Joint Contrastive-Adversarial Representation Learning for Unsupervised Domain Adaptation
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≥1, the bound identifies a sufficiently large unlabeled-sample regime in which the downstream rate is 𝒪̃(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.
Related
- Rethinking Representation Geometry in Supervised Contrastive Learning under Long-Tailed Data
- Prior-Informed Generalizable Low-Rank Representation Learning via Regularized Adversarial Reweighting
- Anchor Divergence for Semantic Geometry in Contrastive Learning
- Fisher Contrastive Learning to Mitigate Feature Suppression: A Case Study in Acute Promyelocytic Leukemia Diagnosis
- RobustLDS: Learning linear dynamical systems under adversarial corruptions
BibTeX
@misc{jiao2026joint,
title = {{Joint Contrastive-Adversarial Representation Learning for Unsupervised Domain Adaptation}},
author = {Yuling Jiao and Lican Kang and Jerry Zhijian Yang and Zhou Yu and Fusheng Zhou},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.15613},
eprint = {2610.15613},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.15613},
}
RIS
TY - UNPB
TI - Joint Contrastive-Adversarial Representation Learning for Unsupervised Domain Adaptation
AU - Yuling Jiao
AU - Lican Kang
AU - Jerry Zhijian Yang
AU - Zhou Yu
AU - Fusheng Zhou
PY - 2026
DA - 2026/10/11
PB - arXiv
DO - 10.48550/arxiv.2610.15613
UR - https://arxiv.org/abs/2610.15613
AB - 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≥1, the bound identifies a sufficiently large unlabeled-sample regime in which the downstream rate is 𝒪̃(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.
ER -