Pred&Guide: Labeled Target Class Prediction for Guiding Semi-Supervised Domain Adaptation
This addresses domain adaptation for classification tasks where labeled target data is scarce, offering an incremental improvement over existing methods.
The paper tackles semi-supervised domain adaptation by proposing Pred&Guide, a framework that uses inconsistency between predicted and actual labels of few labeled target examples to guide adaptation, achieving state-of-the-art results on Office-Home and DomainNet datasets.
Semi-supervised domain adaptation aims to classify data belonging to a target domain by utilizing a related label-rich source domain and very few labeled examples of the target domain. Here, we propose a novel framework, Pred&Guide, which leverages the inconsistency between the predicted and the actual class labels of the few labeled target examples to effectively guide the domain adaptation in a semi-supervised setting. Pred&Guide consists of three stages, as follows (1) First, in order to treat all the target samples equally, we perform unsupervised domain adaptation coupled with self-training; (2) Second is the label prediction stage, where the current model is used to predict the labels of the few labeled target examples, and (3) Finally, the correctness of the label predictions are used to effectively weigh source examples class-wise to better guide the domain adaptation process. Extensive experiments show that the proposed Pred&Guide framework achieves state-of-the-art results for two large-scale benchmark datasets, namely Office-Home and DomainNet.