Modelagem de variáveis qualitativas por meio de redes neurais artificiais: avaliação do uso de análise de correspondência como técnica de codificação

Abstract

In many problems, a qualitative description of the data is needed (using for example variables such as age or sex of a patient). For use on artificial neural networks, however, these variables must be recoded quantitatively. In this study, simulations were made with six wellknown techniques for recoding qualitative variables: Dummy 1-of-c, Dummy 1-of-(c-1), Thermometer, Numerical, Gray, and Binary. The performance of these techniques was compared to the performance obtained using the factors of Correspondence Analysis (CA) instead of the original qualitative variables. The use of these factors as inputs to the neural network has not been reported in the literature. Simulations were made with three dataset. Two of them involve classification problems, with two classes (performance was measured by the percentage of correct classifications); the third dataset involves a problem of function approximation (performance was measured by MAPE and MSE). For two of the datasets, the results for CA are equivalent to those of the other techniques applied; for the third, the performance was not satisfactory, do not showing any advantages over other techniques.