unsupervised domain adaptation

This work addresses unsupervised domain adaptation (UDA) for image classification, where labeled source-domain data and unlabeled target-domain data are available. We systematically evaluate mainstream UDA methods on standard benchmarks—Office-31 and Office-Home—within a unified experimental framework. Our key contribution is the first comparative analysis of Transformer-based UDA algorithms (e.g., SSRT) under varying data scales and domain shifts, revealing their robustness boundaries and failure modes. Implementing adversarial training, feature alignment, self-training, and safe self-refinement (SSRT) in PyTorch, we enable reproducible large-scale ablation studies. Results show SSRT achieves 91.6% accuracy on Office-31; however, it suffers significant degradation under small-batch settings—dropping to 72.4% on Office-Home—highlighting a critical practical limitation. This empirical finding provides essential guidance for deploying UDA methods in real-world scenarios with constrained computational resources.

This work addresses the challenge of selecting appropriate source domains and pre-trained models for unsupervised domain adaptation when target-domain labels are unavailable—a critical yet underexplored problem that often limits adaptation performance. To tackle this, the authors propose the PAS (Pre-trained Adaptation Suitability) scoring mechanism, which evaluates source–target domain compatibility and model transferability by analyzing the geometry of pre-trained feature embedding spaces. Remarkably, PAS accurately predicts post-adaptation target accuracy without requiring any target labels. This approach enables, for the first time, joint unsupervised selection of both source domains and pre-trained models. Extensive experiments on multiple image classification benchmarks demonstrate a strong correlation between PAS scores and actual adaptation accuracy, leading to significantly improved performance while substantially reducing computational overhead.

Existing domain adaptation methods (e.g., DSN) separate domain-invariant representations (DIReps) from domain-specific representations (DDReps) via orthogonality constraints. However, weak orthogonality often leaves discriminative information in DDReps—causing “information leakage” that degrades target-domain generalization. This paper proposes a KL-divergence-driven DDRep minimization mechanism that explicitly compresses the DDRep distribution within deep networks, thereby enhancing the purity and transferability of DIReps. The constraint is compatible with pretrained models and enables end-to-end domain-separation learning. Evaluated on standard image benchmarks (e.g., Office), the method achieves or surpasses state-of-the-art performance. Synthetic-data experiments demonstrate robustness to initialization and substantial improvements in cross-domain classification accuracy. The core innovation lies in replacing conventional orthogonality constraints with a distribution-level KL-divergence penalty, fundamentally mitigating representation leakage.

This paper addresses unsupervised model adaptation from a source domain to a target domain without access to any source-domain data or labels—termed *source-free* adaptation—aiming solely to improve the generalization of a pre-trained source model on unlabeled target data. To this end, we propose a Collaborative Class-Conditional Generative Adversarial Network (CC-GAN) framework: it models the semantic structure of the target domain via class-conditional generation; enforces weight constraints derived from the source model to preserve its discriminative capability; and incorporates clustering-driven feature regularization to enhance the discriminability of target-domain representations. Evaluated across multiple cross-domain vision tasks, our method achieves significant performance gains over conventional source-dependent adaptation approaches—using only unlabeled target data. It is the first to empirically validate the effectiveness, robustness, and scalability of model adaptation under the source-free setting.

Unsupervised domain adaptation (UDA) under label distribution shift suffers from performance degradation, as conventional methods rely on the covariate shift assumption and fail to guarantee discriminative domain-invariant features. Method: This paper proposes Conditional Adversarial Support Alignment (CASUAL), a theoretically grounded framework that jointly optimizes support alignment of conditional feature distributions and classification discriminability via conditional adversarial training. Contribution/Results: CASUAL introduces the first theoretical analysis of conditional support alignment and derives a tighter upper bound on the target risk—rigorously proving its superiority over classical marginal alignment. Extensive experiments across multiple UDA benchmarks with label shift demonstrate that CASUAL consistently outperforms state-of-the-art methods, validating both the effectiveness and generalizability of theory-driven alignment strategies.

Graph domain adaptation faces complex, multi-dimensional distribution shifts, and existing methods rely on handcrafted alignment criteria and graph filters, limiting their generalization. This work proposes ADAlign, a framework that leverages Neural Spectral Discrepancy (NSD) to uniformly capture arbitrary-order feature-structure dependencies in the spectral domain. ADAlign introduces a learnable frequency sampler that, through minimax optimization, adaptively focuses on critical spectral components, enabling scene-aware alignment without manual specification. The method establishes an end-to-end adaptive distribution alignment mechanism, significantly outperforming state-of-the-art approaches across 10 datasets and 16 transfer tasks, while simultaneously reducing memory consumption and improving training efficiency.

This work addresses the challenge of lacking identifiability guarantees in unsupervised multi-source domain adaptation with high-dimensional data by proposing a general domain adaptation framework grounded in Markov blanket structure. The method learns compact latent representations that capture task-relevant distribution shifts and leverages the Markov blanket of the label—comprising its parents, children, and spouses—to guide identifiable representation learning. It reveals for the first time that representations relying solely on complete predictive information are underdetermined under general settings, thereby circumventing strong assumptions commonly required by existing approaches, such as independent latent variables or invariant label distributions. By integrating causal representation learning with nonparametric deep models, the framework achieves theoretically guaranteed identifiability and significantly improves target-domain generalization across diverse distribution shift scenarios.

This paper addresses unsupervised domain adaptation under unobserved subgroup structure in the source domain: the source is partitioned into four subgroups by binary label $Y$ and binary background variable $A$, with the subgroup $A=1,Y=1$ entirely missing. Ignoring this missingness induces prediction bias on the target domain. To address this, we propose a background-specific–global joint modeling framework—the first method enabling recoverable target-domain prediction under partial subgroup unobservability—accompanied by an upper bound on estimation error and asymptotic consistency guarantees. Our approach estimates latent subgroup proportions via distribution matching, without imposing strong assumptions on the missingness mechanism. Experiments on synthetic data and real-world medical and image datasets demonstrate that our method significantly outperforms naive baselines that ignore the missing structure, achieving higher accuracy and improved robustness on the target domain.

本文提出了一种针对数据可用性随时间演变的领域迁移学习问题（TrED），并探讨了现有方法在处理整个演化过程中的不足。

This work addresses the privacy risks in source-free domain adaptation, where models trained on the source domain may inadvertently leak information about source-private classes into the target domain. To tackle this issue, we formally introduce and solve the problem of source-private class forgetting under a novel setting termed SCADA-UL, extending it to scenarios involving continual forgetting and unknown forgotten classes. By leveraging adversarially generated forgetting samples, a label rescaling strategy, and adversarial optimization, our method achieves effective machine unlearning under distribution shift. Experimental results demonstrate that the proposed approach attains forgetting performance comparable to full retraining across multiple benchmark datasets, significantly outperforming existing baselines.