Priorização de variáveis explicativas na modelagem de acidentes de trânsito utilizando técnicas de aprendizado de máquina

Abstract

The prioritization of variables in the process of accident modeling can contribute to the optimization of resource and indication of which data are priority to collect. This study aimed to investigate the influence of the prioritization of variables in the adjustment of accident predictive models of multivariate response (number of accidents without victims, number of accidents with victims and number of accidents with deaths). Two approaches were used: decision tree grouping techniques (Random Forest and Boosted Trees) for initial prioritization and later modeling using artificial neural networks (ANN); and, direct use of ANN for both, prioritization and modeling. Overall results, however, indicated a worse fitness of the models when the number of explanatory variables was reduced. Despite this, it is believed that the evolution of machine learning techniques that best deal with multivariate response, lead to the adequate identification of the most important variables for modeling.