Towards Discriminability with Distribution Discrepancy Constrains for Multisource Domain Adaptation

Abstract

Multisource domain adaptation (MDA) is committed to mining and extracting data concerning target tasks from several source domains. Many recent studies have focused on extracting domain-invariant features to eliminate domain distribution differences. However, there are three aspects that require further consideration. (1) Efforts should be made to ensure the maximum correlation in the potential subspace between the source and target domains. (2) While aligning the marginal distribution, the conditional distribution must also be considered. (3) Merely aligning the source distribution and target distribution cannot guarantee sufficient differentiation for classification tasks. To address these problems, we propose a novel approach named towards discriminability with distribution discrepancy constrains for multisource domain adaptation (TD-DDC). Specifically, TD-DDC first mines features of maximal relations learned from all domains while constructing domain data distribution mean distance metrics for interdomain distribution adaptation. Simultaneously, we integrate discriminability into domain alignment, which means increasing the distance among labels that are distinct from one another while reducing the distance among labels that are the same. Our proposed method not only reduces the interdomain distributional differences but also takes into account the preservation of interdomain correlation and inter-category discrimination. Numerous experiments have shown that TD-DDC performs much better than its competitors on three visual benchmark test databases.

Keywords

• multisource domain adaptation
• correlation learning
• class discriminative feature learning