| Quick Start | Website | Paper | Video | Media Coverage | Call For Contribution |
🌟ProG🌟 (Prompt Graph) is a library built upon PyTorch to easily conduct single or multi-task prompting for pre-trained Graph Neural Networks (GNNs). You can easily use this library to conduct various graph workflows like supervised learning, pre-training and prompting, and pre-training and finetuning for your node/graph-level tasks. The starting point of this library is our KDD23 paper All in One (Best Research Paper Award, which is the first time for Hong Kong and Mainland China).
The ori
branch of this repository is the source code of All in One, in which you can conduct even more kinds of tasks with more flexible graph prompts. BeyondAll in One, the main
branch of this library now supports more than 5 graph prompt models (e.g. All-in-One, GPPT, GPF Plus, GPF, GraphPrompt, etc) with more than 6 pre-training strategies (e.g. DGI, GraphMAE, EdgePreGPPT, EdgePreGprompt, GraphCL, SimGRACE, etc), and have been tested on more than 15 graph datasets, covering both homophilic and heterophilic graphs from various domains with different scales. Click here to see the full and latest supportive list (backbones, pre-training strategies, graph prompts, and datasets).
- Leader: Dr. Xiangguo SUN
- Consultants: Prof. Jia LI, Prof. Hong CHENG
- Developers: Mr. Chenyi ZI, Mr. Haihong ZHAO, Dr. Xiangguo SUN
- Insight Suggestions: Miss. Xixi WU (who also contributes a lot to our survey, repository, etc.)
- Clik Here to See Other Contributors
- 2024/06/08: We use our developed ProG to extensively evaluate various graph prompts, and released our analysis report as follows:
-
Chenyi Zi, Haihong Zhao, Xiangguo Sun, Yiqing Lin, Hong Cheng, Jia Li. ProG: A Graph Prompt Learning Benchmark. https://arxiv.org/abs/2406.05346
-
- 2024/05/28: We are so happy to announce that we have finished most of the updating works for ProG! (the
main
branch of this repository. If you wish to find the original ProG package, go to theori
branch) - 2024/01/01: A big updated version released!
- 2023/11/28: We released a comprehensive survey on graph prompt!
-
Xiangguo Sun, Jiawen Zhang, Xixi Wu, Hong Cheng, Yun Xiong, Jia Li. Graph Prompt Learning: A Comprehensive Survey and Beyond https://arxiv.org/abs/2311.16534
-
- 2023/11/15: We released a 🦀repository🦀 for a comprehensive collection of research papers, datasets, and readily accessible code implementations.
Pypi
From ProG 1.0 onwards, you can install and use ProG. For this, simply run
pip install prompt-graph
Or you can git clone our repository directly.
Before you begin, please make sure that you have Anaconda or Miniconda installed on your system. This guide assumes that you have a CUDA-enabled GPU.
# Create and activate a new Conda environment named 'ProG'
conda create -n ProG
conda activate ProG
# Install Pytorch and DGL with CUDA 11.7 support
# If your use a different CUDA version, please refer to the PyTorch and DGL websites for the appropriate versions.
conda install numpy
conda install pytorch==2.0.1 pytorch-cuda=12.2 -c pytorch -c nvidia
# Install additional dependencies
pip install torch_geometric pandas torchmetrics Deprecated
In addition, You can use our pre-train GNN directly or use our pretrain module to pre-train the GNN you want by
pip install torch_cluster -f https://data.pyg.org/whl/torch-2.3.0+cu121.html
the torch and cuda version can refer to https://data.pyg.org/whl/
The Architecture of ProG is shown as follows:
We have provided scripts with hyper-parameter settings to get the experimental results
In the pre-train phase, you can obtain the experimental results by running the parameters you want:
python pre_train.py --task Edgepred_Gprompt --dataset_name 'PubMed' --gnn_type 'GCN' --hid_dim 128 --num_layer 2 --epochs 1000 --seed 42 --device 0
In downstream_task, you can obtain the experimental results by running the parameters you want, for example,
python downstream_task.py --pre_train_model_path './Experiment/pre_trained_model/Cora/Edgepred_Gprompt.GCN.128hidden_dim.pth' --task NodeTask --dataset_name 'Cora' --gnn_type 'GCN' --prompt_type 'GPF-plus' --shot_num 1 --hid_dim 128 --num_layer 2 --lr 0.02 --decay 2e-6 --seed 42 --device 0
python downstream_task.py --pre_train_model_path './Experiment/pre_trained_model/BZR/DGI.GCN.128hidden_dim.pth' --task GraphTask --dataset_name 'BZR' --gnn_type 'GCN' --prompt_type 'All-in-one' --shot_num 1 --hid_dim 128 --num_layer 2 --lr 0.02 --decay 2e-6 --seed 42 --device 1
Perform a random search of hyperparameters for the GCN model on the Cora dataset. (NodeTask)
python bench.py --pre_train_model_path './Experiment/pre_trained_model/Cora/GraphCL.GCN.128hidden_dim.pth' --task NodeTask --dataset_name 'Cora' --gnn_type 'GCN' --prompt_type 'GPF-plus' --shot_num 1 --hid_dim 128 --num_layer 2 --seed 42 --device 0
Table of The Following Contents
Supportive graph prompt approaches currently (keep updating):
- [All in One] X. Sun, H. Cheng, J. Li, B. Liu, and J. Guan, “All in One: Multi-Task Prompting for Graph Neural Networks,” KDD, 2023
- [GPF Plus] T. Fang, Y. Zhang, Y. Yang, C. Wang, and L. Chen, “Universal Prompt Tuning for Graph Neural Networks,” NeurIPS, 2023.
