GALA is a novel source-free graph domain adaptation method that uses graph diffusion models to reconstruct source-style graphs from target data, combined with a graph jigsaw strategy to enhance model robustness. The method effectively addresses domain shift and label scarcity challenges in graph neural networks without requiring access to source data during adaptation.
@ARTICLE{gala2024,
author={Luo, Junyu and Gu, Yiyang and Luo, Xiao and Ju, Wei and Xiao, Zhiping and Zhao, Yusheng and Yuan, Jingyang and Zhang, Ming},
journal={ IEEE Transactions on Pattern Analysis & Machine Intelligence },
title={ GALA: Graph Diffusion-based Alignment with Jigsaw for Source-free Domain Adaptation },
year={2024},
volume={},
number={01},
ISSN={1939-3539},
pages={1-14},
doi={10.1109/TPAMI.2024.3416372},
month={June}
}
RNA is a novel source-free graph domain adaptation method that combines spectral seriation for ranking graph similarities and harmonic graph alignment for effective knowledge transfer. The method addresses the challenges of label scarcity and domain shifts by leveraging robust pairwise rankings and extracting domain-invariant subgraphs, without requiring access to source domain data.
@inproceedings{rna2024,
title = {Rank and Align: Towards Effective Source-free Graph Domain Adaptation},
author = {Luo, Junyu and Xiao, Zhiping and Wang, Yifan and Luo, Xiao and Yuan, Jingyang and Ju, Wei and Liu, Langechuan and Zhang, Ming},
booktitle = {Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence},
pages = {4706--4714},
year = {2024}
}
