Empirical Study of Support Vector Machine Kernels with Applications to Microarray Data

Support vector machines(SVMs) have demonstrated good performance to correctly classify samples into appropriate classes which contain tens of thousands of genes. The key to the success of using SVMs is choosing an appropriate kernel. Widely used kernels are linear, polynomial, radial basis function and sigmoidal. We compared the performance of kernels when all genes were used and when fewer numbers of good genes were used. Good genes were selected using a nearest shrunken centroid method. We report that linear kernel performance is as good as non linear kernels or slightly better in most cases and scaling data values to small numbers between 0 and 1 didn’t improve SVM performances.