Comment by dannykwells

Can someone explain to me - assuming they have enough data - why not train different models explicitly for each group / subgroup you want to model? You even could then just take the top N% of each group by score, effectively guaranteeing equal treatment for each group. Why would this not work?