Schema of the reverse selection method. In the first three steps a binary classifier is used to predict classification on all datasets, including the training one. If the probability of belonging to the noninteresting P-class is greater than a threshold τ the source is discarded before proceeding to the next step. By doing so all datasets decrease in size and, most importantly, they are rebalanced toward the interesting sources, namely the high-z QSOs. The last step is a simple multi-label classification, just like the direct selection method.