- [GraphPrompt] Liu Z, Yu X, Fang Y, et al. Graphprompt: Unifying pre-training and downstream tasks for graph neural networks. The Web Conference, 2023.
- [GPPT] M. Sun, K. Zhou, X. He, Y. Wang, and X. Wang, “GPPT: Graph Pre-Training and Prompt Tuning to Generalize Graph Neural Networks,” KDD, 2022
- [GPF] T. Fang, Y. Zhang, Y. Yang, and C. Wang, “Prompt tuning for graph neural networks,” arXiv preprint, 2022.
Supportive graph pre-training strategies currently (keep updating):
- For node-level, we consider
DGI
andGraphMAE
, whereDGI
maximizes the mutual information between node and graph representations for informative embeddings andGraphMAE
learns deep node representations by reconstructing masked features. - For edge-level, we introduce
EdgePreGPPT
andEdgePreGprompt
, whereEdgePreGPPT
calculates the dot product as the link probability of node pairs andEdgePreGprompt
samples triplets from label-free graphs to increase the similarity between the contextual subgraphs of linked pairs while decreasing the similarity of unlinked pairs. - For graph-level, we involve
GraphCL
,SimGRACE
, whereGraphCL
maximizes agreement between different graph augmentations to leverage structural information andSimGRACE
tries to perturb the graph model parameter spaces and narrow down the gap between different perturbations for the same graph.
Supportive graph backbone models currently (keep updating):
- Graph Convolutional Network (GCN), GraphSAGE, GAT, and Graph Transformer (GT).
Beyond the above graph backbones, you can also seamlessly integrate nearly all graph models implemented by PyG.
**Click [here] to see more details information on these graph prompts, pre-training strategies, and graph backbones. **
We have designed four pre_trained class (Edgepred_GPPT, Edgepred_Gprompt, GraphCL, SimGRACE), which is in ProG.pretrain module, you can pre_train the model by running pre_train.py
and setting the parameters you want.
Or just unzip to get our dataset pretrained model which is already pre-trained.
