Learning from label proportions via an iterative weighting scheme and discriminant analysis