I found this question but I need an answer to the other direction.
Example: Let's say we want to predict if a person with a certain profile wants to buy product A and/or B. So we have 2 binary classes A and B that don't exclude each other:
A B 0 1 1 1 0 0 1 0 ...
(We don't want to predict how likely it is for a person to buy B if the person has already bought A.) Does it - in general or under certain conditions - make sense to transform this problem into a single-class multi-label problem with 4 labels (4=#combinations of A and B)? What if the number of binary classes is larger than 2?