unzip Experiment.zip
import prompt_graph as ProG
from ProG.pretrain import Edgepred_GPPT, Edgepred_Gprompt, GraphCL, SimGRACE, NodePrePrompt, GraphPrePrompt, DGI, GraphMAE
from ProG.utils import seed_everything
from ProG.utils import mkdir, get_args
from ProG.data import load4node,load4graph
args = get_args()
seed_everything(args.seed)
if args.task == 'SimGRACE':
pt = SimGRACE(dataset_name = args.dataset_name, gnn_type = args.gnn_type, hid_dim = args.hid_dim, gln = args.num_layer, num_epoch=args.epochs, device=args.device)
if args.task == 'GraphCL':
pt = GraphCL(dataset_name = args.dataset_name, gnn_type = args.gnn_type, hid_dim = args.hid_dim, gln = args.num_layer, num_epoch=args.epochs, device=args.device)
if args.task == 'Edgepred_GPPT':
pt = Edgepred_GPPT(dataset_name = args.dataset_name, gnn_type = args.gnn_type, hid_dim = args.hid_dim, gln = args.num_layer, num_epoch=args.epochs, device=args.device)
if args.task == 'Edgepred_Gprompt':
pt = Edgepred_Gprompt(dataset_name = args.dataset_name, gnn_type = args.gnn_type, hid_dim = args.hid_dim, gln = args.num_layer, num_epoch=args.epochs, device=args.device)
if args.task == 'DGI':
pt = DGI(dataset_name = args.dataset_name, gnn_type = args.gnn_type, hid_dim = args.hid_dim, gln = args.num_layer, num_epoch=args.epochs, device=args.device)
if args.task == 'NodeMultiGprompt':
nonlinearity = 'prelu'
pt = NodePrePrompt(args.dataset_name, args.hid_dim, nonlinearity, 0.9, 0.9, 0.1, 0.001, 1, 0.3, args.device)
if args.task == 'GraphMultiGprompt':
nonlinearity = 'prelu'
pt = GraphPrePrompt(graph_list, input_dim, out_dim, args.dataset_name, args.hid_dim, nonlinearity,0.9,0.9,0.1,1,0.3, 0.1, args.device)
if args.task == 'GraphMAE':
pt = GraphMAE(dataset_name = args.dataset_name, gnn_type = args.gnn_type, hid_dim = args.hid_dim, gln = args.num_layer, num_epoch=args.epochs, device=args.device,
mask_rate=0.75, drop_edge_rate=0.0, replace_rate=0.1, loss_fn='sce', alpha_l=2)
pt.pretrain()
Before we do the downstream task, we need to load the nessary data. For some specific prompt, we need to choose function load_induced_graph to the input of our tasker
def load_induced_graph(dataset_name, data, device):
folder_path = './Experiment/induced_graph/' + dataset_name
if not os.path.exists(folder_path):
os.makedirs(folder_path)
file_path = folder_path + '/induced_graph_min100_max300.pkl'
if os.path.exists(file_path):
with open(file_path, 'rb') as f:
print('loading induced graph...')
graphs_list = pickle.load(f)
print('Done!!!')
else:
print('Begin split_induced_graphs.')
split_induced_graphs(data, folder_path, device, smallest_size=100, largest_size=300)
with open(file_path, 'rb') as f:
graphs_list = pickle.load(f)
graphs_list = [graph.to(device) for graph in graphs_list]
return graphs_list
args = get_args()
seed_everything(args.seed)
print('dataset_name', args.dataset_name)
if args.task == 'NodeTask':
data, input_dim, output_dim = load4node(args.dataset_name)
data = data.to(args.device)
if args.prompt_type in ['Gprompt', 'All-in-one', 'GPF', 'GPF-plus']:
graphs_list = load_induced_graph(args.dataset_name, data, args.device)
else:
graphs_list = None
if args.task == 'GraphTask':
input_dim, output_dim, dataset = load4graph(args.dataset_name)
In downstreamtask.py
, we designed two tasks (Node Classification, Graph Classification). Here are some examples.
import prompt_graph as ProG
from ProG.tasker import NodeTask, LinkTask, GraphTask
if args.task == 'GraphTask':
input_dim, output_dim, dataset = load4graph(args.dataset_name)
if args.task == 'NodeTask':
tasker = NodeTask(pre_train_model_path = args.pre_train_model_path,
dataset_name = args.dataset_name, num_layer = args.num_layer,
gnn_type = args.gnn_type, hid_dim = args.hid_dim, prompt_type = args.prompt_type,
epochs = args.epochs, shot_num = args.shot_num, device=args.device, lr = args.lr, wd = args.decay,
batch_size = args.batch_size, data = data, input_dim = input_dim, output_dim = output_dim, graphs_list = graphs_list)
if args.task == 'GraphTask':
tasker = GraphTask(pre_train_model_path = args.pre_train_model_path,
dataset_name = args.dataset_name, num_layer = args.num_layer, gnn_type = args.gnn_type, hid_dim = args.hid_dim, prompt_type = args.prompt_type, epochs = args.epochs,
shot_num = args.shot_num, device=args.device, lr = args.lr, wd = args.decay,
batch_size = args.batch_size, dataset = dataset, input_dim = input_dim, output_dim = output_dim)
_, test_acc, std_test_acc, f1, std_f1, roc, std_roc, _, _= tasker.run()
Kindly note that the comparison takes the same pre-trained pth. The absolute value of performance won't mean much because the final results may vary depending on different pre-training states.It would be more interesting to see the relative performance with other pre-training paradigms.
