These PDA methods are different in the following four aspects, including whether toreweight theimportance ofsource datainthecross-entropyloss, themetric ofadopted distribution distance, the strategy to obtain the source data weights, and whether to utilize the conditionalentropyloss. If the strategy is based on the discriminatorD that is trainedtopredictthesource(resp. The full implementation details are provided in this section. Forthenoiselevelp [0,1],wesimulatetheweightof thei-th sample that belongs to source-only classes bypwi/ P jI(yj Ys\Yt)wj,forisuch that i Ys\Yt (i.e., thei-th source sample that belongs to source-only classes is assigned with noisy weight inproportional totheweight learned byouradversarial reweighting model).