Is your feature request related to a problem? Please describe.
Kinda both. Currently, each sample can only be member of one class for the stratification, but if we want to balance multiple classes across the splits this is not possible.
Describe the solution you'd like
Open the stratification argument to take a dictionary mapping strings to a list of classes and adjust the problem formulation accordingly. From the theoretical perspective this may not even require a change.
Describe alternatives you've considered
A possible workaround may be to collate all class labels into one class and to treat this as a new class to be balanced, but this leads to a different problem, i.e., it has a different solution than the one original problem.
Is your feature request related to a problem? Please describe. Kinda both. Currently, each sample can only be member of one class for the stratification, but if we want to balance multiple classes across the splits this is not possible.
Describe the solution you'd like Open the stratification argument to take a dictionary mapping strings to a list of classes and adjust the problem formulation accordingly. From the theoretical perspective this may not even require a change.
Describe alternatives you've considered A possible workaround may be to collate all class labels into one class and to treat this as a new class to be balanced, but this leads to a different problem, i.e., it has a different solution than the one original problem.
Additional context N.A.