The official implementation for ICML2024 paper "Learning Divergence Fields for Shift-Robust Graph Representations"
Related material: [Paper], [Blog (Chinese)], [Blog (English)], [Slides]
The implementation of training pipeline for observed (Arxiv and Twitch) is based on EERM.
The implementation of training pipeline for unobserved (STL and CIFAR) is based on DIFFormer.
[2024.06.03] We release the code for our model on real-world datasets that involve Observed Geometries, Partially Observed Geometries and Unobserved Geometries. More detailed info will be updated soon.
[2024.06.04] We upload the datasets (Arxiv, Twitch, Cifar, STL, DPPIN).
One can download the datasets (Arxiv, Twitch, Cifar, STL, DPPIN) from the google drive link below:
https://drive.google.com/drive/folders/1LctHB8_8fRqp3jq9kU3DryHXwA5PCihC
We propose a geometry diffusion model that is optimized by a new learning objective (comprised of a supervised term and a regularization term) for the generalization problem with interdependent data.
The following tables present the results of generalization with different data geometries.
Python 3.8, PyTorch 1.13.0, PyTorch Geometric 2.1.0, NumPy 1.23.4
Please refer to the bash script run.sh in each folder for running the training and evaluation pipeline on different datasets.
If you find our code and model useful, please consider citing our work. Thank you!
@inproceedings{wu2024glind,
title = {Learning Divergence Fields for Shift-Robust Graph Representations},
author = {Qitian Wu and Fan Nie and Chenxiao Yang and Junchi Yan},
booktitle = {International Conference on Machine Learning (ICML)},
year = {2024}
}