Improved Sparse Least Square Support Vector Machines for the Function Estimation

Least square support vector machines (LS-SVMs) are deemed good methods for classification and function estimation. Compared with the standard support vector machines (SVMs), a drawback is that the sparseness is lost in the LS-SVMs. The sparseness is imposed by omitting the less important data in the training process, and retraining the remaining data. Iterative retraining requires more intensive computations than training the non-sparse LS-SVMs. In this paper, we will describe a new pruning algorithm for LS-SVMs: the width of ε-insensitive zone is introduced in the process of the training; in addition, the amount of the pruning points is adjusted according to the performance of training, not specified using the fixed percentage of training data; furthermore, the cross training is applied in the training. The performance of improved LS-SVMs pruning algorithm, in terms of computational cost and regress accuracy, is shown by several experiments based on the same data sets of chaotic time series.