Research map / Label-efficient and robust learning / Domain adaptation and generalization

On Unsupervised Domain Adaptation: Pseudo Label Guided Mixup for Adversarial Prompt Tuning

AAAI 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Feature alignment for unsupervised adaptation · 9 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PL-Mix, a pseudo-label-confidence-guided Mixup added to adversarial soft prompt tuning for unsupervised domain adaptation on Amazon Reviews, with a generalization bound justifying it.

Problem

Adversarial UDA methods align source and target features without class-dependent alignment, so source data may align with wrongly labeled target data, hurting target classification; pseudo labels for target data are noisy.

Approach

Adjusts the Beta Mixup ratio between source and target samples using the confidence of target pseudo labels (larger movement toward same-label target data, less toward different-label data, and dropping low-confidence labels), applied to both the label classifier and the domain discriminator within adversarial prompt tuning; supported by an information-theoretic generalization bound.

Key results

• PL-Mix with RoBERTa-base reaches 93.84% average accuracy over 12 Amazon Reviews tasks, versus 93.02% for AdSPT (reproduced/reported SOTA).
• With BERT-base-uncased, PL-Mix averages 90.80%, compared with 90.13% for AdSPT and 90.55% for reproduced COBE.
• In the ablation, the classifier-only variant PL-Mix-C reaches 93.64% and the discriminator-only variant PL-Mix-D reaches 93.41% on RoBERTa, both above AdSPT's 93.02%.

Measured effects reported in the paper

Surprising findings

• Reported gains over AdSPT are modest on BERT (about 0.7 points), and the paper reports that its reproduced AdSPT numbers differ from the published ones.

Limitations

• Theoretical analysis does not characterize the effect of prompt tuning on marginal alignment, which the authors state is beyond scope.
• Evaluation is limited to the Amazon Reviews sentiment benchmark with small encoders (BERT-base, RoBERTa-base).

Experimental setup

Datasets: Amazon Reviews

Benchmarks: Amazon Reviews (12 single-source tasks) Amazon Reviews multi-source setting

Metrics: accuracy

Baselines: DANN COBE AdSPT DM-ADA IIMT DMRL

Builds on: AdSPT DANN Mixup DM-ADA Wang and Mao (2023) information-theoretic UDA bound

Closest papers by meaning

• Adversarial Domain Adaptation with Domain Mixup (AAAI 2020 · similarity 0.90)
• Implicit Class-Conditioned Domain Alignment for Unsupervised Domain Adaptation (ICML 2020 · similarity 0.89)
• Connect, Not Collapse: Explaining Contrastive Learning for Unsupervised Domain Adaptation (ICML 2022 · similarity 0.89)
• CASUAL: Conditional Support Alignment for Domain Adaptation with Label Shift (AAAI 2025 · similarity 0.89)
• Can We Evaluate Domain Adaptation Models Without Target-Domain Labels? (ICLR 2024 · similarity 0.88)
• Toward Improving Robustness and Accuracy in Unsupervised Domain Adaptation (AAAI 2025 · similarity 0.88)