterns provide important context that conventional text-based POI representations may not capture.
What Is ME-POIs?
ME-POIs is designed to add this missing behavioral dimension to POI representations.
Instead of relying only on descriptions and textual information, the framework incorporates aggregate mobility information to learn how people interact with different places.
The central idea is relatively simple:
Text can help AI understand what a place is, while mobility patterns can help it understand how the place is used.
By combining these two sources of information, ME-POIs aims to create richer representations of locations.
How ME-POIs Works
The framework uses several important techniques to turn mobility information into useful representations of places.
1. Learning From Aggregate Human Movement
ME-POIs incorporates aggregate and anonymized mobility information to identify broader patterns in how people visit locations.
The objective is not to understand the behavior of a particular individual. Instead, the system looks for population-level patterns associated with places.
For example, a location may show a strong pattern of short visits during weekday mornings, while another location may attract longer visits primarily during weekends.
These patterns can provide clues about how different locations are used.
2. Creating Contextualized Visit Representations
A visit is more informative than simply counting how many people went to a location.
ME-POIs can represent visits using contextual information such as:
- Time of day
- Duration
- Visit patterns
- Nearby locations
- Places visited before or afterward
This additional context can help distinguish different types of visits.
For example, two locations might both receive hundreds of visitors per day, but their usage patterns could be completely different.
One might attract commuters for short visits, while another might attract visitors who stay for several hours.
3. Using Contrastive Learning
ME-POIs also uses contrastive learning to learn meaningful relationships between mobility patterns and POI representations.
In simple terms, the model learns to identify which behavioral patterns are more closely associated with particular places or types of places.
Over time, this can allow the system to recognize recurring usage patterns such as:
- Morning coffee visits
- Commuter activity
- Lunch-time business visits
- Evening dining
- Weekend recreation
- Longer study sessions
The objective is to build a more detailed behavioral representation of each location.
Tackling the Long-Tail Problem
One of the more interesting aspects of ME-POIs is its approach to the long-tail problem.
Mapping platforms contain millions of locations, but not all POIs have the same amount of information or user activity.
Popular locations may have:
- Large numbers of reviews
- Extensive descriptions
- Significant visitor activity
- Strong historical data
New or less popular locations may have very limited information.
This creates a challenge known as the cold-start problem.
ME-POIs addresses this challenge through a form of transfer learning. Information from data-rich locations, or “anchors,” can help provide useful behavioral information for less-documented POIs.
The approach considers relationships across different spatial scales, allowing information from relevant locations to contribute to understanding locations with limited data.
For example, a newly opened bookstore with very little historical information could potentially benefit from behavioral patterns learned from established bookstores with similar characteristics.
Reported Results
The reported evaluation provides encouraging evidence for the approach.
In experiments involving five map-enrichment tasks and mobility data from Los Angeles and Houston, incorporating ME-POIs improved performance across 34 of 35 evaluated metrics.
These results suggest that adding mobility information can provide meaningful improvements over representations based primarily on conventional textual information.
The findings also highlight a broader opportunity for AI systems to combine semantic information with behavioral signals when attempting to understand physical locations.
Potential Applications of Mobility-Informed AI
The ability to understand how places are used could have applications across mapping, recommendations, urban planning, and business intelligence.
Smarter Maps and Navigation
Future mapping systems could potentially provide more contextual information about locations rather than simply identifying their category.
For example, a system could distinguish between places that are typically busy at different times or identify locations associated with particular usage patterns.
This could make local search and navigation more context-aware.
More Personalized Recommendations
Recommendation systems could move beyond simple categories such as “restaurants near me” or “cafés nearby.”
Instead, AI could potentially learn behavioral characteristics associated with places.
For example, someone looking for a quiet location to work might receive recommendations based on usage patterns rather than relying solely on reviews describing a venue as “quiet.”
Urban Planning and Smart Cities
Understanding how public spaces and commercial areas are used throughout the day could provide valuable insights for city planners.
Mobility-informed analysis could potentially help evaluate:
- Public-space utilization
- Transportation demand
- Pedestrian movement
- Commercial activity
- Infrastructure requirements
- Timing of public services
Such information could contribute to more efficient planning and resource allocation.
Retail and Business Intelligence
Businesses could potentially use location intelligence to better understand when and how customers interact with different areas.
This could support decisions involving:
- Store locations
- Staffing
- Operating hours
- Inventory planning
- Customer experience
- Local market analysis
Location-Based Services
More detailed place representations could also improve location-based services and advertising by providing richer contextual information about locations and their surrounding environments.
However, applications involving individual-level tracking would require particularly careful attention to privacy, consent, data governance, and applicable regulations.
Privacy and Responsible Use Matter
Mobility data can be extremely valuable for understanding cities, but it is also sensitive.
The important distinction is between aggregate behavioral patterns and identifying or tracking individual people.
Systems built using mobility information must be designed with strong privacy protections and appropriate safeguards. Responsible data handling will be essential if mobility-informed AI becomes increasingly integrated into consumer services, urban planning, and commercial applications.
The long-term value of technologies such as ME-POIs will therefore depend not only on their technical performance but also on how responsibly the underlying data is collected, processed, and used.
Google Research’s Broader AI Direction
ME-POIs reflects a broader evolution in artificial intelligence: moving from understanding isolated pieces of information toward building richer representations of the real world.
Text tells an AI system what a place is. Images can show what it looks like. Maps provide its geographic context. Mobility patterns can reveal how people interact with it.
Combining these different forms of information could allow AI systems to develop a much more comprehensive understanding of physical environments.
This could eventually lead to digital systems that are better at answering not only “What is this place?” but also “How is this place typically used?” and “Why might it be relevant right now?”
The Future of AI-Powered Place Understanding
The development of ME-POIs points toward a future in which mapping and location intelligence become increasingly dynamic and context-aware.
Instead of treating a Point of Interest as a static entry in a database, AI could represent it as a changing entity whose characteristics vary according to time, activity, surrounding locations, and human behavior.
That could make maps and recommendation systems significantly more useful.
The technology is still part of a broader research journey, and questions around reliability, scalability, privacy, and real-world deployment remain important. Nevertheless, the underlying concept is compelling.
Conclusion
Google Research’s ME-POIs demonstrates how AI can go beyond the traditional “what” of a place and begin modeling the “how”.
By combining text-based POI information with aggregate mobility patterns, the framework aims to create richer representations of locations and improve the way AI understands the physical world.
The reported results across mapping-related tasks suggest that mobility information can add meaningful context to conventional place embeddings.
As AI increasingly connects digital information with real-world environments, approaches such as ME-POIs could help create smarter maps, more relevant recommendations, better urban insights, and more context-aware location services.
The next generation of location intelligence may not simply tell us where a place is or what it is. It could also help us understand how people use it, when they use it, and what that behavior tells us about the place itself.