Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Contrastive Conditional Alignment based on Label Shift Calibration for Imbalanced Domain Adaptation

Xiaona Sun, Zhenyu Wu, Zhiqiang Zhan, Yang Ji

TL;DR

This work tackles imbalanced domain adaptation by addressing coexisting covariate and label shifts. It introduces CCA-LSC, a two-stage framework that first applies Contrastive Conditional Alignment to learn domain-invariant and class-discriminative features, then uses Label Shift Calibration to adjust target pseudo-labels according to per-class label shift via the metric $M_{ls}$. The method combines domain adversarial learning, weighted centroid alignment, and discriminative feature alignment with calibrated pseudo-labels to improve pseudo-label reliability and reduce misalignment, achieving state-of-the-art results on OfficeHome and DomainNet. Extensive ablations and analysis demonstrate the effectiveness of each component and the two-stage learning strategy, with practical implications for robust, real-world IDA where label distributions differ across domains.

Abstract

Many existing unsupervised domain adaptation (UDA) methods primarily focus on covariate shift, limiting their effectiveness in imbalanced domain adaptation (IDA) where both covariate shift and label shift coexist. Recent IDA methods have achieved promising results based on self-training using target pseudo labels. However, under the IDA scenarios, the classifier learned in the source domain will exhibit different decision bias from the target domain. It will potentially make target pseudo labels unreliable, and will further lead to error accumulation with incorrect class alignment. Thus, we propose contrastive conditional alignment based on label shift calibration (CCA-LSC) for IDA, to address both covariate shift and label shift. Initially, our contrastive conditional alignment resolve covariate shift to learn representations with domain invariance and class discriminability, which include domain adversarial learning, sample-weighted moving average centroid alignment and discriminative feature alignment. Subsequently, we estimate the probability distribution of the target domain, and calibrate target sample classification predictions based on label shift metrics to encourage labeling pseudo-labels more consistently with the distribution of real target data. Extensive experiments are conducted and demonstrate that our method outperforms existing UDA and IDA methods on benchmarks with both label shift and covariate shift. Our code is available at https://github.com/ysxcj-hub/CCA-LSC.

Contrastive Conditional Alignment based on Label Shift Calibration for Imbalanced Domain Adaptation

TL;DR

This work tackles imbalanced domain adaptation by addressing coexisting covariate and label shifts. It introduces CCA-LSC, a two-stage framework that first applies Contrastive Conditional Alignment to learn domain-invariant and class-discriminative features, then uses Label Shift Calibration to adjust target pseudo-labels according to per-class label shift via the metric . The method combines domain adversarial learning, weighted centroid alignment, and discriminative feature alignment with calibrated pseudo-labels to improve pseudo-label reliability and reduce misalignment, achieving state-of-the-art results on OfficeHome and DomainNet. Extensive ablations and analysis demonstrate the effectiveness of each component and the two-stage learning strategy, with practical implications for robust, real-world IDA where label distributions differ across domains.

Abstract

Many existing unsupervised domain adaptation (UDA) methods primarily focus on covariate shift, limiting their effectiveness in imbalanced domain adaptation (IDA) where both covariate shift and label shift coexist. Recent IDA methods have achieved promising results based on self-training using target pseudo labels. However, under the IDA scenarios, the classifier learned in the source domain will exhibit different decision bias from the target domain. It will potentially make target pseudo labels unreliable, and will further lead to error accumulation with incorrect class alignment. Thus, we propose contrastive conditional alignment based on label shift calibration (CCA-LSC) for IDA, to address both covariate shift and label shift. Initially, our contrastive conditional alignment resolve covariate shift to learn representations with domain invariance and class discriminability, which include domain adversarial learning, sample-weighted moving average centroid alignment and discriminative feature alignment. Subsequently, we estimate the probability distribution of the target domain, and calibrate target sample classification predictions based on label shift metrics to encourage labeling pseudo-labels more consistently with the distribution of real target data. Extensive experiments are conducted and demonstrate that our method outperforms existing UDA and IDA methods on benchmarks with both label shift and covariate shift. Our code is available at https://github.com/ysxcj-hub/CCA-LSC.
Paper Structure (28 sections, 2 theorems, 12 equations, 3 figures, 5 tables)

This paper contains 28 sections, 2 theorems, 12 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Unsupervised Domain Adaptation With Covariate Shift
  4. Unsupervised Domain Adaptation With Label Shift
  5. Imbalanced Domain Adaptation
  6. Method
  7. Problem Setup
  8. Contrastive Conditional Alignment (CCA)
  9. Domain Adversarial Learning
  10. Sample-weighted Moving Average Centroid Alignment
  11. Discriminative Feature Alignment
  12. Label Shift Calibration (LSC)
  13. Label Shift Metrics $M_{ls}$
  14. Label Shift Calibration (LSC)
  15. Overall Optimization and Analysis
  16. ...and 13 more sections

Key Result

theorem thmcountertheorem

Denote $h \in \mathcal{H}$ as the hypothesis. Given two domains $\mathcal{S}$ and $\mathcal{T}$, the target error $\varepsilon_{\mathcal{T}}$ is bounded by three terms: (i) $\varepsilon_{\mathcal{S}}$: source error, (ii) $d_{\mathcal{H}\Delta \mathcal{H}}(\mathcal{S},\mathcal{T})$: the discrepancy d

Figures (3)

  • Figure 1: (a)Top: In cases of substantial label shift, classifier learned from source domain may mislabel target samples due to the unknown target label distribution. This can result in error accumulation and misalignment in IDA methods that use self-training with pseudo-labels. (a)Bottom: Our approach rectifies the classification boundary to predict target samples based on the label shift metric $M_{ls}$, effectively reducing the error rates in estimating target pseudo-labels. We employ calibrated pseudo-labels in CCA to learn feature representations that are both domain-invariant and class-discriminative. (b): Label distributions on DomainNet and OfficeHome
  • Figure 2: Analysis of label shift calibration. (a) Accuracy under different degrees of imbalance on Cl$\to$Pr. (b) The proportion of target samples with calibrated pseudo labels($\widehat{y} \neq \widehat{y}^m$) via CCA-LSC. (c) and (d) :The accuracy of the target pseudo labels $\widehat{y}$ (obtained by the classifier) and $\widehat{y}^m$ (calibrated based on label shift metric $M_{ls}$) in all calibrated target samples ($\widehat{y} \neq \widehat{y}^m$) on Cl$\to$Pr and S$\to$P respectively.
  • Figure 3: The impact of selecting target samples with different confidence levels on the estimation of target label distribution

Theorems & Definitions (2)

  • theorem thmcountertheorem: ben2010theory
  • theorem thmcountertheorem: xie2018learning