Data mining for constructing ellipsoidal fuzzy classifier with various input features using GRBF neural networks

This paper aims at developing a theoretical framework for constructing ellipsoidal fuzzy classifiers with various input features from a data mining viewpoint. The proposed methodology for constructing fuzzy classification systems with ellipsoidal regions contains four parts: 1) rule-set initialization using a fully connected RBF neural network with an APC-III learning algorithm and cross entropy criterion; 2) feature selection by using a simple and practical algorithm; 3) determination of rule-set structure and representation using a generalized RBF neural network, where a fuzzy plus operator is employed as the activation function of the neurons at the output layer; and 4) a regularization cost function addressing the trade-off between misclassification, recognition and generalization for optimizing the initial rule-set.