{Yu Zhang Jinjing Zhu Pengxin Guo}

Abstract

Partial Domain Adaptation (PDA), which assumes that the label space of the target domain is a subset of that in the source domain, has attracted much attention in recent years. Due to the difference in the label space of these two domains, it is hard to directlyalign these two domains in PDA. To solve this problem, we propose a Selective Partial Domain Adaptation (SPDA) method, which selects useful data for the adaptation to the target domain. Specifically, we firstly design a Maximum of Cosine (MoC) similarityfunction customized for PDA to select useful data in the source domain to decrease the domain discrepancy. In the MoC similarity function, for each target sample, we select the source sample with the maximal cosine similarity for adaptation. Moreover, a selective training method is designed to add useful target data into the source domain. In detail, the selective training method firstly assigns pseudo-labels to target samples with the selftraining strategy and then adds target samples with high confidence in terms of pseudolabels to the source domain. Based on these two selection operations, the proposed SPDA method can select useful data for domain adaptation. Experiments on several datasetsdemonstrate the effectiveness of the proposed SPDA method.