Joint Clustering and Discriminative Feature Alignment for Unsupervised Domain Adaptation

Unsupervised Domain Adaptation (UDA) aims to learn a classifier for the unlabeled target domain by leveraging knowledge from labeled source with different but related distribution. Many existing approaches typically domain-invariant representation space directly matching marginal distributions of two domains. However, they ignore exploring underlying discriminative features data and align cross-domain features, which may lead suboptimal performance. To tackle these issues simultaneously, this paper presents Joint Clustering Discriminative Feature Alignment (JCDFA) approach UDA, is capable naturally unifying mining alignment class-discriminative into one single framework. Specifically, in order mine intrinsic information data, JCDFA jointly learns shared encoding tasks: supervised classification clustering where can guide learning locate object category. We then conduct feature separately optimizing new metrics: 1) an extended contrastive learning, i.e., semi-supervised 2) Maximum Mean Discrepancy (MMD), conditional MMD, explicitly minimizing intra-class dispersion maximizing inter-class compactness. When procedures, are integrated framework, tend benefit each other enhance final performance cooperative perspective. Experiments conducted on four real-world benchmarks (e.g., Office-31, ImageCLEF-DA, Office-Home VisDA-C). All results demonstrate that our obtain remarkable margins over state-of-the-art adaptation methods. Comprehensive ablation studies also verify importance key component proposed algorithm effectiveness combining strategies