The notion of fairness you are discussing is 1) vague 2) very similar to standard training. If you have an optimization problem that represents your fairness, i

The notion of fairness you are discussing is 1) vague 2) very similar to standard training. If you have an optimization problem that represents your fairness, it’d be more useful to just compare that with the other problems in the literature. Otherwise, it sounds like you want max accuracy and max fairness, which is unlikely to occur on non-degenerate datasets. Anyway, if I’m understanding correctly, you would end up with “individual fairness” if you did formalize your fairness, which is discussed in the linked papers.