Quick Start

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 3 --epochs 50 --seed 42 --device 5

In downstream_task, you can obtain the experimental results by running the parameters you want:

python downstream_task.py --pre_train_path 'None' --task GraphTask --dataset_name 'MUTAG' --gnn_type 'GCN' --prompt_type 'None' --shot_num 10 --hid_dim 128 --num_layer 3 --epochs 50 --seed 42 --device 5

Pre-train your GNN model

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.

import prompt_graph as ProG
from ProG.pretrain import Edgepred_GPPT, Edgepred_Gprompt, GraphCL, SimGRACE
from ProG.utils import seed_everything
from ProG.utils import mkdir, get_args


args = get_args()
seed_everything(args.seed)
mkdir('./pre_trained_gnn/')

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)

pt.pretrain()

Do the Downstreamtask

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 == 'NodeTask':
    tasker = NodeTask(pre_train_model_path = './pre_trained_gnn/Cora.Edgepred_GPPT.GCN.128hidden_dim.pth', 
                    dataset_name = 'Cora', num_layer = 3, gnn_type = 'GCN', prompt_type = 'GPrompt', epochs = 150, shot_num = 5)
    tasker.run()


if args.task == 'GraphTask':
    tasker = GraphTask(pre_train_model_path = './pre_trained_gnn/MUTAG.SimGRACE.GCN.128hidden_dim.pth', 
                    dataset_name = 'MUTAG', num_layer = 3, gnn_type = 'GCN', prompt_type = 'GPrompt', epochs = 150, shot_num = 5)
    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 training paradigms.

Dataset

Our experiments are conducted on a diverse set of datasets, covering both node classification and graph classification tasks:

Node Classification Datasets

Dataset	Description	Task
Cora, CiteSeer, and PubMed	Citation networks where nodes represent documents, and edges represent citation links. Each node has a feature vector from the document text.	Classify nodes into academic topics.
Flickr	Social network dataset where nodes represent users, and edges represent follower relationships. Node features are based on user activity and metadata.	Classify users into interest groups.
ogbn-arxiv	Citation network of computer science papers from arXiv, part of the Open Graph Benchmark. Nodes represent papers with feature vectors based on content.	Predict the subject area of each paper.

Graph Classification Datasets

Dataset	Description	Task
PROTEINS	Graphs representing proteins, with nodes as secondary structure elements and edges as spatial or sequential proximities.	Classify proteins into categories.
COX2	Molecular graphs where nodes are atoms and edges are chemical bonds.	Predict the biological activity of molecules.
ENZYMES	Graphs of enzyme structures.	Predict the enzyme commission number.
BZR	Molecular graphs.	Classify molecules based on biochemical properties.
MUTAG	Graphs of mutagenic compounds, with nodes as atoms and edges as bonds.	Classify compounds by mutagenicity.
DD	Graphs representing protein structures, with nodes as amino acids and edges as interactions.	Categorize proteins into structural families.
COLLAB	Scientific collaboration graphs, with nodes as researchers and edges as co-authorships.	Classify ego-networks into research fields.

These datasets provide a comprehensive evaluation of our methods across various types of graphs and classification tasks, ensuring robustness and generalizability of the results.