Codes and datasets for Handling Distribution Shifts on Graphs: An Invariance Perspective. This work focuses on distribution shifts on graph data and proposes a new approach Explore-to-Extrapolate Risk Minimization (EERM) for out-of-distribution generalization.
In our experiment, we consider three types of distribution shifts. You can make a directory ./data and download the datasets according the following details.
-
Artificial Transformation: We use Cora and Amazon-Photo datasets to construct spurious node features. The data construction script is provided in
./synthetic/synthetic.py. The original datasets can easily accessed via Pytorch Geometric package. To download our preprocessed data, please go to the Google drive:https://drive.google.com/drive/folders/15YgnsfSV_vHYTXe7I4e_hhGMcx0gKrO8?usp=sharing -
Cross-Domain Transfer: We use Twitch-Explicit and Facebook-100 datasets. These two datasets both contain multiple graphs. We use different graphs for training/validation/testing. To download the dataset:
-
Temporal Evolution: We use Elliptic and OGBN-Arxiv datasets. One can download the Elliptic data from Kaggle dataset. For OGB dataset, see the OGB website for more details.
More information will be updated.
@inproceedings{wu2022eerm,
title = {Handling Distribution Shifts on Graphs: An Invariance Perspective},
author = {Qitian Wu and Hengrui Zhang and Junchi Yan and David Wipf},
booktitle = {International Conference on Learning Representations (ICLR)},
year = {2022}
}