Code for Calibrating and Improving Graph Contrastive Learning

Environment

The running environment of the repo is listed in requirements.txt.

Reproducing results

GCA

Please first enter the GCA folder,

cd GCA

Then run the following command

python train.py --dataset WikiCS --param local:wikics.json --device cuda:0 --reg --reg-ratio 0.0001 --num_seeds 10
python train.py --dataset Amazon-Photo --param local:amazon_photo.json --device cuda:0 --reg --reg-ratio 0.0001 --num_seeds 10
python train.py --dataset Amazon-Computers --param local:amazon_computers.json --device cuda:0 --reg --reg-ratio 0.01 --num_seeds 10
python train.py --dataset Coauthor-CS --param local:coauthor_cs.json --device cuda:0 --reg --reg-ratio 0.1 --num_seeds 10
python train.py --dataset Cora --param local:cora.json --device cuda:0 --reg --reg-ratio 8.0 --num_seeds 10
python train.py --dataset CiteSeer --param local:citeseer.json --device cuda:0 --reg --reg-ratio 8.0 --num_seeds 10
python train.py --dataset PubMed --param local:pubmed.json --device cuda:0 --reg --reg-ratio 8.0 --num_seeds 10

ML/LC

To replicate the ML/LC results presented in our paper, we apply the GCN encoder on Cora, Citeseer, Pubmed, Wiki, Computers, and Photo, setting the number of layer to 1 and implementing full-batch training. The Contrast-Reg ratio is set to 1.

For Cora (LC), we set a learning rate of 3e-3 and enable curriculum learning. To reproduce the experiments, execute the following command:

python main.py --dataset cora --contrast-model LC --lr 3e-3 --use-curri --curri-round 5 --reg --ratio 1 --bilinear --num-layers-to 1

For Cora (ML), we set the learning rate to 3e-3. To reproduce the experiments, execute the following command:

python main.py --dataset cora --contrast-model ML --lr 3e-3 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 1000

For Citeseer, Pubmed, Computers, Photo, Wiki dataset, run the following commands:

python main.py --dataset citeseer --contrast-model LC --lr 1e-4 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 500 --use-curri --curri-round 3 --pre-norm --final-norm
python main.py --dataset citeseer --contrast-model ML --lr 5e-4 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 300 --pre-norm --final-norm
python main.py --dataset pubmed --contrast-model LC --lr 1e-2 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 300 --final-norm
python main.py --dataset pubmed --contrast-model ML --lr 1e-2 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 1000 --final-norm
python main.py --dataset computers --contrast-model LC --lr 3e-4 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 1000
python main.py --dataset computers --contrast-model ML --lr 1e-4 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 1000 
python main.py --dataset photo --contrast-model LC --lr 3e-4 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 1000
python main.py --dataset photo --contrast-model ML --lr 1e-4 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 1000 
python main.py --dataset wiki --contrast-model LC --lr 1e-3 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 1000
python main.py --dataset wiki --contrast-model ML --lr 1e-4 --reg --ratio 1 --bilinear --num-layers-to 1 --epochs 1000 

For the link prediction task, the necessary codes are readily available in the 'link-prediction' folder. To replicate the LC outcomes for the Cora, Citeseer, Pubmed, and Wiki datasets, execute the commands outlined below:

python link_exete.py --dataset cora --reg --lr 0.003 --epochs 300 --use_curri --curri_round 5 --lr_deduct 0.5 --same_layer
python link_exete.py --dataset citeseer --reg --lr 0.0001 --epochs 500 --use_curri --curri_round 3 --pre_norm --final_norm --lr_deduct 0.5 --same_layer
python link_exete.py --dataset pubmed --reg --lr 0.01 --epochs 300 --final_norm --same_layer
python link_exete.py --dataset wiki --reg --lr 0.001 --epochs 1000 --same_layer

The hyperparameters for link prediction tasks adhere rigorously to the settings derived from node classification, with no additional fine-tuning applied.