Abstract
Model selection for support vector machines (SVMs) involves tuning SVM hyperparameters,
such as C, which controls the amount of overlap, and the kernel parameters. Several criteria
developed for doing so do not take C into account. In this paper, we propose a unified framework for SVM model selection which makes it possible to include C in the definition of the
kernel parameters. This makes tuning hyperparameters for SVMs equivalent to choosing the
best kernel parameter values. We tested this approach using empirical error and radius margin
criteria. Our experiments on the Challenge Benchmarks dataset show promising results which
confirm the usefulness of our method.
Keywords: Model Selection, SVM, Support vector machine, hyperparameter, kernel.