A Genetic Algorithms-based Approach for Selecting the Most Relevant Input Variables in Classification Tasks

The paper deals with the design and development of classifiers and, in particular, with the problem of selecting the most relevant input variables to be used as inputs for classification purpose in practical applications. In many real problems the selection of input variables is a very important task: often real datasets used for developing a classifier contain a high number of inputs but no a priori knowledge is available on the considered application, which allow to distinguish which variables among them are actually relevant inputs for a classifier to develop. In this paper an automatic input selection method is proposed which exploits genetic algorithms: the classifier is implemented through a neural network, in particular a self organizing map, which is subjected to a supervised training procedure. This approach provides satisfactory results in terms of classification accuracy and it is also very useful because it selects the input variables that are interesting for the proposed task, by thus contributing to increase the knowledge and comprehension of the considered problem. In order to demonstrate the effectiveness of the proposed method, some tests have been performed by exploiting a dataset belonging to the UCI repository and the results are presented and discussed in the paper.