Multi-Label Classification
A classification task where any number of labels can apply to one example at once, handled with independent sigmoid outputs instead of one competing softmax.
Softmax forces every class's probability to be taken from the others, since they must sum to one, which makes it wrong for labels that genuinely co-occur — a support ticket that is both a bug report and a refund request. Independent sigmoids, one per label with its own threshold, let any number of labels fire at once.
Scoring it multiplies the usual macro-versus-micro averaging question by however many labels exist, plus a new axis: averaging per label or per example answers different questions about the same predictions.