An original idea for incorporating implicit decision-maker preferences in multi-objective evolutionary optimization is presented. The preferences are captured in a training set formed by solutions assigned to preference-ordered categories. New solutions are classified by a multi-criteria sorting method using the training set. This approach creates a selection pressure towards the region of interest of the decision-maker instead of the whole Pareto front. Test examples confirm that our method achieves a good characterization of the region of interest for the public project portfolio problem.