Time-Series Features for Predictive Policing

Forecasting when and where crimes are more likely to occur based on years of historical record analysis is becoming a task which is increasingly helping cities' safety departments with capacity planning, goal setting, and anomaly detection. Crime is a geographically concentrated phenomena and varies in intensity and category over time. Despite its importance, there are serious challenges associated with producing reliable forecasts such as sub-regions with sparse crime incident information. In this work, we address these challenges proposing a crime prediction model which leverages features extracted from time series patterns of criminal records based on spatial dependencies. Our results benchmarked against the state of the art and evaluated on two real world datasets, one from San Francisco, US, and another from Natal, Brazil, show how crime forecasting can be enhanced by leveraging Spatio-Temporal dependencies improving our understanding of such models.