Geometrical modification of wavelet SVM kernels and its application in microarray analysis

The selection and design of appropriate kernel functions play a key role in effective support vector machine (SVM) leaning. A general strategy is to customize the existent kernel functions to fit into the data property and structure. Wavelet kernels have been developed to approximate arbitrary nonlinear functions for signal processing. In this paper, we propose novel wavelet kernels based on the Riemannian geometrical structure theory, by constructing a hyperplane with better spatial resolution. This wavelet kernel SVM approach was applied to the yeast time course microarray dataset and outperformed the traditional Gaussian kernel and polynomial kernel.