In 1970, Waldo Tobler formulated the First Law of Geography: "Everything is related to everything else, but near things are more related than distant things." For decades, this principle has anchored how we model the world. If you want to predict traffic in a city, you look at the neighboring streets. If you want to predict the temperature in a valley, you look at the surrounding peaks.
However, a growing body of research suggests that in complex, modern systems, Tobler’s law is increasingly insufficient. A residential suburb in New Jersey may have more in common with a suburb in Illinois than with the industrial shipyard just two miles away. In these scenarios, functional similarity—how an entity behaves—is a more powerful predictor than geographic proximity.
The Failure of the "City as a Video"
Traditional spatial-temporal models often treat urban data like a video: a grid of pixels where each cell represents a city block. As Jinguo Cheng, Yuxuan Liang, and Lijun Sun argue in Rethinking Urban Mobility Prediction, this "computer vision" lens prioritizes local spatial features but fails to capture the long-distance commuting flows that define a city’s pulse.
When we rely solely on Euclidean distance, we become "topology blind." As noted by Guoheng Sun and colleagues in NetSight, this blindness prevents models from recognizing that two physically distant base stations might exhibit identical traffic patterns because they both serve similar commercial hubs.
Constructing Semantic Adjacency
To move beyond physical constraints, researchers are turning to Graph Neural Networks (GNNs) to define a new kind of neighborhood: the semantic neighbor. Instead of connecting nodes based on how many meters apart they are, these models connect nodes based on their "behavioral fingerprints."
- Functional Roles: In Spatial-Temporal Demand Forecasting and Competitive Supply via Graph Convolutional Networks, Bolong Zheng and Jilin Hu demonstrate that connecting distant business districts with identical demand cycles allows a model to share representational knowledge across the city.
- Dynamic Correlations: In Graph WaveNet, Zonghan Wu and Shirui Pan introduce a self-adaptive adjacency matrix that learns these hidden relationships directly from the data, bypassing the need for a pre-defined map.
- Pattern Similarity: For meteorological forecasting, Xun Zhu uses Weather2K to show that distant stations with similar historical data sequences provide better predictive signals than nearby stations with different microclimates.
Teleconnections and Global Dependencies
This shift isn't just for cities. In climate science, Nils Bochow and Niklas Boers highlight the importance of teleconnections—physical links between climate patterns in geographically remote regions. In Reconstructing Historical Climate Fields With Deep Learning, they show that without accounting for these long-range dependencies, models cannot accurately fill gaps in historical records where local data is missing.
Similarly, in neuroscience, Orhan Firat developed a framework for Discriminative Functional Connectivity that identifies "functional neighborhoods" in the brain. These are groups of voxels that work in tandem to process concepts, even if they are located in entirely different lobes.
Why It Works: Knowledge Sharing
The intuition behind functional similarity is simple: if two things act the same, they can learn from each other.
In Deep Multi-View Spatial-Temporal Network for Taxi Demand Prediction, the authors use graph embeddings to represent semantic proximity. This allows the model to "transfer" what it knows about a busy morning peak in one residential zone to another similar zone across town.
This approach is particularly effective for "cold-start" problems. Peter A. Flach explores this in Gaussian Process Model Re-Use, where new bike-sharing stations are matched with older ones that share a similar "temporal fingerprint." By recycling models from behavioral twins rather than geographic neighbors, researchers can make accurate predictions even when data at a specific site is scarce.
The Future of Relational Modeling
From predicting epidemic spreads via growth trajectories to optimizing 6G networks by clustering base stations with similar traffic profiles, the research direction is clear. We are moving away from rigid, Euclidean grids toward flexible, relational graphs. By prioritizing functional behavior over physical location, these models capture the true, non-local essence of the complex systems that define our world.