Multiclass vs multilabel vs multioutput
These three concepts are related and sometimes easy to confuse.
A multiclass problem is a problem where each instance receives a single label, but where the label consists of >2 classes. E.g. when we analyze images of animals and we can classify each image as one of $n$ animals.
A multioutput problem is one where for each instance we have multiple output class labels, where each class label has a fixed set of available classes. That is, each instance is labelled per output with one of the multiple available classes for that output.
If we have a multioutput problem but each output is a binary label, then this is sometimes referred to as a special case called multilabel (e.g. in scikit-learn).
Multioutput problems often require decomposition of the problem into multiple single-output but multi-class problems. An obvious limitation of this approach is that in practice the multiple outputs may be correlated.
Some algorithms, such as decision trees, can be extended to multioutput problems without requiring problem decomposition.