Machine Learning Models for Predicting Crime Hotspots in Medellin City

Crime is one of the most severe social problems in developed and emerging countries, and it affects welfare, economy, and regional development. The study of criminal behavior is significant to the formulation of new law enforcement strategies for crime prevention and control. In addition, some governments have implemented projects in a smart cities context to make available some databases related to various phenomena that occur inside their cities, including those associated with criminal acts. Medellin, Colombia, has several public databases where it is possible to get information about mobility, education, infrastructure, health, and security. In this work, some Machine Learning models are developed and evaluated for crime hotspots prediction using available geographical, temporal and demographic information available in Medellin for a period between 2015 and 2019. The work analyzes the implications of selecting geographic divisions and the separation of time intervals in the models. As a result, it was found that the models based on Decision Trees present an F1-Score of 88.2 % and have better performance than models based on Logistic Regression and Neural Networks (MLP).