Journal of Geographical Systems География, планирование и развитие
How inter-city differences affect cross-city crime spatiotemporal prediction performance under few-shot scenarios: evidence from 12 U.S. cities
Аннотация
Transfer learning has been recognized as an effective approach to enhance model performance in urban spatiotemporal prediction under few-shot conditions. However, how inter-city structural characteristics influence its effectiveness in cross-city crime prediction remains underexplored. To address this issue, we conducted a systematic evaluation across 12 major U.S. cities at two spatial resolutions, covering property and violent crime types over multiple time periods. The results were interpreted from the perspectives of Routine Activity Theory and Crime Pattern Theory to understand cross-city transferability. Findings suggest substantial heterogeneity in cross-city crime spatiotemporal prediction performance across crime types and spatial scales. Overall, transfer learning consistently improves predictive performance for both property and violent crime, with significant improvements observed in most evaluated scenarios. However, the magnitude of improvement varies considerably across cities and spatial resolutions. For property crime, transfer performance is relatively stable across scales and is associated with urban functional organization and spatial opportunity structures in the source city. This pattern is broadly consistent with the opportunity-based mechanisms and stable routine activity patterns described by these theoretical frameworks. In contrast, violent crime prediction exhibits scale dependency, with transfer performance more strongly associated with local urban characteristics at the 1 km scale, while exhibiting different association patterns at the 2 km scale. In summary, this research uncovers the complex interactions among urban environments, spatial scales, and the transferability of crime spatiotemporal prediction models, highlighting the roles of spatial scale and urban structural heterogeneity in shaping the generalizability of crime spatiotemporal patterns across cities.