I have a multi-label classification problem-- millions of records that potentially could hold more than one label. I'm running into issues related to lack of research/examples online, and am unable to create a strong model. Would it be advised against to build a multi-class classification model, and take any label with a predicted probability over 0 to be considered a potential label? Even with problem transformations, like binary relevance and classifier chains, the accuracy is terribly low (~12%) as opposed to using Naive Bayes and getting ~57% accuracy.
Perhaps I understood the question wrong, but multi-label and multi-class problems are fundamentally different, as the multi-class is exclusive (no instance can be classified into two classes a the same time), while the multi-label relaxes this condition. As Wikipedia on multi-label states:
in the multi-label problem there is no constraint on how many of the classes the instance can be assigned to.
So, using multi-class or multi-label depends on the problem itself and not on how you solve it.