Inter-Class Distance Based Kernel Parameter Evaluating Method for RBF-SVM

Selecting the optimal parameters of the Support Vector Machines (SVM) is very important in practice. This paper detailedly analyzes the effects given by the Radial Basis Function (RBF) kernel parameter on the feature space, and proposes a novel kernel parameter evaluating method, which is based on the Inter-Class Mean Distance (ICMD). Theoretical and experimental analysis is made on the proposed method. The proposed method makes it possible to select the kernel parameter and the penalty parameter by two stages, which significantly decreases the time cost of the parameters selection. Experiments are made to compare the “two stage” method with the grid search method, results show that the former can select the optimal parameters with greatly shortened time cost.