Pruning methods to MLP neural networks considering proportional apparent error rate for classification problems with unbalanced data

• It is proposed the APERP metric, which is suitable to deals with imbalanced classes.
• The proposed metric is applied in conjunction to pruning of MLP neural networks.
• Three new pruning methods are presented to do pruning of MLP hidden neurons.
• The new CHI and KAPPA pruning methods had good results on E. coli unbalanced problem.
• The new pruning methods are easier to implement computationally than other methods.

This article deals with classification problems involving unequal probabilities in each class and discusses metrics to systems that use multilayer perceptrons neural networks (MLP) for the task of classifying new patterns. In addition we propose three new pruning methods that were compared to other seven existing methods in the literature for MLP networks. All pruning algorithms presented in this paper have been modified by the authors to do pruning of neurons, in order to produce fully connected MLP networks but being small in its intermediary layer. Experiments were carried out involving the E. coli unbalanced classification problem and ten pruning methods. The proposed methods had obtained good results, actually, better results than another pruning methods previously defined at the MLP neural network area.