GRAVER: Generative Graph Vocabularies for Robust Graph Foundation Models Fine-tuning

This repository contains the official implementation of GRAVER, proposed in the paper GRAVER: Generative Graph Vocabularies for Robust Graph Foundation Models Fine-tuning (NeurIPS 2025), which is a robust and efficient fine-tuning framework for Graph Foundation Models (GFMs) under few-shot cross-domain settings.

1. Environment Requirements

Core Dependency Versions

The following packages are required to run the project:

• torch == 1.10.1+cu113
• torch-geometric == 2.5.3
• torch-cluster == 1.5.9
• torch-scatter == 2.0.9
• torch-sparse == 0.6.12
• torch-spline-conv == 1.2.1
• scikit-learn == 1.3.2
• numpy == 1.24.4
• pandas == 2.0.3
• matplotlib == 3.7.5
• scipy == 1.10.1
• huggingface-hub == 0.29.3
• wandb == 0.19.8
• tqdm == 4.67.1
• protobuf == 5.29.3
• requests == 2.32.3
• typing_extensions == 4.12.2
• attrs == 25.2.0
• dgl == 0.5.0

To install all dependencies:

pip install -r requirements.txt

---

2. Dataset Download & Organization

You can download the full dataset from HuggingFace:

📦 Click to Download GRAVER Dataset

Then unzip and organize it under the data/ directory as follows:

data/
├── ori_data/                 # Original graph data
├── save_model/               # Pretrained models
├── test_split/graph/         # Test indices for graph classification
├── few_data/                 # Few-shot data for node classification
└── few_data_graph/           # Few-shot data for graph classification

---

3. Quick Start

Node Classification

Run a 1-shot cross-dataset node classification task:

cd scripts
python main.py --dataset cora --shot_num 1 --experiment_type cross-dataset

Graph Classification

Run a 1-shot cross-domain graph classification task:

cd scripts_graph
python main.py --dataset wikics --shot_num 1 --experiment_type cross-domain

---

4. Full Pipeline: Pretraining + Finetuning

To run the full pretrain-then-finetune pipeline:

python main.py --dataset cora --shot_num 1 --experiment_type cross-dataset

---

5. Recommended Settings to Reproduce Results

Node Classification (e.g., Cora)

python main.py --dataset cora --shot_num 1 --experiment_type cross-dataset

Recommended parameters:

--lr 0.00001 
--hid_units 256 
--lambda_entropy 0.204 
--downstreamlr 0.001

Graph Classification (e.g., Wiki-CS)

python main.py --dataset wikics --shot_num 1 --experiment_type cross-domain

Recommended parameters:

--lr 0.00001 
--hid_units 256 
--lambda_entropy 0.078 
--downstreamlr 0.001

---

6. Project Structure

GRAVER/
├── scripts/                   # Node classification scripts
├── scripts_graph/             # Graph classification scripts
├── GRAVER/                    # Node classification model code
├── GRAVER_graph/              # Graph classification model code
├── config.py                  # Central configuration
└── requirements.txt           # Dependency list

---

7. Using wandb (Optional)

To enable experiment tracking with wandb:

wandb login [YOUR_API_KEY]

Metrics tracked:

• Training loss & accuracy
• Top-20 group accuracy rankings
• Path of best saved pretrained model

---

8. Full Argument List

Here are the supported command-line arguments:

Argument	Description	Default
--dataset	Target dataset name	"cora"
--seed	Random seed	39
--gpu	GPU index	0
--experiment_type	cross-dataset or cross-domain	"cross-dataset"
--lr	Learning rate (pre-train)	0.00001
--l2_coef	L2 regularization	0.0
--hid_units	Hidden dimension	256
--lambda_entropy	Entropy loss weight	0.204
--downstreamlr	Learning rate (fine-tune)	0.001
--combinetype	Feature combination strategy (add, mul, etc.)	"mul"
--model_path	Path to load pre-trained model	"unwork"
--nb_epochs	Pre-training epochs	10000
--shot_num	Few-shot samples per class	1
--fw_epochs	Fine-tuning epochs	1
--prompt_times	Number of prompt averaging times	20
--disenconv_init_k	Initial factor count	2
--disenconv_delta_k	Factors reduced per layer	0
--disenconv_routit	Routing iterations	1
--disenconv_tau	Temperature for softmax	1.0
--disenconv_dropout	Dropout rate	0.2
--disenconv_num_layers	Number of DisenGCN layers	1
--groups	Number of candidate groups	20
--test_num_nodes	Number of test nodes	300

For more examples and help, please refer to the scripts/ and scripts_graph/ folders.

9. Citation

If you find this work useful, please cite:

@inproceedings{yuan2025graver,
  author    = {Yuan, Haonan and Sun, Qingyun and Shi, Junhua and Fu, Xingcheng and Hooi, Bryan and Li, Jianxin and Yu, Philip S},
  title     = {{GRAVER}: Generative Graph Vocabularies for Robust Graph Foundation Models Fine-tuning},
  booktitle = {The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year      = {2025}
}

---

📬 Contact

For questions or discussions, please contact Haonan Yuan or open an issue in this repository.