Abstract:

Spatial prediction of street robbery is a central task in criminal geography, providing critical guidance for the efficient governance of urban public security. However, most existing spatial crime prediction models rely on correlation analysis to establish relationships between environmental variables and criminal incidents, often overlooking the underlying causal mechanisms linking urban environments to street robbery. To bridge this gap, this study proposes a novel spatial prediction framework that integrates causal inference with geographic machine learning. First, the greedy equivalence search (GES) algorithm is employed to identify points of interest (POI) features that exert potential causal effects on robbery. Second, the geographically weighted random forest (GWRF) model is utilized to quantify the nonlinear and spatially heterogeneous relationships between the environmental features and robbery incidents. Using empirical data from Manhattan, New York city, the proposed framework demonstrates superior performance. The experimental results reveal that the GES algorithm effectively filters out POIs with causal implications, while the GWRF model successfully captures complex spatial variations. By incorporating causal inference, the proposed method not only enhances the precision of spatial street robbery prediction but also provides more scientific and actionable decision support for urban crime prevention and targeted interventions.

Key words: street crime prediction, causal discovery, greedy equivalence search algorithm, geographic machine learning