Abstract
Traditional Origin-Destination (OD) matrices often fail to capture the spatial and structural nuances of human mobility, treating location pairs as independent statistics. This paper introduces the "Mobility Tableau," a vector graph representation of population flow, and proposes a novel similarity metric based on the minimum cost to transform one graph into another. By decomposing transformation costs into volume, spatial, mass, and structural components, this approach enables more robust model evaluation and cross-city urban planning.
Introduction
Urban planners and researchers typically rely on Origin-Destination (OD) matrices to analyze human mobility. However, these matrices treat every location pair as an independent data point. A prediction error of one city block is penalized as harshly as an error of ten miles. Existing alternatives like MSSIM or Wasserstein distances often ignore geographic correlation or are computationally prohibitive. There is a clear need for a holistic metric that captures not just volume, but the spatial patterns and structural integrity of movement.
Methodology: The Mobility Tableau
The authors propose the Mobility Tableau, a vector graph where arrows represent population flow between geographic grids. To compare two tableaus, they calculate the minimum spatial cost to transform one into the other using three basic operations: Add, Delete, and Shift.
The cost of these operations is measured using Manhattan distance, framed as a bipartite graph maximum weight matching problem solved via a Point Set oriented Kuhn-Munkres algorithm. This allows for the derivation of normalized metrics: NMD (volume similarity), SP (spatial similarity), NMA (mass inclusiveness), and RRNSA (structural similarity).
Key Insights and Results
A primary advantage of this method is that geography is inherent to the error calculation. The "Shift" operation recognizes that proximity matters, unlike traditional RMSE.
As shown in Figure 2, traditional metrics like MAE and RMSE often fail to distinguish between different types of mobility errors. Mobsimilarity, however, can isolate spatial shifts from volume changes, providing a more nuanced understanding of model performance.
Robustness and Sparse Data
In tests using Tokyo GPS data and simulated sparse Call Detail Record (CDR) data, the structural similarity metric (RRNSA) proved remarkably robust. While traditional $R^2$ dropped significantly with sparse data (to 0.31), RRNSA remained high (96.47%), indicating that sparse data still captures the underlying movement patterns of a city even when raw volume counts are inaccurate.
Cross-City Application
The metrics also facilitate cross-city transfer learning. By decoupling structure from volume, researchers identified Osaka's Chuoku district as the structural equivalent of Tokyo's Shinjuku district—a match that traditional volume-based metrics failed to uniquely identify due to differences in total population density.
Conclusion and Future Work
Mobsimilarity provides a multi-dimensional lens for mobility analysis, moving beyond simple volume counts to capture the spatial and structural essence of urban flow. While the current $O(n^3)$ complexity of the Kuhn-Munkres algorithm poses a challenge for very large grids, future optimizations and the expansion into 3D spatial-temporal graphs promise to further enhance its utility for urban science.
Summary updated