Towards enhancing prototypes driven by graph convolutional network for domain adaptation

Abstract

Domain adaptation (DA) is essential for transferring knowledge across domains with differing distributions, yet challenges like domain shifts and scarce labeled data limit performance. Prototype-based methods show promise on the DA task. This work introduces a prototype-based method, termed enhanced prototypical network (EnPro), for unsupervised domain adaptation (UDA) and semi-supervised domain adaptation (SSDA) settings with consistent architecture and training. We provide a theoretical analysis dividing the DA mapping space into consensus, vicinal, and vulnerable spaces. This improves classification by expanding the consensus and vicinal spaces while reducing the vulnerable space. To achieve this, we use a graph convolutional network (GCN) to increase labeled target samples through reliable pseudo-labels and enhanced prototypes. Experiments on UDA and SSDA benchmark datasets demonstrate state-of-the-art performance.