That is, if ??p is a set of the true positive labels of xt, then a successful learning system will tend to rank the labels in the ??p higher than any other labels not in ??p.To address the requirements mentioned above, we adopt the following procedure.