Self-exciting Point Processes with Image Features as Covariates for Robbery Modeling

State-of-the-art crime prediction models exploit the spatio-temporal clustering patterns and the self-exciting nature of criminality to predict vulnerable crime areas. However, omitting spatial covariates correlated with the occurrence of crimes potentially bias the estimated parameters. This research combines self-exciting point processes, generalized additive models and environmental attributes extracted through convolutional neural networks from street-level images to predict robbery hotspots across the locality of Chapinero in Bogota, Colombia. Our model using image features as covariates outperforms a standard self-exciting point process and shed light on the association between crime occurrence and the socioeconomic and environmental conditions of the city.