• Python, Matlab
• Python package: tensorflow=1.1
• Other version of tensorflow might work, but it might disable the gpu.

1. cd MRCGAN_network/amg-master
2. matlab -nodisplay -nosplash -nodesktop -r mexall
3. conda env create -f environment.yml
4. conda activate miscgan

1. Training: python main_network.py --demo --stage training --gpu

2. Testing and evaluation: python main_network.py --demo --stage testing --gpu

• --demo: Demo
• --dataset_A: the path of the graph in domain A. Domain A usually refers to the original graph.
• --dataset_B: the path of the graph in domain B. Domain B usually refers to the synthetic graph.
• --epoch: the number of epochs
• --Starting_layer: Considering the memory limitation, we start the training process at the second or third layer to reduce the run time and memory. The value should be in the range of 1-4.
• --stage: training stage or testing stage
• --iter: Iterations of residule_block function for generator in the training stage.
• --gpu: Use GPU to train the model or not. The default value is False.
• --clusters: the number of clusters. Here, we cluster the nodes in the graph into 2 group
• --which_direction: domain A to domain B or domain B to domain A.
• --kernel_number: number of kernels in the initial convolutional neural network

There are two evaluation methods shown in the testing stage.

1. Plot the generated graph.
2. KL divergence of graph degree.

The final results will be displayed in the terminal and the graphs are saved at './graph'. The performance increases as we set the parameter 'starting_layer' to be 1 or 2. However, when we set the parameter 'starting_layer' to be 3 or 4, the runtime reduces rapidly. There is a trade-off between performance and the runtime.

@article{zhou2019misc,
title={Misc-GAN: A Multi-scale Generative Model for Graphs},
author={Zhou, Dawei and Zheng, Lecheng and Xu, Jiejun and He, Jingrui},
journal={Frontiers in Big Data},
volume={2},
pages={3},
year={2019},
publisher={Frontiers}
}