Corrections to "Joint Domain Adaptation Based on Adversarial Dynamic Parameter Learning"

Yumeng Yuan, Yuhua Li, Zhenlong Zhu, Ruixuan Li, Xiwu Gu

Introduction

Domain Adaptation (DA) is a critical challenge in machine learning, focusing on the transfer of knowledge from a labeled source domain to an unlabeled target domain with a different but related distribution. A common example is sentiment analysis, where a model trained on book reviews must adapt to the specific vocabulary of electronics reviews. The Adversarial Dynamic Parameter Learning (ADPL) framework was proposed to address "Joint Domain Adaptation" by aligning both marginal and conditional distributions. This paper provides essential corrections and optimizations to the original ADPL implementation, establishing a new state-of-the-art performance benchmark.

Methodology

The ADPL framework utilizes an adversarial learning structure, similar to Generative Adversarial Networks (GANs), to extract domain-invariant features. The goal is to train a feature extractor that produces representations from which a domain discriminator cannot distinguish the source from the target.

A pivotal component of this framework is the Weighted GAN Loss Function. This formulation is designed to refine the quality of learned features by dynamically adjusting the importance of different samples during the alignment process. To ensure a rigorous evaluation, the authors also re-tuned competitive baselines, such as MEDA, to provide a fair comparison against the optimized ADPL code.

Experimental Results

The effectiveness of the ADPL framework was evaluated using the Amazon review dataset (12 cross-domain tasks) and image datasets (CIFAR-10 and STL-10).

Quantitative Performance

The optimized ADPL achieved an average accuracy of 81.70% on the Amazon review dataset, outperforming all traditional and deep learning baselines. It secured the highest accuracy in 11 out of 12 individual tasks. Notably, it surpassed EasyTL by 2.2%, the updated MEDA by 4.63%, and WDGRL by 2.4%. While EasyTL performed slightly better on the "Electronics to Books" task due to its flexibility in high-discrepancy scenarios, ADPL remained the dominant method overall.

Qualitative Visualization

The impact of the weighted GAN loss is most evident in the t-SNE feature visualizations for the $uc \rightarrow us$ task.

t-SNE visualization comparing source domain (a), target domain with weighted GAN loss (b), and target domain with unweighted GAN loss (c).

As shown in the figure above, the weighted GAN loss (b) produces significantly tighter and more distinct clusters in the feature space compared to the unweighted version (c). This visual evidence confirms that the proposed weighting mechanism successfully aligns the target domain features with the source domain distribution, which is the primary objective of effective domain adaptation.

Conclusion

The corrections and code optimizations presented here solidify ADPL as a robust framework for joint domain adaptation. By refining the adversarial learning process and the associated loss functions, the authors demonstrate significant performance gains over existing benchmarks, supported by both quantitative metrics and qualitative feature visualizations.

Summary updated

BibTeX
@article{yuan2021corrections,
  author = {Yumeng Yuan and Yuhua Li and Zhenlong Zhu and Ruixuan Li and Xiwu Gu},
  title = {{Corrections to "Joint Domain Adaptation Based on Adversarial Dynamic Parameter Learning"}},
  journal = {IEEE Trans. Emerg. Top. Comput. Intell.},
  year = {2021},
  doi = {10.1109/tetci.2021.3086063},
}
Figure