Seleção de características de dados utilizando redes neurais artificiais

Abstract

Artificial Neural Networks have been disseminated over the years and its use has grown because of the good results in solving many real world problems. But the high degree of complexity present in Artificial Neural Networks makes its computation difficult, one of the factors that affect the performance of the neural network is the presence of redundant input variables that add nothing to their learning process thus making training more difficult and time consuming. The feature selection methods aim to determine which variables (characteristics) of the input that are most relevant or most contribute to the determination of output or response of the network, this enables the reduction of the number of entries to be excluded from the network variables less significant. In this work we implemented five methods of feature selection, the method of Garson, perturb, DBP, Sensitivity Analysis and Correlation. They chose three problems (Iris, CPU Performance, Strength of Concrete) for training the networks trained with the backpropagation algorithm following methods were performed yielding the importance of each entry, the entries were excluded and less important networks retrained obtaining A new mean square error which was compared to the original in order to evaluate the performance of the method. For the problem considered simpler Iris all methods yielded similar results. But for the most complex problems with the presence of more variables such as CPU Performance and Strength of Concrete methods perturb and Correlation showed the worst results, the method of Garson obtained a satisfactory result and Pad and Sensitivity Analysis showed better results and highlighted in relation to others.