Mining large-scale human mobility data for long-term crime prediction

Traditional crime prediction models based on census data are limited, as they fail to capture the complexity and dynamics of human activity. With the rise of ubiquitous computing, there is the opportunity to improve such models with data that make for better proxies of human presence in cities. In this paper, we leverage large human mobility data to craft an extensive set of features for crime prediction, as informed by theories in criminology and urban studies. We employ averaging and boosting ensemble techniques from machine learning, to investigate their power in predicting yearly counts for different types of crimes occurring in New York City at census tract level. Our study shows that spatial and spatio-temporal features derived from Foursquare venues and checkins, subway rides, and taxi rides, improve the baseline models relying on census and POI data. The proposed models achieve absolute R^2 metrics of up to 65% (on a geographical out-of-sample test set) and up to 89% (on a temporal out-of-sample test set). This proves that, next to the residential population of an area, the ambient population there is strongly predictive of the area's crime levels. We deep-dive into the main crime categories, and find that the predictive gain of the human dynamics features varies across crime types: such features bring the biggest boost in case of grand larcenies, whereas assaults are already well predicted by the census features. Furthermore, we identify and discuss top predictive features for the main crime categories. These results offer valuable insights for those responsible for urban policy or law enforcement.