Open koaning opened 4 years ago
Is this a feature that can be added to the current approach or is this a new approach? I like your idea, but you might not need it right now.
The current approach also corrects for something so with that in mind it can still be added, only to be improved later.
So I was thinking about starting this, but then realized that the way I currently had the implementation might not make that much sense. Currently I pick thresholds for each group to get to a target true positive rate, but it might make more sense to calibrate the probabilities for each subgroup over their entire range.
What's your take @koaning?