Machine Unlearning For Alleviating Negative Transfer In Partial-Set Source-Free Unsupervised Domain Adaptation

Source-free Unsupervised Domain Adaptation (SFUDA) aims to adjust a source model trained on the source domain to a related but unlabeled target domain without accessing the source data. Many SFUDA methods are studied in closed-set scenarios where the target domain and source domain categories are perfectly aligned. However, a more practical scenario is a partial-set scenario where the source label space subsumes the target one. In this paper, we prove that reducing the differences between the source and target domains in the partial-set scenario helps to achieve domain adaptation. And we propose a simple yet effective SFUDA framework called the Machine Unlearning Framework to alleviate the negative transfer problem in the partial-set scenario, thereby allowing the model to focus on the target domain category. Specifically, we first generate noise samples for each category that only exists in the source domain and generate pseudo-labeled samples from the target domain. Then, in the forgetting stage, we use these samples to train the model, making it behave like the model has never seen the class that only exists in the source domain before. Finally, in the adaptation stage, we use only the pseudo-labeled samples to conduct self-supervised training on the model, making it more adaptable to the target domain. Our method is easy to implement and pluggable, suitable for various pre-trained models. Experimental results show that our method can well alleviate the negative transfer problem and improve model performance under various target domain category settings.