Least Squares Support Vector Machine Parameter Selection

Support Vector Machine (SVM) was initially presented by Vapnik et al. in the last decade of the20th century. SVM is based on statistical learning theory and excellent learning performance. SVM is widely used in many fields, such as pattern recognition, Image retrieval, and protein data analysis.The support vector machine is a kernel-based learning methods, it can handle non-linear samples, it will enter the space through the kernel function is mapped to feature space, making it conducive to problem solving linear performance. The least squares support vector machine is a variant of support vector machines and is also a kernel-based learning methods. The kernel function is the crucial ingredient of least squares support vector machine, which directly affect the performance of least squares support vector machine, kernel parameter is the main elements of the kernel function, so kernel parameters play an crucial role in improving the learning and generalization ability of model. If only have high performance kernel functions, but also lack of proper regularization parameter, will also affect the performance of least squares support vector machine, so very important nuclear parameters and the regularization parameter.In this paper, a novel algorithm of parameter selection is proposed based on distance measure for least squares support vector machine. This method first to determine the optimal kernel parameter by calculating the ratio between the spacing of the distance and class within the class of training samples in feature space, and then use the grid search method to determine the optimal regularization parameter. The experiment proved the feasibility of this method compared with the traditional grid search method greatly reduces training time, in the case of no loss of measurement accuracy.