Joint Contrastive-Adversarial Representation Learning for Unsupervised Domain Adaptation

Yuling Jiao; Lican Kang; Jerry Zhijian Yang; Zhou Yu; Fusheng Zhou

stat.MLOct 11, 2026

Joint Contrastive-Adversarial Representation Learning for Unsupervised Domain Adaptation

Authors: Yuling Jiao, Lican Kang, Jerry Zhijian Yang, Zhou Yu, Fusheng Zhou

Abstract

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.

Figures & tables

Methods	A $\to$ W	D $\to$ W	W $\to$ D	A $\to$ D	D $\to$ A	W $\to$ A
ERM	10.1	28.9	31.3	9.2	6.1	6.0
CDAN	12.1	29.1	28.3	7.8	5.9	4.7
DANN	16.4	63.4	67.5	10.2	7.2	9.6
Ours (ResNet)	16.7	44.6	51.4	12.1	10.5	13.0
ViT-B	14.8	37.5	47.8	12.5	5.9	9.9
Ours (ViT)	17.9	39.2	42.9	11.0	7.1	11.2

Table 1: Top-1 accuracy (%) of the proposed method and baselines trained from scratch on Office-31.

Methods	Ar $\to$ Rw	Pr $\to$ Rw	Cl $\to$ Pr	Ar $\to$ Cl	Pr $\to$ Ar	Cl $\to$ Rw
ERM	10.0	16.4	10.9	4.4	7.1	9.5
CDAN	9.4	16.0	12.5	3.9	5.0	8.4
DANN	9.6	16.4	10.3	4.9	4.4	10.2
Ours (ResNet)	7.6	10.4	10.0	4.3	5.0	7.2
ViT-B	9.2	14.3	3.1	5.8	7.0	9.1
Ours (ViT)	9.4	14.5	8.2	4.5	7.8	10.5

Table 2: Top-1 accuracy (%) of the proposed method and baselines trained from scratch on Office-Home.

Panel of Figure 2 — Figure 2: Two-dimensional t-SNE projections for the Office-31 W $\to$ D task. Source and target samples are shown in green and purple, respectively. Compared with DAN ( 2(b) ), ERM ( 2(c) ), and ViT ( 2(d) ), JCAR ( 2(a) ) exhibits substantial cross-domain overlap while maintaining a broadly distributed feature geometry.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

\hat{f}\in\arg\min_{f\in\mathcal{F}}\max_{h\in\mathcal{H}}\hat{\mathcal{R}}_{ad}(f,h).

Appendix

Algorithm 1 Domain adaptation with common representation

Data	Domain	Class	Total sample
Office-31	3	31	4,625
Office-Home	4	65	15,588
DomainNet	6	345	586,753

Appendix

Table 3: The dataset information for domain adaptation experiments.

Figure 3: The framework for representation learning.

Data	Epoch	Batch Size	Learning rate	Feature dimension
Office-31	800	16(8)	0.01, 0.1, 0.1(1.0)	256
Office-Home	800	16(8)	0.01, 0.1, 0.1(1.0)	512
DomainNet	800	16(8)	0.01, 0.1, 0.1(1.0)	512

Appendix

Table 4: Hyperparameter settings

Methods	A $\to$ W	D $\to$ W	W $\to$ D	A $\to$ D	D $\to$ A	W $\to$ A
ERM	74.8	95.0	93.9	69.7	43.2	55.4
CDAN	61.6	95.0	97.0	66.7	53.3	52.8
DANN	71.7	95.0	97.0	70.7	52.8	53.5
ViT-B	95.0	99.2	99.9	93.9	81.8	81.4
Ours (ViT)	92.0	98.9	99.5	82.0	79.1	79.8

Appendix

Table 5: The Top-1 accuracy (%) comparison of proposed method with others on the Office-31 dataset using pretrained model weights.

