关键词: 支持向量机, 线性支持向量机, 核方法, 近似高斯核, 再生核希尔伯特空间

Abstract: Training support vector machine (SVM) with nonlinear kernel functions on large-scale data is usually very time consuming. In contrast, there exist faster solvers to train the linear SVM. To utilize the computational efficiency of linear SVM without sacrificing the accuracy of nonlinear ones, in this paper, we present a method for solving large-scale nonlinear SVM based on an explicit description of an approximate Gaussian kernel. We first give the definition of the approximate Gaussian kernel, and establish the connection between approximate Gaussian kernel and Gaussian kernel, and also derive the error bound between these two kernel functions. Then, we present an explicit description of the reproducing kernel Hilbert space (RKHS) induced by the approximate Gaussian kernel. Thus, we can exactly depict the structure of the solutions of SVM, which can enhance the interpretability of the model and make us more deeply understand this method. Finally, we explicitly construct the feature mapping induced by the approximate Gaussian kernel, and use the mapped data as input of linear SVM. In this way, we can utilize existing efficient linear SVM to solve non-linear SVM on large-scale data. Experimental results show that the proposed method is efficient, and can achieve comparable classification accuracy to a normal nonlinear SVM.

Key words: support vector machine, linear support vector machine, kernel methods, approximate Gaussian kernel, reproducing kernel Hilbert space