Universal Semi-Supervised Domain Adaptation by Mitigating Common-Class Bias

TL;DR

This work introduces Universal Semi-Supervised Do-main Adaptation (UniSSDA), a practical yet challenging setting where the target domain is partially labeled, and the source and target label space may not strictly match.

Abstract

Domain adaptation is a critical task in machine learning that aims to improve model performance on a target domain by leveraging knowledge from a related source domain. In this work, we introduce Universal Semi-Supervised Domain Adaptation (UniSSDA), a practical yet challenging setting where the target domain is partially labeled, and the source and target label space may not strictly match. UniSSDA is at the intersection of Universal Domain Adaptation (UniDA) and Semi-Supervised Domain Adaptation (SSDA): the UniDA setting does not allow for fine-grained categorization of target private classes not represented in the source domain, while SSDA focuses on the restricted closed-set setting where source and target label spaces match exactly. Existing UniDA and SSDA methods are susceptible to common-class bias in UniSSDA settings, where models overfit to data distributions of classes common to both domains at the expense of private classes. We propose a new prior-guided pseudo-label refinement strategy to reduce the reinforcement of common-class bias due to pseudo-labeling, a common label propagation strategy in domain adaptation. We demonstrate the effectiveness of the proposed strategy on benchmark datasets Office-Home, DomainNet, and VisDA. The proposed strategy attains the best performance across UniSSDA adaptation settings and establishes a new baseline for UniSSDA.

Figures (7)

• Figure 1: Adaptation settings in Universal SSDA. Existing SSDA works conventionally focus on closed-set settings.
• Figure 2: (a) and (b) show the percentage of samples in target private classes ($\mathcal{Y}_T \cap \mathcal{Y}_S^c$) misclassified as common classes ($\mathcal{Y}_T \cap \mathcal{Y}_S$) under open-partial setting, demonstrating the effect of common-class bias on existing SSDA and UniDA methods. (a) is implemented on DomainNet-126 $C\rightarrow P$ with ResNet-34 backbone. (b) is implemented on DomainNet-345 $C\rightarrow P$ with frozen ViT foundation model encoder and learnable classifier.
• Figure 3: T-SNE visualization of common versus target private class samples shows negative transfer due to common-class bias. Incorrect mapping between common and target private classes can persist and be reinforced by naive pseudo-labeling. Example taken from DomainNet-126 $C\rightarrow P$ in open-partial setting, with visualization on 15 shared and 15 target private classes.
• Figure 4: The semi-supervised classifier trained with naive pseudo-labels is more vulnerable to common-class bias than the supervised classifier is. Using supervised classifier outputs as priors to refine the pseudo-labels significantly improves the performance of the resulting semi-supervised classifier. Example taken from DomainNet $C\rightarrow P$ in open-partial setting.
• Figure 5: Proposed Prior-Guided Pseudo-label Refinement (PGPR) strategy: During adaptation, we add a supervised classification head $\tilde{h}$ trained only on labeled samples to provide prior distribution estimates. We align the per-instance class group distribution output by $h \circ f$ to the estimate by $\tilde{h} \circ f$ and then aggregate the two classifier decisions. The class with the highest resulting probability is taken as the refined pseudo-label for training. The supervised classfier $\tilde{h}$ is discarded at the end of training.