On the Complexity of Discrete Feature Selection for Optimal Classification