Explicitly fusing plug-and-play guidance of source prototype into target subspace for domain adaptation

The commonly used maximum mean discrepancy (MMD) criterion has two main drawbacks when reducing cross-domain distribution gaps: firstly, it reduces the distribution discrepancy in a global manner, potentially ignoring local structural information between domains, and secondly, its performance heavily relies on the often-unstable pseudo-label refinement process. To solve these problems, we introduce two universal plug-and-play modules: dynamic prototype pursuit (DPP) regularization and bi-branch self-training (BST) mechanism. Firstly, DPP introduces a new inter-class perspective to stabilize MMD by assigning a source prototype to each target sample. This allows us to utilize inter-class data structure information for better alignment. Next, BST is a novel non-parametric pseudo-label refinement mechanism that updates pseudo labels of target data using a classifier trained on the same distribution as the target domain. This avoids the distribution gap issue, making BST more likely to generate accurate target pseudo labels. Importantly, DPP and BST are universal plug-and-play modules for shallow domain adaptation methods. To demonstrate this, experiments of 3 MMD-based models incorporated with DPP and BST are conducted on Office-Caltech, Reuters21578, and Berlin-Emovo-Tess datasets. Experimental results show that these models incorporated with DPP and BST generally achieve better results compared to not using DPP and BST in terms of multiple metrics including accuracy, F1-score, MCC, and false positive rates. Code of 3 different DA methods enhanced by the plug-and-play DPP and BST is available at: https://github.com/Evelhz/DPP-and-BST.