Refinement of the feedforward network in multi-class classification problems using a hybrid approach combining supervised clustering and a fuzzy classifier

Abstract

A FeedForward Neural Network (FFNN) with a single hidden layer (Single Feedforward Neural Networks, SFNN) represents a classical discriminative model able to estimate conditional probabilities (Bayesian probabilities) and provides an efficient alternative in multi-class classification problems. This paper presents an approach to improve the accuracy of SFNN in recognizing multiple patterns in static data focusing on classification problems which systematically pose difficulties in obtaining high levels of accuracy (low misclassification rates) using different classification methods. In an innovative way, the method comprises the joint application of supervised clustering and fuzzy classifier, starting from the probability distribution predicted by the SFNN. The proposed approach​ (Refinement based on Supervised Clustering and Fuzzy Classifier, RSCFC) is applied in real datasets widely used as a benchmark for multi-class classification. The results were compared with other methods involving different categories of neuro-fuzzy systems (cooperative and hybrid) and other networks with tailored topologies for the classification of static data. The RSCFC method proved to be viable, able to improve the results obtained with the SFNN using classical settings (e.g., softmax function, crossed entropy) and provided correct classification rates higher or similar to those achieved with existing methods. Additionally, the method is flexible as any classifier capable of estimating conditional probabilities for data classification can be adopted as a starting point (primary classifier) for the refinement and achievement of greater accuracy.