In the documentation of sklearn, it says that several algorithms inherentrly support multilabel classification, such as RandomForest or MLP : https://scikit-learn.org/stable/modules/multiclass.html
Does it come from the implementation or the structure of the models? Moreover, how does it works for these algorithm? Is is a One-vs-the-Rest strategy or something else?
Thanks.
If the algorithm inherently supports multi-label classification, then it's usually an implicit feature of the algorithm rather than an implementation detail.
For example, MLPs inherently support multi-label classification because the output layer has a perceptron for each class, and each of these perceptrons output a probability for that class. The vector of outputs will predict an example's membership among all labels. Similarly, the the leaves of a Random Forest (or any other tree-based algorithm) can contain arbitrary-length vectors that describe the probability of the example belonging to each label.
The one-vs-rest strategy is used generalize binary classifiers (e.g. logistic regression) to multinomial problems. Multinomial problems are distinct from multi-label problems.
Edit:
Clarified my answer w.r.t. the differences between multinomial and multi-label problems.
