Correct implementation of binary relevance

Question

I have questions regarding Binary Relevance implementation.

How to treat multi-label problem with empty class? (I have instance that cannot have any label). Should I introduce new label "NoLabel" in this case?

Answer 1

In this case, binary relevance won't help. Since the decision planes are always going to split a all negative space. and especially when you add a "No Label" Label.. So the True Negative is going to be mixed with False Negative.

But there is an expensive way to solve this by using 2^{# Labels} -1 +1 multi-class classifier. 2^{# Labels} -1 is for direct mapping to binary relevance, and +1 is for the "No Label" Label.