In our bench
Dataset | Graphs | Avg.nodes | Avg.edges | Features | Node classes | Task (N / G) | Category |
---|---|---|---|---|---|---|---|
Cora | 1 | 2,708 | 5,429 | 1,433 | 7 | N | Homophilic |
Pubmed | 1 | 19,717 | 88,648 | 500 | 3 | N | Homophilic |
CiteSeer | 1 | 3,327 | 9,104 | 3,703 | 6 | N | Homophilic |
Actor | 1 | 7600 | 30019 | 932 | 5 | N | Heterophilic |
Wisconsin | 1 | 251 | 515 | 1703 | 5 | N | Heterophilic |
Texas | 1 | 183 | 325 | 1703 | 5 | N | Heterophilic |
ogbn-arxiv | 1 | 169,343 | 1,166,243 | 128 | 40 | N | Homophilic & Large scale |
Dataset | Graphs | Avg.nodes | Avg.edges | Features | Graph classes | Task (N / G) | Domain |
---|---|---|---|---|---|---|---|
MUTAG | 188 | 17.9 | 19.8 | 7 | 2 | G | small molecule |
IMDB-BINARY | 1000 | 19.8 | 96.53 | 0 | 2 | G | social network |
COLLAP | 5000 | 74.5 | 2457.8 | 0 | 3 | G | social network |
PROTEINS | 1,113 | 39.1 | 72.8 | 3 | 2 | G | proteins |
ENZYMES | 600 | 32.6 | 62.1 | 18 | 6 | G | proteins |
DD | 1,178 | 284.1 | 715.7 | 89 | 2 | G | proteins |
COX2 | 467 | 41.2 | 43.5 | 3 | 2 | G | small molecule |
BZR | 405 | 35.8 | 38.4 | 3 | 2 | G | small molecule |
Note Current experimental datasets: Node/Edge:Cora/Citeseer/Pubmed; Graph:MUTAG
- Write a comprehensive usage document(refer to pyG)
- Write a tutorial, and polish data code, to make our readers feel more easily to deal with their own data. That is to: (1) provide a demo/tutorial to let our readers know how to deal with data; (2) polish data code, making it more robust, reliable, and readable.
- Pre_train: implementation of InfoGraph, contextpred, AttrMasking, ContextPred, GraphLoG, JOAO
- Add Prompt: prodigy (NeurIPS'2023 Spotlight)
- induced graph(1.better way to generate induced graph/2.simplify the 3 type of generate-func)
- support deep GNN layers by adding the feature DeepGCNLayer
Thanks! / 谢谢! / ありがとう! / merci! / 감사! / Danke! / спасибо! / gracias! ...
@inproceedings{sun2023all,
title={All in One: Multi-Task Prompting for Graph Neural Networks},
author={Sun, Xiangguo and Cheng, Hong and Li, Jia and Liu, Bo and Guan, Jihong},
booktitle={Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery \& data mining (KDD'23)},
year={2023},
pages = {2120–2131},
location = {Long Beach, CA, USA},
isbn = {9798400701030},
url = {https://doi.org/10.1145/3580305.3599256},
doi = {10.1145/3580305.3599256}
}
@article{zi2024prog,
title={ProG: A Graph Prompt Learning Benchmark},
author={Chenyi Zi and Haihong Zhao and Xiangguo Sun and Yiqing Lin and Hong Cheng and Jia Li},
year={2024},
journal = {arXiv:2406.05346},
eprint={2406.05346},
archivePrefix={arXiv}
}
@article{sun2023graph,
title = {Graph Prompt Learning: A Comprehensive Survey and Beyond},
author = {Sun, Xiangguo and Zhang, Jiawen and Wu, Xixi and Cheng, Hong and Xiong, Yun and Li, Jia},
year = {2023},
journal = {arXiv:2311.16534},
eprint = {2311.16534},
archiveprefix = {arxiv}
}
@inproceedings{zhao2024all,
title={All in One and One for All: A Simple yet Effective Method towards Cross-domain Graph Pretraining},
author={Haihong Zhao and Aochuan Chen and Xiangguo Sun and Hong Cheng and Jia Li},
year={2024},
