graphgan

Graph Generative Adversarial Nets (GraphGAN)

• Paper link: https://arxiv.org/pdf/1711.08267.pdf
• Author's code repo: https://github.com/hwwang55/GraphGAN. Note that the original code is implemented with Tensorflow for the paper.

Dataset Statics

Dataset	# Nodes	# Edges
arXiv-GrQc	5242	14496

Files Description

• gan_datasets : folds to save training data

  • CA-GrQc_train : training data
  • CA-GrQc_test : positive sampling of test data
  • CA-GrQc_test_neg : negative sampling of test data

  The data should be an undirected graph in which node IDs start from 0 to N-1 (N is the number of nodes in the graph). Each line contains two node IDs indicating an edge in the graph.

  txt file sample : 0 1 3 2 ...

  • CA-GrQc_pre_train.emb : pre-trained node embeddings

  Note: the dimension of pre-trained node embeddings should equal parameter 'n_emb'

• gan_cache

  • CA-GrQc.pkl : save constructed BFS-trees

• gan_results

  • CA-GrQc.txt : evaluation results
  • CA-GrQc_best_acc_gen_.emb : embeddings of generator with the best evaluation
  • CA-GrQc_best_acc_dis_.emb : embeddings of discriminator with the best evaluation
  • CA-GrQc_gen_.emb : learned embeddings of generator during training
  • CA-GrQc_dis_.emb : learned embeddings of discriminator during training

• checkpoint : folds to save the model weights

Results

TL_BACKEND="paddle" python graphgan_trainer.py --batch_size_dis 1024 --batch_size_gen 1024 --n_epochs 30 --lr_gen 1e-5 --lr_dis 1e-5
TL_BACKEND="tensorflow" python graphgan_trainer.py --batch_size_dis 1024 --batch_size_gen 1024 --n_epochs 30 --lr_gen 1e-5 --lr_dis 1e-5
TL_BACKEND="torch" python graphgan_trainer.py --batch_size_dis 1024 --batch_size_gen 1024 --n_epochs 30 --lr_gen 1e-5 --lr_dis 1e-5

Dataset	Paper	Our(pd)	Our(tf)	Our(torch)
arXiv-GrQc	0.849	0.8813±0.00069	0.8819±0.00133	0.8820