Margin-Based Feature Selection in Incomplete Data