Un enfoque de aprendizaje automático para la predicción del delito en la Ciudad Autónoma de Buenos Aires

Abstract

Crime is undoubtedly a problem that affects all nations and governments worldwide. Therefore, its prevention is part of the agenda for each of them. The objective of this thesis is to demonstrate, through a machine learning approach, that it is possible to estimate the place and time where a crime will occur in the future. Particularly, it aims to determine whether crimes are truly random or if they are simultaneously affected by a set of spatial-temporal variables in the Autonomous City of Buenos Aires (CABA). A model with these characteristics, if successful, would allow for a more precise allocation of patrol officers and police from CABA’s security forces. The obtained results suggest that, compared to a naive model, machine learning algorithms are vastly superior, and it is possible to determine the number of crimes expected in the following month. This work details the different datasets used to enrich crime records, as well as the efforts made to create a grid that will serve as a starting point for estimating the models. Additionally, it explains the tradeoff generated when choosing a grid size for the analysis.