mhadnanali/febaa

Features Based Adaptive Augmentation for Graph Contrastive Learning

FebAA

Features Based Adaptive Augmentation for Graph Contrastive Learning (https://www.sciencedirect.com/science/article/pii/S1051200423004074)

BGRL+FebAA

BGRL+FebAA can be executed using BGRL official code available at the github repository, files uploaded in BGRL+FebAA folder, needed to be placed in relevant folders given in BGRL code . The main script file is BGRL_FebAA.py used for training on the transductive task datasets and configuration files can be found in ./config folder.

Regenerate Training Results:

To run BGRL on a dataset from the transductive setting, use BGRL_FebAA.py and one of the configuration files that can be found in config/. For example, to train on the wiki-cs dataset, use the following command:

python BGRL_FebAA.py --flagfile=config/*-wiki-cs_FeBAA.cfg

Above same command can be used to regenerate the results as seeds values are given in .cfg files, while * will be replaced with inf or rand.

Verify Test Results:

The runs folder contains log files, Get_Results.py can be executed to get the results from log files. Note that our reported results are based on an average of 20 runs. Test accuracies under linear evaluation are reported on TensorBoard. To start the tensorboard server run the following command:

tensorboard --logdir runs

	WikiCS	Amazon Computers	Amazon Photos	CoAuthorCS	CoAuthorPhy
Inf	80.59±0.58	91.07±0.20	93.74±0.19	93.55±0.14	95.90±0.08
Rand	80.57±0.50	90.94±0.23	93.80±0.23	93.58±0.13	95.90±0.09

GRACE+FebAA

Grace+FebAA is implemented using PyGCL. To execute the codes, one need to install PyGCL and place the augmentors folder files from this repository to PyGCL augmentors folder while GRACE+FebAA.py and Features folder should be kept in examples folder of PyGCL folder. Then execute the below command to get results. Make sure to enter relevent seeds given in the last table, dataset in the code (if you are trying to recreate our results).

python GRACE+FebAA.py

We intentionally did not create any configration ( .cfg or .yaml ) file for input to keep it same as PyGCL.

	Cora	CiteSeer	Actor
Inf	87.30±1.12	76.26±1.46	30.58±1.06
Rand	87.48±0.50	75.36±1.29	30.35±1.09

Hyper-parameters to Recreate Results

Below table contains the hyper-paramter values to recreate the results, where 1 & 2 indicates the graph view 1 and graph view 2.

Dataset	Edge Drop Prob. 1 & 2	Feature Ratio 1 & 2	Feature Drop Prob. 1 & 2	manual_seed	random.seed	Least or Most
Cora	0.4 & 0.2	100% & 80%	0.4 & 0.375	125656	896146	Least
CiteSeer	0.4 & 0.2	100% & 70%	0.4 & 0.43	553358	559648	Most
Actor	0.3 & 0.3	100% & 30%	0.3 & 1	8833511	7396411	Most

febaa's Issues

Some questions about training

Thank you for your wonderful work. I encountered a problem. When running multiple experiments for report, do we fix the random seed and run it again, or change the random seed and run it again? And for the results, taking bgrl+FebAA for example, is it the evaluation result from the pre training to the last round or the best evaluation result in the 10000 rounds? I look forward to your answers, thank you!

Grace+FebAA

麻烦请问可以提供Grace+FebAA的全部代码和运行命令吗，谢谢