Support Vector Machines are models widely used in supervised classification. The classical model minimizes a compromise between the structural risk and the empirical risk. In this paper, we consider the Support Vector Machine with feature selection and we design and implement a bi-objective evolutionary algorithm for approximating the Pareto optimal frontier of the two objectives. The metaheuristic is based on the nondominated sorting genetic algorithm and includes problem-specific knowledge. To demonstrate the efficiency of the algorithm proposed, we have carried out extensive computational experiments comparing the Paretofrontiers given by the exact method AUGMECON2 and the metaheuristic approach respectively in a set of well known instances. In this paper, we also discuss some properties of the points in the Pareto frontier.