Randomized and Dimension Reduced Radial Basis Features for Support Vector Machine
Hidaka Akinori, Kurita Takio
pp. 1-8
DOI:
10.5687/iscie.29.1抄録
Generally, the dimension of the Kernel matrices of the kernel Support Vector Machines (SVM) increases as the number of training samples increases. However high dimensional features often bring redundant computation and decline of the generalization ability. Also kernel functions have several hyper-parameters which are fixed to the same values for all training samples. By considering the kernel matrix of Radial Basis Function (RBF) as a new high dimensional nonlinear feature for SVM, there is no limitation of which those hyper-parameters have to be a single fixed number. In this paper, we develop a nonlinear feature extraction method based on the selection of kernel seeds and the fine tuning of the kernel parameters, called randomized and dimension Reduced Radial Basis Features (RRBF).