Classifier Adaptation Based on Modified Label Propagation for Unsupervised Domain Adaptation

Abstract

Unsupervised domain adaptation endeavors to learn a desirable classifier for a target domain by transferring knowledge learned from a related (source) domain. Existing approaches focus on deriving domain-invariant feature representations by aligning the domain distributions. However, those works often require an extra classifier. In contrast, this paper proposes a classifier adaptation method based on modified label propagation (CAMLP) for unsupervised domain adaptation. Inspired by pseudolabeling, CAMLP proposes a simple but effective measurement for relationships among cross-domain samples. Thus, samples from distinct domains are constructed in a same graph. The true labels can then propagate from the source domain to the target one along the graph. We also propose a consistency-aware pseudolabel annotation to alleviate the problem of negative transfer caused by unreliable pseudo labels. Extensive experiments on several benchmark datasets confirm that the proposed method performs favorably against the state-of-the-art approaches.