Joint distribution alignment on Lie group manifolds for domain adaptation

Abstract

Domain adaptation (DA) aims to learn a discriminative domain-invariant classifier by aligning the disparity between the source and target domains. Existing DA methods usually face one or several of the following problems: (1) using unreliable pseudo-labels of the target data directly learnt in the source domain; (2) aligning global distributions while neglecting local geometric structures of data; and (3) providing suboptimal local distribution alignments. Targeting these problems, this paper proposes a DA method called joint distribution alignment on Lie group manifold (JDALG). JDALG first extracts transformation-invariant Lie group features which embed intrinsic geometric structures of data. It then iteratively learns soft class labels for target data and aligns marginal and conditional distributions in a shared manifold subspace, minimizing global disparities while preserving local source domain consistency. This process culminates in a robust domain-invariant classifier. Extensive experiments on 6 DA datasets demonstrated that JDALG ranks the highest in 28 out of 54 transfer tasks, achieves average accuracy improvements of 7.6 %, 4.6 %, 0.7 %, 0.8 %, and 1.8 % over the best comparing methods on 5 datasets, and ranks the second on 1 dataset. The experiment results and comprehensive evaluations validate the effectiveness of JDALG compared to state-of-the-art DA methods.