Feature Disentanglement and Fusion Model for Multi-source Domain Adaptation with Domain-Specific Features

Chengrong Yang, Qiwen Jin, Xiaoguo Zhang, Yujue Zhou

Abstract: Multi-source domain adaptation aims to transfer knowledge from multiple source domains with diverse data distributions to a target domain. However, most existing methods overlook the unique characteristics of the target domain, which reduces the discriminability of target domain features. In this paper, we propose the Feature Disentanglement and Fusion Model (FDFM) to effectively leverage target-domain-specific features. Specifically, FDFM consists of two key components: a Feature Disentanglement block (FD) and a Feature Fusion block (FF). The FD block disentangles features into domain-invariant components and target-domain-specific components, while the FF block integrates these components to generate more discriminative feature representations for the target domain. Additionally, we design two classifiers: one trained on domain-invariant features and another on fused features with pseudo-labels. The final predictions are obtained by combining the outputs of both classifiers. Extensive experiments on four popular transfer learning benchmark datasets show that FDFM surpasses other state-of-the-art methods. For example, FDFM improves average accuracy from 74.4% to 76.6% on the OFFICE-HOME dataset.