I have the following problem: Given two classes A and B which should be predicted using Sklearn or XGBoost. The function which has to be maximised is f(x) = A'^2/(A'+B').
Meaning that A' is the number of passing As from the dataset and B' the number of passing Bs. So the cost for removing a A is higher than removing a B. (Maximised should be TruePositive^2/(TruePositive+FalsePositive)
Defining a custom evaluation metric is not the problem, my problem is that I do not know which objective function/loss function (sorry, not sure what is the correct label is) I should use so that the model explicitly trains to maximise TP^2/(TP+FP) and not train sth. different.
Thanks in advance!