Methods	{D,W} $\to$ A	{A, D } $\to$ W	{A, W} $\to$ D
ERM	56.5	91.8	93.9
CDAN	56.7	92.5	89.9
DANN	56.5	81.1	90.9
ViT-B	81.4	99.2	100.0
Ours (ViT)	78.5	98.4	98.9

Appendix

Table 6: The Top-1 accuracy (%) comparison of proposed method with others on the Office-31 dataset using pretrained model weights.

Methods	Ar $\to$ Rw	Pr $\to$ Rw	Cl $\to$ Pr	Ar $\to$ Cl	Pr $\to$ Ar	Cl $\to$ Rw
ERM	66.5	63.5	56.6	39.5	41.2	55.3
CDAN	64.3	65.4	54.3	36.4	41.6	50.2
DANN	65.0	67.3	57.4	39.1	41.2	51.7
ViT-B	86.8	87.8	82.8	62.3	78.2	85.6
Ours (ViT)	84.5	82.9	76.2	58.5	67.0	78.9

Appendix

Table 7: The Top-1 accuracy (%) comparison of proposed method with others on the Office-Home dataset using pretrained model weights.

Methods	{Cl,Pr, Rw}	{Ar, Rw, Pr }	{Ar, Cl, Pr }	{Ar, Cl, Rw}
	$\downarrow$	$\downarrow$	$\downarrow$	$\downarrow$
	Ar	Cl	Rw	Pr
ERM	63.3	46.8	77.7	76.1
CDAN	67.4	51.5	77.2	73.8
DANN	62.7	48.3	75.5	73.3
ViT-B	83.7	64.4	90.0	89.6

Appendix

Table 8: The Top-1 accuracy (%) comparison of proposed method with others on the Office-Home dataset using pretrained model weights.

Methods	r $\to$ c	r $\to$ p	p $\to$ c	c $\to$ s	s $\to$ p	r $\to$ s
ERM	36.5	40.7	29.4	32.1	27.5	26.2
CDAN	51.4	50.1	41.1	40.1	41.0	40.7
DANN	48.9	48.8	35.9	41.1	40.2	39.8
ViT-B	47.4	53.5	37.8	44.7	47.1	34.8
Ours (ViT)	44.2	47.8	30.9	31.2	33.7	34.7

Appendix

Table 9: The Top-1 accuracy (%) comparison of proposed method with others on the DomainNet dataset using pretrained model weights.

Panel of Figure 4 — Figure 4: The t-SNE plots for the representation of the source and target domains in Office-31, the top row is our method, and other rows are the baselines. The first two columns correspond to single-source domain scenarios, and the remaining two correspond to multi-source domains.

Panel of Figure 5 — Figure 5: The t-SNE plots for the representation of the source and target domains in Office-Home, the top row is our method, and other rows are the baselines. The first two columns correspond to single-source domain scenarios, and the remaining two correspond to multi-source domains.

Unsupervised Domain Adaptation

Unsupervised domain adaptation transfers models from labeled source data to unlabeled target domains with different data distributions.

See all 183 papers

Joint Distribution Alignment for Universal Domain Adaptation
Aug 25, 2026Shizhe Li, Hongshan Pu, Mengying Xie +2Domain Adaptation Distribution Matching
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
A Step Towards Robust Unsupervised Domain Adaptation via Fine-Tuning and Reinforcement Learning
Jul 3, 2026Sushant Dagaji Desale, Rahul Mishra, Ashutosh Kumar SinhaAdversarial Training Unsupervised Domain Adaptation
Indian Institute of Technology Patna
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

Aug 25, 2026J/K move · Enter open · S save

Joint Distribution Alignment for Universal Domain Adaptation

Shizhe Li, Hongshan Pu, Mengying Xie, Yi Xiang+1

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

Domain Adaptation Distribution Matching

Open paper AskPDF arXiv

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.

Abstract

Unsupervised Domain Adaptation

Explore similar work

Joint Distribution Alignment for Universal Domain Adaptation

A Step Towards Robust Unsupervised Domain Adaptation via Fine-Tuning and Reinforcement Learning

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