Crime risk prediction incorporating geographical spatiotemporal dependency into machine learning models

The spatiotemporal distribution of crime is closely related to the environment, exhibiting a typical characteristic of "spatiotemporal autocorrelation". However, most of the existing machine learning-based crime prediction methods have difficulty in simulate the spatiotemporal dependence of crime. In this study, we mitigate the spatiotemporal dependence embedded in crime data by introducing a spatiotemporal lag variable. To verify the feasibility of the proposed methods, four machine learning methods were used to determine whether considering spatiotemporal dependency could improve model prediction accuracy and explore the impact of various factors (i.e., environmental factors and demographical factors) on crime risk intensity in different locations using crime data collected from June 2014 to May 2018 in Dallas. The results indicated the following: (1) incorporating spatiotemporal lag variables can effectively improve the prediction accuracy of machine learning models; (2) variables predicting crime are highly nonlinear over time and space, and tree-based nonlinear models greatly outperform linear models in predicting crime; and (3) interpretable machine learning models can reveal the unique contribution of each variable to researchers and practitioners. These findings contribute to our understanding of the mechanism of crime occurrence and may guide the development of crime prevention strategies.