Joint Distribution Alignment for Universal Domain Adaptation

Shizhe Li; Hongshan Pu; Mengying Xie; Yi Xiang; Xiaowei Yang

cs.LGAug 25, 2026

Joint Distribution Alignment for Universal Domain Adaptation

Authors: Shizhe Li, Hongshan Pu, Mengying Xie, Yi Xiang, Xiaowei Yang

Organizations: School of Software Engineering, South China University of Technology, Guangzhou 510006, China · School of Future Technology, South China University of Technology, Guangzhou 511442, China · College of Computer Science, Chongqing University, Chongqing 400044, China

Abstract

Unsupervised domain adaptation (UDA) has been widely concerned in the fields of machine learning, pattern recognition, and computer vision. Traditional UDA learning usually assumes that the label spaces of the source and target domains are exactly the same and only needs to solve the problem of sample distribution drift existing between two domains. However, in real world applications, the label spaces between two domains may be different. In this case, there are both sample distribution drift and class spatial difference between domains, namely Universal Domain Adaptation (UniDA) learning scenario. At present, existing works rarely offer theoretical analysis for universal domain adaptation. In this paper, we provide an upper bound of the generalization error for universal domain adaptation. According to the proposed generalization error bound, we propose a novel UniDA algorithm called Joint Distribution Alignment for Universal Domain Adaptation (JAUA), which aligns the joint distributions by minimizing the distribution discrepancy calculated by Chi-Square divergence. Furthermore, we propose a progressive pseudo-labeling method to assign the pseudo labels to unlabeled target samples. The experiment results on six public image datasets demonstrate the superiority of JAUA in handling the UniDA problem.

Figures & tables

Symbols	Descriptions
$\mathcal{X}\in\mathcal{R}^{D}$	Feature space
$C$	The number of common classes
$\mathcal{Y}^{u_{s}}=\{y^{u_{s}}\}=\{C+1\}$	Private label space
$\mathcal{Y}^{u_{t}}=\{y^{u_{t}}\}=\{C+2\}$	Unknown label space
$\mathcal{Y}^{u_{G}}=\{y_{1},y_{2},...,y_{c}\}$	Common label space
$\mathcal{Y}^{s}=\{y_{1},y_{2},...,y_{c},y^{u_{s}}\}$	Source label space

TABLE I: Symbols and notations

Fig. 3: The framework of JAUA. The input dataset is first classified into $1,2,\ldots,C+1$ classes using SVM. Through progressive pseudo-labeling, the dataset’s classification is progressively updated. Target domain samples with confidence $E$ below $\tau$ are classified as $C+2$ , and the label set is updated. This iterative training process is repeated until the final target domain classification results are obtained.

datasets	Domain	Sample number	Category
Office-31	Amazon	2,817	31
	Dslr	795	31
	Webcam	498	31
Office-Home	Art	2,427	65
	Clipart	4,365	65
	Product	4,439	65

TABLE II: Dataset description

Parameters	Office-31	Office-Home	Visda	Domainnet	ImageCLEF	PACS
$\lambda_{tt}$	1	1	1	1	1	1
$\lambda_{s}$	0.1	0.1	0.07	0.1	0.07	0.1
$\lambda_{t}$	0.1	0.1	0.07	0.1	0.07	0.1
$\lambda_{gss}$	0.1	0.01	0.4	0.3	0.4	0.3
$\lambda_{gst}$	0.3	0.01	0.2	0.4	0.2	0.3
$\lambda$	0.0001	0.0001	0.0001	0.0001	0.0001	0.0001

TABLE III: The parameter settings in JAUA

Methods	ACC	UNK	HOS
UAN	60.8	20.4	30.5
CMU	61.4	24.1	34.6
DANCE	69.2	31.0	42.8
DCC	64.2	32.3	43.0
OSBP	30.3	49.9	37.7
IWAN	58.7	18.0	27.6

TABLE IV: Comparison of the results on Visda dataset(Unit:%)

Methods	ACC	UNK	HOS
UAN	89.3	49.3	63.5
CMU	91.1	61.1	73.1
DANCE	93.0	78.2	84.7
DCC	93.2	70.7	80.2
IWAN	86.7	36.7	51.6
OVANet	84.2	88.0	85.9

