Domain compensatory adversarial networks for partial domain adaptation

Recently, domain adaptation has stimulated interest in the community of machine learning since it can improve the performance of the model in a new domain (target domain) by borrowing knowledge from a labeled source domain. At the same time, the presence of large-scale labeled datasets also raised significant attention in this scenario: the class labels in the new domain are only a subset of those in the source domain. We propose an adversarial-net-based method in this paper, called domain compensatory adversarial network (DCAN). The critical difficulty of this problem is to reduce the negative impact of source instances with weak discriminability and filter out outlier source classes by exploiting the prior probability of classes. DCAN can identify source instances with weak discriminability and reverse its domain label to compensate for the target label space, which alleviates the negative effect of mismatching label space. Besides, DCAN reweights outlier source classes with the class prior distributions of source data for both category classifier and domain classifier to promote positive transfer. Experiments have revealed the superiority of DCAN compared to the existing methods.