For a dataset with multi-label judgement, e.g., coco dataset but where we only want to predict the most-possible label. There're multiple ways, for example : 1) train as a multi-label learning(each label as a binary-task) and predict as a multi-class problem using sigmoid and predict the one with largest sigmoid score; 2) train as a multi-class learning (e.g., change the data) and predict in a multi-class way.
My question is for such problem, how to choose which methods to use ? And what is the standard handling for such problems ?
It seems not much literature to support the related connection. It would be much appreciated to see if there's reference to point to related research or some study on more general connection between multi-label learning and multi-class learning in both theory and practice.