Training Least-Square SVM by a Recurrent Neural Network Based on Fuzzy c-mean Approach

摘要

An algorithm to solve the least square support vector machine (LSSVM) is presented. The underlying optimization problem for LSSVM follows a system of linear equations. The proposed algorithm incorporates a fuzzy c-mean (FCM) clustering approach and the application of a recurrent neural network (RNN) to solve the system of linear equations. First, a reduced training set is obtained by the FCM clustering approach and used to train LSSVM. Then a gradient system with discontinuous righthand side, interpreted as an RNN, is designed by using the corresponding system of linear equations. The fusion of FCM clustering approach and RNN overcomes the loss of spareness of LSSVM. The efficiency of the algorithm is empirically shown on a benchmark data set generated from the University of California at Irvine (UCI) machine learning database.