Collaborate to Adapt: Source-Free Graph Domain Adaptation via Bi-directional Adaptation (WWW 2024)
This is a PyTorch implementation of the GraphCTA algorithm, which tries to address the domain adaptation problem without accessing the labelled source graph. It performs model adaptation and graph adaptation collaboratively through a series of procedures: (1) conduct model adaptation based on node's neighborhood predictions in target graph considering both local and global information; (2) perform graph adaptation by updating graph structure and node attributes via neighborhood constrastive learning; and (3) the updated graph serves as an input to facilitate the subsequent iteration of model adaptation, thereby establishing a collaborative loop between model adaptation and graph adaptation.
- python3.8
- pytorch==2.0.0
- torch-scatter==2.1.1+pt20cu118
- torch-sparse==0.6.17+pt20cu118
- torch-cluster==1.6.1+pt20cu118
- torch-geometric==2.3.1
- numpy==1.24.3
- scipy==1.10.1
- tqdm==4.65.0
Datasets used in the paper are all publicly available datasets. You can find Elliptic, Twitch and Citation via the links.
Just execuate the following command for source model pre-training:
python train_source.py
Then, execuate the following command for adaptation:
python train_target.py
If you find GraphCTA useful for your research, please consider citing the following paper:
@article{zhang2024collaborate,
title={Collaborate to Adapt: Source-Free Graph Domain Adaptation via Bi-directional Adaptation},
author={Zhang, Zhen and Liu, Meihan and Wang, Anhui and Chen, Hongyang and Li, Zhao and Bu, Jiajun and He, Bingsheng},
journal={arXiv preprint arXiv:2403.01467},
year={2024}
}