Optimizing the classification cost using SVMs with a double hinge loss