Multi-label classification: do Hamming loss and subset accuracy really conflict with each other?