Reweighting data in penalized optimization models: an approach to maximize subgroup fairness