In a multilabel setting a training example could be a, b, (a, b), d, c, (d, c), etc. This makes it a bit hard to come up with a helpful confusion matrix because the number of columns or rows could be very large - as I understand it, you wouldn't have a column or row for just a or b, but also (a, b).
What do people usually do in these cases? Do they usually create a column or row for every possible combination, or do they simplify somehow?
In case you'd like to provide code I am using Weka, but my question is primarily about the best practice.
