Kernel variable selection for multicategory support vector machines

Variable selection is important in statistical learning because it can increase predictive performances and yield interpretable models. Since support vector machines construct a classification model through a mapping from an original space to a high-dimensional feature space, it is difficult to select informative variables and interpret the relation between covariates and class labels. In this paper, we suggest a variable selection method for support vector machines, focusing on the multicategory problem. We study asymptotic properties of the proposed method. Also we illustrate that our method can accurately select relevant variables and yield interpretable models on both simulated and real data sets.