[Paper Review] A Balanced and Uncertainty-aware Approach for Partial Domain Adaptation

Jian Liang, Yunbo Wang+3 more|arXiv (Cornell University)|Mar 5, 2020

Domain Adaptation and Few-Shot Learning50 references5 citations

TL;DR

This paper proposes BA3US, a novel domain adaptation method for partial domain adaptation that addresses negative transfer and uncertainty propagation through Balanced Adversarial Alignment (BAA) and Adaptive Uncertainty Suppression (AUS). BAA balances label distributions by augmenting the target domain with sampled source data, while AUS uses a confidence-weighted complement entropy objective to suppress high prediction scores for incorrect classes, achieving state-of-the-art performance on multiple benchmarks.

ABSTRACT

This work addresses the unsupervised domain adaptation problem, especially in the case of class labels in the target domain being only a subset of those in the source domain. Such a partial transfer setting is realistic but challenging and existing methods always suffer from two key problems, negative transfer and uncertainty propagation. In this paper, we build on domain adversarial learning and propose a novel domain adaptation method BA$^3$US with two new techniques termed Balanced Adversarial Alignment (BAA) and Adaptive Uncertainty Suppression (AUS), respectively. On one hand, negative transfer results in misclassification of target samples to the classes only present in the source domain. To address this issue, BAA pursues the balance between label distributions across domains in a fairly simple manner. Specifically, it randomly leverages a few source samples to augment the smaller target domain during domain alignment so that classes in different domains are symmetric. On the other hand, a source sample would be denoted as uncertain if there is an incorrect class that has a relatively high prediction score, and such uncertainty easily propagates to unlabeled target data around it during alignment, which severely deteriorates adaptation performance. Thus we present AUS that emphasizes uncertain samples and exploits an adaptive weighted complement entropy objective to encourage incorrect classes to have uniform and low prediction scores. Experimental results on multiple benchmarks demonstrate our BA$^3$US surpasses state-of-the-arts for partial domain adaptation tasks. Code is available at \url{https://github.com/tim-learn/BA3US}.

Motivation & Objective

• To address the challenge of negative transfer in partial domain adaptation, where source-only classes mislead target domain alignment.
• To mitigate uncertainty propagation, where high prediction scores for incorrect classes in source data degrade alignment with target data.
• To transform the partial domain adaptation problem into a more tractable unsupervised domain adaptation task via target domain augmentation.
• To improve model robustness and accuracy by jointly optimizing for label distribution balance and uncertainty suppression.

Proposed method

• BAA balances label distributions by randomly sampling a few source samples to augment the smaller target domain during adversarial training, ensuring symmetric class representation across domains.
• AUS introduces an adaptive weighted complement entropy loss that emphasizes uncertain samples (low confidence) and encourages uniform, low prediction scores for incorrect classes.
• The method integrates BAA and AUS into a unified framework built on domain adversarial networks, using gradient reversal for domain alignment.
• Class-level weighting is applied to filter out source-only classes, focusing adaptation on shared classes.
• The model uses a confidence-weighted complement entropy objective that dynamically adjusts based on prediction confidence, suppressing uncertainty propagation.
• Training involves iterative optimization with adaptive sampling rate ρ and periodic re-estimation of class weights m to improve alignment and generalization.

Experimental results

Research questions

• RQ1 Can augmenting the target domain with a few source samples effectively balance label distributions and reduce negative transfer in partial domain adaptation?
• RQ2 How does uncertainty propagation from high prediction scores of incorrect classes affect domain alignment, and can it be mitigated?
• RQ3 Does an adaptive complement entropy objective that emphasizes low-confidence samples improve generalization in partial domain adaptation?
• RQ4 Can the proposed method achieve state-of-the-art performance on standard partial domain adaptation benchmarks like Office-Home and Office31?
• RQ5 Is the uncertainty suppression mechanism effective beyond partial domain adaptation, including in standard closed-set domain adaptation?

Key findings

• BA3US achieves state-of-the-art performance on the Office-Home and Office31 benchmarks, with 75.98% accuracy on Ar→Cl and 97.81% on A→D in the Office31 dataset.
• The method shows robustness to hyperparameter choices, with stable performance across different values of β and ξ, indicating low sensitivity.
• The adaptive complement entropy objective outperforms the standard complement entropy (ξ=0) by 0.4–0.5% in accuracy, demonstrating its effectiveness.
• t-SNE visualizations show that BA3US produces more discriminative and domain-aligned features compared to baseline methods.
• The convergence analysis confirms that BA3US converges quickly and outperforms E-DANN and 'source only' baselines, especially after 1,000 iterations.
• The method achieves significant gains over existing SOTA methods, with improvements of up to 1.5% on challenging partial domain adaptation tasks.