In domain adaptation, maximum mean discrepancy (MMD) has been widely adoptedas a discrepancy metric between the distributions of source and target domains.However, existing MMD-based domain adaptation methods generally ignore thechanges of class prior distributions, i.e., class weight bias across domains.This remains an open problem but ubiquitous for domain adaptation, which can becaused by changes in sample selection criteria and application scenarios. Weshow that MMD cannot account for class weight bias and results in degradeddomain adaptation performance. To address this issue, a weighted MMD model isproposed in this paper. Specifically, we introduce class-specific auxiliaryweights into the original MMD for exploiting the class prior probability onsource and target domains, whose challenge lies in the fact that the classlabel in target domain is unavailable. To account for it, our proposed weightedMMD model is defined by introducing an auxiliary weight for each class in thesource domain, and a classification EM algorithm is suggested by alternatingbetween assigning the pseudo-labels, estimating auxiliary weights and updatingmodel parameters. Extensive experiments demonstrate the superiority of ourweighted MMD over conventional MMD for domain adaptation.
展开▼
