Abstract
Urban region embedding is a critical task for predicting neighborhood characteristics such as crime rates and land usage. However, raw human mobility data is often redundant and fails to capture the underlying functional identity of a region. We present Multi-Graph Fusion Networks (MGFN), a framework that distills complex mobility data into distinct spatio-temporal patterns. By clustering similar mobility states and employing a dual-attention mechanism, MGFN learns generalized embeddings that significantly outperform state-of-the-art models in urban characteristic prediction.
Introduction
Urban planners rely on mathematical vectors known as "region embeddings" to represent the characteristics of city neighborhoods. While human mobility data (e.g., taxi trips) provides a rich source of information, it presents two major challenges. First, existing models often treat mobility as a raw time series, leading to redundancy rather than generalization. Second, a region's identity is dynamic; a business district behaves differently on a weekday morning than on a weekend night. Capturing these shifting "personalities" into a single, robust representation remains a significant technical hurdle.
Methodology
The MGFN framework addresses these challenges through a two-stage process: Mobility Graph Fusion and Mobility Pattern Joint Learning.
1. Mobility Graph Fusion (MGF)
Instead of processing data hour-by-hour, MGFN uses a novel metric called Mobility Graph Distance (MGD) to cluster similar time periods into $N$ distinct "Mobility Patterns." MGD evaluates unidirectional flow (the balance of people entering vs. leaving) and spatial structure (high-intensity connections). This allows the model to automatically group functional states, such as "weekday morning peaks" or "weekend late nights."
As shown in the visualization below, these patterns capture the inherent variation and similarity between regions across different days of the week, moving beyond simple sequential analysis.
2. Mobility Pattern Joint Learning
Once patterns are identified, the model utilizes a two-tier attention architecture:
- Intra-Pattern Message Passing: Within a specific pattern (e.g., Morning Peak), regions share information with neighbors based on mobility flow strength, utilizing a Graph Attention Network (GAT) approach.
- Inter-Pattern Cross-Attention: This layer correlates a region's behavior across different patterns, asking how its role during the workday relates to its role during the weekend.
The final output is a unified embedding that captures the multi-faceted nature of the urban space.
Evaluation and Key Insights
The researchers evaluated MGFN using Manhattan taxi records (~10 million trips) and validated it against crime counts, Foursquare check-ins, and land-use data.
Key Results
- Superior Accuracy: MGFN outperformed the second-best model (MVURE) by 13.11% in $R^2$ for check-in prediction and achieved the lowest error rates (MAE and RMSE) for crime prediction.
- The "Uniflow" Signal: Net flow (inflow minus outflow) proved to be a powerful indicator of functionality. Residential areas typically "drain" in the morning, while office hubs "fill up."
- Generalization: MGFN, trained only on mobility data, outperformed models with access to multiple data sources (POI and check-in data), suggesting that well-processed mobility data is a sufficient proxy for many urban features.
Conclusion
MGFN demonstrates that clustering mobility data into functional patterns is more effective for region embedding than treating it as a sequential time series. While the current study relies on taxi data in NYC, future work will explore multi-modal graphs (subways, buses) and extend the framework to tasks like house price prediction and epidemic spread modeling.
Summary updated