A stable hyperparameter selection for the Gaussian RBF kernel for discrimination

Abstract

Kernel-based classification methods, for example, support vector machines, map the data into a higher-dimensional space via a kernel function. In practice, choosing the value of hyperparameter in the kernel function is crucial in order to ensure good performance. We propose a method of selecting the hyperparameter in the Gaussian radial basis function (RBF) kernel by considering the geometry of the embedded feature space. This method is independent of the choice of the discrimination algorithm and also computationally efficient. Its classification performance is competitive with existing methods including cross-validation. Using simulated and real-data examples, we show that the proposed method is stable with respect to sampling variability. © 2010 Wiley Periodicals, Inc.