Constructive Preference Elicitation by Setwise Max-margin Learning