Feature selection to enhance a two-stage evolutionary algorithm in product unit neural networks for complex classification problems

This paper combines feature selection methods with a two-stage evolutionary classifier based on product unit neural networks. The enhanced methodology has been tried out with four filters using 18 data sets that report test error rates about 20 % or above with reference classifiers such as C4.5 or 1-NN. The proposal has also been evaluated in a liver-transplantation real-world problem with serious troubles in the data distribution and classifiers get low performance. The study includes an overall empirical comparison between the models obtained with and without feature selection using different kind of neural networks, like RBF, MLP and other state-of-the-art classifiers. Statistical tests show that our proposal significantly improves the test accuracy of the previous models. The reduction percentage in the number of inputs is, on average, above 55 %, thus a greater efficiency is achieved.