paper in Computers, Environment and Urban Systems
Osoro, Ogutu B., Jingjing Li, Ming Tang, Xi Chen, Kelly Merrill, Lauren Forbes, Briana L. Simms, Paris Wheeler, Clementine Adeyemi, and Diego F. Cuadros. “Family Specific Spatial Borrowing in Multimodal Urban Reconstruction: A Leave-One-Neighborhood-Out Study of Cincinnati.” Computers, Environment and Urban Systems 132 (2027): 102541. https://doi.org/10.1016/j.compenvurbsys.2026.102541.
Family Specific Spatial Borrowing in Multimodal Urban Reconstruction: A Leave-One-Neighborhood-Out Study of Cincinnati.
Urban spatial models often assume that nearby borrowing improves prediction across target variables, but this assumption is rarely tested across distinct urban feature families within a city. We evaluate this question on a citywide hexagonal grid of Cincinnati, Ohio (2016 cells, 60 features in 14 feature families) using multimodal masked reconstruction. We compare four models spanning linear, local nonlinear, shared-gate graph, and selective-gate graph architectures under full leave-one-neighborhood-out evaluation across 50 neighborhoods, with primary summaries for the 36 neighborhoods having at least 20 cells. Median held-out-neighborhood R2 is the primary comparison metric. Cross-neighborhood transfer is family-specific, and no single borrowing strategy dominates. Ridge performs best for street structure, indicating that some spatially regular families are best reconstructed from local linear combinations. The local nonlinear masked autoencoder (MLP-MAE) holds the highest median for most families with positive transfer, though these differences are not statistically distinguishable from the graph models, indicating that local nonlinear multimodal fusion is often sufficient. Graph-based models help most clearly for topography, while the selective-gate graph model wins many individual neighborhood comparisons for several families despite lower medians, reflecting different error profiles. Graph models also produce more neighborhood-structured representations: contextual embeddings from the selective multimodal spatial autoencoder (SMSA) recover neighborhood structure better than raw principal components or local-only embeddings. Learned gates are stable across folds and provide interpretable diagnostics of context reliance, but not a universal ranking of borrowing regimes. These results show that the appropriate spatial borrowing strategy is target-specific rather than universal. For neighborhood-scale interpolation, imputation, and urban monitoring, one spatial modeling recipe can systematically misrepresent some feature families even within the same city.
