GrapHoST

This repository provides the official implementation and reproduction instructions for WSDM26 paper Does Homophily Help in Robust Test-time Node Classification?.

---

Requirements

We use Python 3.8. Please install all dependencies listed in requirements.txt:

pip install -r requirements.txt

The required packages include:

deeprobust==0.2.5  
dgl==0.9.1  
dgl_cu102==0.6.1  
GCL==0.6.11  
googledrivedownloader==0.4  
ipdb==0.13.7  
matplotlib==3.5.2  
networkx==2.5  
numpy==1.20.1  
ogb==1.3.5  
pandas==1.2.3  
scikit_learn==1.1.3  
scipy==1.6.2  
torch==1.13.0  
torch_geometric==2.0.1  
torch_scatter==2.0.8  
torch_sparse==0.6.12  
tqdm==4.60.0  
visualization==1.0.0  

---

Download Datasets

To ensure a fair comparison with baseline methods in a consistent setting, we use the datasets provided by EERM and GTrans.

Follow the steps below:

1. Download the datasets from the original shared Google Drive folder:

https://drive.google.com/drive/folders/15YgnsfSV_vHYTXe7I4e_hhGMcx0gKrO8?usp=sharing

2. Place all downloaded folders into ./GraphOOD-EERM/data/.

Your directory should look like:

project/
│   README.md
│   requirements.txt
│   reproduce_results.sh
│   ...
└───GraphOOD-EERM/
    └───data/
        ├── Amazon/
        ├── ...

3. To make it more convenient, we have already included all datasets except the large datasets arxiv, fb100, and elliptic in this repository. You only need to manually download these large datasets.

4. You can freely reproduce results for all other datasets, such as cora, amazon-photo, twitch-e, actor, chameleon, and squirrel, without any additional downloads.

---

Reproduction

After installing the environment and preparing the dataset, you can reproduce the main results in one line:

bash reproduce_results.sh

This script will run the full set of experiments and generate the results reported in the main table of the paper.

GrapHoST improves the test-time performance of a fixed pre-trained GNN, so the final results depending on the base GNN. Therefore we report results before and after applying GrapHoST, and the performance difference demonstrates GrapHoST effectiveness.

---

Acknowledgement

This repository takes credit from GTrans for the feature transformation and EERM for the dataset.

---

Reference

Please cite our paper:

@inproceedings{GrapHoST,
  title = {Does Homophily Help in Robust Test-time Node Classification?},
  author = {Yan Jiang and Ruihong Qiu and Zi Huang},
  booktitle = {ACM International Conference on Web Search and Data Mining (WSDM)},
  year = {2026}
  }

---

Contact

For any questions or issues, please feel free to contact us or open an issue in the repository.