booktitle={Proceedings of the 27th ACM SIGKDD international conference on knowledge discovery \& data mining (KDD'24)}
}
@inproceedings{gao2024protein,
title={Protein Multimer Structure Prediction via {PPI}-guided Prompt Learning},
author={Ziqi Gao and Xiangguo Sun and Zijing Liu and Yu Li and Hong Cheng and Jia Li},
booktitle={The Twelfth International Conference on Learning Representations (ICLR)},
year={2024},
url={https://openreview.net/forum?id=OHpvivXrQr}
}
@article{chen2024prompt,
title={Prompt Learning on Temporal Interaction Graphs},
author={Xi Chen and Siwei Zhang and Yun Xiong and Xixi Wu and Jiawei Zhang and Xiangguo Sun and Yao Zhang and Yinglong Zhao and Yulin Kang},
year={2024},
eprint={2402.06326},
archivePrefix={arXiv},
journal = {arXiv:2402.06326}
}
@article{li2024survey,
title={A Survey of Graph Meets Large Language Model: Progress and Future Directions},
author={Yuhan Li and Zhixun Li and Peisong Wang and Jia Li and Xiangguo Sun and Hong Cheng and Jeffrey Xu Yu},
year={2024},
eprint={2311.12399},
archivePrefix={arXiv},
journal = {arXiv:2311.12399}
}
@article{wang2024ddiprompt,
title={Advanced Drug Interaction Event Prediction},
author={Wang, Yingying and Xiong, Yun and Wu, Xixi and Sun, Xiangguo and Zhang, Jiawei},
journal={arXiv preprint arXiv:2402.11472},
year={2024}
}
Media Reports
- 香港首位學者獲ACM頒最佳研究論文獎, 香港新聞網, 2023-09-20 15:21
- 内地及香港首次!港中大的他们获得这项国际大奖!,香港中文大学官方公众号, 2023-09-11 21:30
- Two CUHK scholars receive Best Paper Award from ACM SIGKDD Conference 2023, CUHK Focus
- Prof. Cheng Hong and her postdoc fellow Dr. Sun Xiangguo won the best paper award at KDD2023, CUHK SEEM
- 港科夜闻|香港科大(广州)熊辉教授、李佳教授分别荣获 ACM SIGKDD2023 服务奖与最佳论文奖(研究)
- 数据科学与分析学域李佳教授荣获SIGKDD2023最佳论文奖(研究)!
- 实时追踪科研动态丨姚期智、Quoc Viet Le等人8.9精选新论文,附ChatPaper综述
- KDD 2023奖项出炉:港中文、港科大等获最佳论文奖,GNN大牛Leskovec获创新奖
- 多篇GNN论文获KDD 2023大奖, 图神经网络与推荐系统 2023-08-09 16:03
- 港科广数据科学与分析学域李佳教授荣获SIGKDD2023最佳论文奖(研究)!
Online Discussion
- LOGS第2023/08/12期||KDD 2023 Best Paper Winner 孙相国 :提示学习在图神经网络中的探索
- Talk预告 | KDD'23 Best Paper 港中文孙相国:All in One - 提示学习在图神经网络中的探索
- All in One Multi-Task Prompting for Graph Neural Networks 论文解读
- kdd2023最佳论文
- All in One: Multi-task Prompting for Graph Neural Networks(KDD 2023 Best Paper
- 怎么评价KDD23的best paper? - 知乎
Other research papers released by us
- 最新图大模型综述:由港科广、港中文、清华联合发布,详述使用大模型处理图任务的进展与挑战
- 大模型和图如何结合?最新《图遇见大型语言模型》综述,详述最新进展
- 香港中文领衔港科广、复旦重磅发布:迈向通用图智能的新方法,图提示学习进展与挑战
- 香港中文领衔港科广、复旦重磅发布:迈向通用图智能的新方法,图提示学习进展与挑战
- 图上如何提示?港中文等最新《图提示学习》全面综述,详述图提示分类体系
Once you are invited as a contributor, you would be asked to follow the following steps:
- step 1. create a temp branch (e.g.
xgTemp
) from themain
branch (latest branch). - step 2. fetch
origin/xgTemp
to your localxgTemp
, and make your own changes via PyCharm etc. - step 3. push your changes from local
xgTemp
to your github cloud branch:origin/xgTemp
. - step 4. open a pull request to merge from your branch to
main
.
When you finish all these jobs. I will get a notification and approve merging your branch to main
.
Once I finish, I will delete your branch, and next time you will repeat the above jobs.
A widely tested main
branch will then be merged to the stable
branch and a new version will be released based on stable
branch.