Improved Domain-Adversarial Neural Networks with Mixup and Pseudo-Labeling

Domain adaptation is essential for transferring knowledge from a labeled source domain to an unlabeled target domain, a process especially critical in healthcare due to the high costs and challenges associated with data collection and annotation. Domain-Adversarial Neural Networks (DANN) use adversarial training to reduce the feature distribution differences between the source and target domains. However, DANN primarily focuses on aligning feature distributions, potentially overlooking the valuable intrinsic characteristics present within the target domain. To overcome these challenges, we introduce an improved DANN framework that incorporates mixup and pseudo-labeling techniques. Mixup, a data augmentation method, generates new training samples by linearly interpolating between pairs of examples, enhancing the model’s generalization capability. Pseudo-labeling, a semi-supervised learning strategy, assigns labels to unlabeled data based on the model’s predictions. By integrating these techniques, our framework effectively leverages target domain data, improving domain adaptation performance. Experimental results indicate that our method surpasses conventional DANN approaches across multiple benchmark datasets.