A Fast Least Squares Support Vector Machine Classifier

Least Squares Support Vector Machines (LS-SVM) acquire the optimal solution by solving a set of linear equations, instead of solving a convex quadratic programming problem. But the solutions in a lose sparsity property. When the number of training sample points is bigger the cost of computation becomes great. The paper presents a new algorithm of Fast Least Squares Support Vector Machines (FLS-SVM). As the same generalization ability, especially when the number of training sample points is bigger the training speed of the new algorithm is faster than that of original LS-SVM algorithm. The new algorithm first selects the samples as reduced training set which have bigger support value from total training set. Then it trains LS-SVM to acquire optimal solution by using the selected samples in reduced training set. The results of experiment verify that the new algorithm not only acquires the same generalization ability with that of the original algorithms, but also is faster than that of the original algorithms.