SAMGPT

We provide the code (in pytorch) and datasets for our paper "SAMGPT: Text-free Graph Foundation Model for Multi-domain Pre-training and Cross-domain Adaptation", which is accepted by WWW2025.

Description

The repository is organised as follows:

• data/: contains data we use.
• cache/: stores intermediate results during pre-training.
• checkpoints/: contains checkpoints of the models after pre-training.
• result/: stores results when downstream task is done.
• src/: implements pre-training and downstream task.

Package Dependencies

• python 3.8.16
• pytorch 1.10.1
• cuda 11.3
• pyG 2.1.0

Running experiments

Data Preparation

To download the datasets and generate fewshot data, run python ./src/utils/generate_fewshot.py

Node Classification and Graph Classification

To train and evaluate SAMGPT on various target domains, you may need to change the corresponding parameters in execute.py.

Pre-training and prompt tuning:

python ./src/execute.py 
  --dataset {Name of the downstream dataset, default: Cora} 
  --pretrain_datasets {List of pretrain datasets, default: ['Citeseer', 'Pubmed', 'Photo', 'Computers', 'FacebookPagePage', 'LastFMAsia']} 
  --downstream_task {Type of downstream task, options: node or graph, default: node}
  --gpu {GPU device number, default: 0} 
  --shot_num {Number of examples for few-shot learning, default: 1} 
  --pretrain_method {Pretraining method, options: GRAPHCL, LP, or splitLP, default: GRAPHCL}  

Mini-batch implementation

An example for mini-batch implementation is shown in ./src_batch.