An efficient computational model for LS-SVM and its applications in time series prediction

Least Squares Support Vector Machine (LS-SVM) is a classic algorithm for regression estimation and classification. But unfortunately, for really large problems, LS-SVM can become highly memory and time consuming. In this paper, we present a simplified algorithm for LS-SVM, called I
_LS-SVM
, which effectively reduces the algorithmic complexity. In order to improve the rate of convergence and overcome instability of numerical value, a preconditioning conjugate gradient method is applied for solving the reduced system of linear equations. To evaluate the effectiveness of I
_LS-SVM
, several experiments for time series prediction are conducted. Compared with the standard LS-SVM, the proposed method is more effective for large training data set.