Develop discrete minimax classifiers to address imbalanced classes and prior probability shifts in multi-label classification, and resolve inconsistencies in conditional probability estimation for non-decomposable losses such as F-measure.