For a dataset with multi-label judgment, e.g., coco dataset but where we only want to predict the most possible label. There're multiple ways:
train as multi-label learning and predict as a multi-class problem;
train as multi-class learning and predict in the same way;
train as multi-label learning and predict top $k$ possibilities and select the majority label.
It seems not much literature to support the related connection. It would be much appreciated to see if there's any related research or some study on a more general connection between multi-label learning and multi-class learning in both theory and practice.