Seleção de principais atributos para Redes Neurais Artificiais do tipo MLP: um estudo de caso sobre mineração de dados para diagnóstico de dengue.

Abstract

This study proposes to investigate which attributes are most significant for predicting the diagnosis of Dengue using attribute selection and MLP neural networks. In this study, a database was used by the SINAN Online - Notification of Injury Information System, of the Ministry of Health, which is a real database and public domain. The data preprocessing step was performed to optimize the base adaptation to the data mining algorithms.Feature selection was provided by selecting a ranking of best attributes according to the eight feature selection algorithms: ChiSquareAttributeEval, FilteredAttributeEval, GainRatioAttributeEval, InfoGainAttributeEval, OneRAttributeEval, ReliefFattributeEval, SVMAttributeEval, and SymmetricalUncertAttributeEval. Through the use of the MLP classifier algorithm, it was possible to identify the best attributes in the cross validation and normal validation approaches (with split 70%), testing the subset of the best attributes defined by the ranking. Thus, it was also possible to verify the performance improvement of the classifier and the consequent reduction of the dimensionality of the data after the selection of attributes.