The complex nature of molecule graphs poses unique challenges to out-of-distribution (OOD) generalization, differentiating them from images and general graphs:
- Molecules with Similar Structures Can Have Different Functions:
- Functional Group: Acts as a causal subgraph.
- Scaffold: Serves as an environmental (noise) subgraph.
- Semantic Environments in OOD:
- Environments are related.
- Environments are hierarchical.
Figure 1. Challenges on molecule graphs in out-of-distribution setting.
Existing methods use flat enviornments to conduct the graph invariant learning. There are two limitations in flat environment infernece:
- Provided (Real): neglect local similarity among the numerous environments.
- Inference (Infer #2): Inferring from a small number of environments may fail to capture global similarity and interrelationships among the environments.
Figure 2. (a) Results on IC50-SCA dataset from DrugOOD. (b) Flat environments from existing approaches. (c) Hierarchical environments from our methods. For visualization, we set #real environments as 10.
Figure 3.Our Framework consists of (a) Hierarchical Stochastic Subgraph Generation, (b) Hierarchical Semantic Environments, (c) Robust GIL with Hierarchical Semantic Environments.
Our code is based on the following libraries:
torch==1.9.0+cu111
torch-geometric==2.0.2
plus the DrugOOD benchmark repo.
The data used in the paper can be obtained following these instructions.
We provide the hyperparamter tuning and evaluation details in the paper and appendix. In the below we give a brief introduction of the commands and their usage in our code. We provide the corresponding running scripts in the script folder.
Simply run
bash run.sh 0 icassay
with corresponding datasets and model specifications.
If you find our paper and repo useful, please cite our paper: -->
@inproceedings{piao2024improving,
title={Improving out-of-distribution generalization in graphs via hierarchical semantic environments},
author={Piao, Yinhua and Lee, Sangseon and Lu, Yijingxiu and Kim, Sun},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={27631--27640},
year={2024}
}Ack: The readme is inspired by CIGA. 😄