TABLE V: Comparison of the results on Office-31 dataset(Unit:%)

Methods	ACC	UNK	HOS
UAN	77.0	45.2	56.6
CMU	77.2	52.0	61.6
DANCE	80.3	54.0	63.9
DCC	78.9	68.4	72.0
OSBP	64.6	34.5	44.5
IWAN	73.3	33.0	45.3

TABLE VI: Comparison of the results on Office-Home dataset(Unit:%)

Fig. 4: Influence of the parameters of JAUA for task: D $\rightarrow$ A on Office-31

Methods	HOS
UAN	41.0
CMU	48.3
DANCE	33.5
DCC	49.2
OSBP	32.0
IWAN	32.8

TABLE VII: Comparison of the results on Domainnet dataset(Unit:%)

Methods	HOS
OVANet	74.3
MLNET	83.3
E-MLNET	85.4
JAUA	89.4

TABLE VIII: Comparison of the results on ImageCLEF dataset(Unit:%)

Methods	ACC
UAN	30.7
CMU	21.0
DANCE	17.6
OVANet	58.5
GUDA	58.2
JAUA	64.2

TABLE IX: Comparison of the results on PACS dataset(Unit:%)

Fig. 5: The change of the whole loss of JAUA for task: D $\rightarrow$ A on Office-31

Fig. 6: Ablation experiment for task: D $\rightarrow$ A on Office-31

Domain Adaptation

Domain adaptation transfers machine-learning models across changing data distributions, often using unlabeled or limited target-domain examples.

See all 252 papers

Joint Contrastive-Adversarial Representation Learning for Unsupervised Domain Adaptation
Oct 11, 2026Yuling Jiao, Lican Kang, Jerry Zhijian Yang +2Unsupervised Domain Adaptation Representation Learning
A General Representation-Based Approach to Multi-Source Domain Adaptation
Apr 26, 2026Ignavier Ng, Yan Li, Zijian Li +3Multi-Source Domain Adaptation Unsupervised Domain Adaptation
Carnegie Mellon University · Mohamed bin Zayed University of Artificial Intelligence
Variance-reduced Domain Adaptation using Paired Sampling
Jul 22, 2026Andrea NapoliUnsupervised Domain Adaptation Maximum Mean Discrepancy
Department of Electronics and Computer Science University of Southampton, UK

Oct 11, 2026J/K move · Enter open · S save

Joint Contrastive-Adversarial Representation Learning for Unsupervised Domain Adaptation

Yuling Jiao, Lican Kang, Jerry Zhijian Yang, Zhou Yu+1

Unsupervised Domain Adaptation Representation Learning Cross-Domain Generalization Domain-Adversarial Learning Contrastive Learning

Open paper AskPDF arXiv

Effective unsupervised domain adaptation (UDA) requires a representation that captures transferable structure from both the labeled source domain and the unlabeled target domain. We introduce Joint Contrastive-Adversarial Representation Learning (JCAR), a two-stage framework that learns a common representation from the union of unlabeled source and target samples by combining contrastive learning with adversarial domain alignment, and then freezes the representation to fit a source-supervised predictor. We establish a non-asymptotic excess-risk bound for this procedure that couples representation and domain-discriminator estimation while separating downstream prediction, with explicit neural-network approximation errors. Under the stated assumptions and common Hölder smoothness

β\geq1

, the bound identifies a sufficiently large unlabeled-sample regime in which the downstream rate is

\widetilde{\mathcal{O}}(m^{-β/(2β+d)})

, where

m

counts labeled source samples and logarithmic factors are suppressed. The labeled-sample exponent is governed by the latent dimension

d

rather than the ambient dimension

p

, which enters the pretraining cost. Experiments on Office-31, Office-Home, and DomainNet demonstrate clear transfer gains across diverse domain shifts, with particularly strong performance when representations are learned from scratch.

Abstract

Domain Adaptation

Explore similar work

Joint Contrastive-Adversarial Representation Learning for Unsupervised Domain Adaptation

A General Representation-Based Approach to Multi-Source Domain Adaptation

Variance-reduced Domain Adaptation using Paired Sampling