Heterogeneous domain adaptation with statistical distribution alignment and progressive pseudo label selection

Heterogeneous domain adaptation (HDA) aims to adapt a trained model on a source domain with different input feature space to an unlabeled target domain. In fact, HDA is a challenging issue, since there exists feature and distribution discrepancies across domains. In this paper, we propose a novel approach named as heterogeneous domain adaptation with statistical distribution alignment and progressive pseudo label selection (SDA-PPLS). SDA-PPLS learns two projection matrices for source and target domains to map them into a latent subspace to have a shared feature space. Moreover, to mitigate the distribution gap, SDA-PPLS aligns both first-order and second-order statistical information, simultaneously, to improve the target classification model performance. In addition, to discriminate instances into distinct classes, SDA-PPLS aligns the class conditional distributions by pseudo label refinement of target domain data. Finally, to prevent the propagation of inaccurate pseudo labels to the next iteration, a progressive technique is proposed to select instances with higher probability. Experimental results on several real-word datasets on image to image, text to text and text to image tasks with different feature representations, demonstrate that the proposed method outperforms other state-of-the-art HDA methods.

Automated summary

The paper introduces a novel method called SDA-PPLS for heterogeneous domain adaptation, which aligns statistical distributions and progressively selects pseudo labels to map source and target domains into a shared feature space effectively.