The article presents a novel graph Spectral Alignment (SPA) framework for unsupervised domain adaptation (UDA) in machine learning. The SPA framework aims to extend the in-domain model to distinctive target domains where data distributions differ. The core of the method includes a coarse graph alignment mechanism and a fine-grained message propagation module for enhanced discriminability in the target domain. The method surpasses existing domain adaptation methods in terms of performance, efficacy, robustness, discriminability, and transferability.
Publication date: 26 Oct 2023
Project Page: https://github.com/CrownX/SPA
Paper: https://arxiv.org/pdf/2310.17594
