DETERMINANTES E PREDIÇÃO DE CRIMES DE HOMICÍDIOS NO BRASIL: UMA ABORDAGEM DE APRENDIZADO DE MÁQUINA

Resumo

Throughout history, organized societies have attempted to prevent crime by various approaches. In this context, the objective of this study is to identify the economic, social and demographic determinants of homicide and property crime in Brazil. As a methodology, this is a quantitative study where Regression Tree, Random Forests, Boosting and K-Nearest Neighbors methods were used as alternative tools to traditional linear models, such as least squares regression. The main justification focuses on the limitations of linear methods, such as non-normality of data, multicollinearity between social factors, exogeneity and wrong selection of variables. The data analyzed indicate that among the 32 covariates used to represent social, economic and demographic factors, 8 had the greatest impacts on violence at the national level, such as the size of the young population, basic sanitation, total population size, economically active population, urban population, GDP, female heads in the family, poor people between 0 and 14 years old and the proportion of people earning up to half a minimum wage, where each factor was discussed according to the economic literature of the crime. In addition, the Random Forests model exaplain, in averaged, 85% of homicide and estate crime variability at the national level. We believe that this approach helps to produce more robust responses on the effects of social, economic, and demographic factors on crime and is therefore a new tool for guiding policymakers.