danielzuegner / netgan Goto Github PK

View Code? Open in Web Editor NEW

188.0 6.0 66.0 48.64 MB

Implementation of the paper "NetGAN: Generating Graphs via Random Walks".

Home Page: https://www.kdd.in.tum.de/netgan

License: MIT License

Python 10.21% Jupyter Notebook 89.79%

machine-learning graph-mining generative-adversarial-network deep-learning

netgan's Introduction

NetGAN: Generating Graphs via Random Walks

Implementation of the method proposed in the paper:
NetGAN: Generating Graphs via Random Walks

by Aleksandar Bojchevski, Oleksandr Shchur, Daniel Zügner, Stephan Günnemann
Published at ICML 2018 in Stockholm, Sweden.

This implementation is written in Python 3.6 and uses Tensorflow 1.4.1.

Requirements

Install the reqirements via
pip install -r requirements.txt

Note that the modules powerlaw and python-igraph are only needed to compute the graph statistics. If you only want to run NetGAN, feel free to comment out the respective parts of the code.

Run the code

To try our code, the best way to do so is to use the IPython notebook demo.ipynb

Pre-trained models used in the paper

Run graph_generation_pretrained.ipynb and link_prediction_pretrained.ipynb to try our pre-trained models on Cora-ML.

Latent variable interpolation

Run latent_interpolation.ipynb to run latent variable interpolation experiments as in the paper.

Installation

To install the package, run python setup.py install.

Citation

Please cite our paper if you use the model or this code in your own work:

@inproceedings{DBLP:conf/icml/BojchevskiSZG18,
  author    = {Aleksandar Bojchevski and
               Oleksandr Shchur and
               Daniel Z{\"{u}}gner and
               Stephan G{\"{u}}nnemann},
  title     = {NetGAN: Generating Graphs via Random Walks},
  booktitle = {Proceedings of the 35th International Conference on Machine Learning,
               {ICML} 2018, Stockholmsm{\"{a}}ssan, Stockholm, Sweden, July
               10-15, 2018},
  pages     = {609--618},
  year      = {2018},
}

References

Cora dataset

In the data folder you can find the Cora-ML dataset. The raw data was originally published by

McCallum, Andrew Kachites, Nigam, Kamal, Rennie, Jason, and Seymore, Kristie. "Automating the construction of internet portals with machine learning." Information Retrieval, 3(2):127–163, 2000.

and the graph was extracted by

Bojchevski, Aleksandar, and Stephan Günnemann. "Deep gaussian embedding of attributed graphs: Unsupervised inductive learning via ranking." ICLR 2018.

Contact

Please contact [email protected] in case you have any questions.

netgan's People

Contributors

Stargazers

Watchers

netgan's Issues

Covariates + Single Graph

I have a custom graph with some covariates, how can I train on these to compare to ERGM? Particularly, what would you recommend for using covariates for netGAN generation?

What is the best way to supply my own single graph to this pipeline with a set of covariates?

ERROR: Object arrays cannot be loaded when allow_pickle=False (+Fix)

Hi, I tried running demo.ipynb but got the following error when running the 2nd cell:

ERROR: Object arrays cannot be loaded when allow_pickle=False

I made a small fix within netgan/Utils.py.
If you want I can send a PR request for this.

Best,

Nick

Graph generation for custom dataset

I want to take a random graph dataset and convert it into the format used in the repository. But I am unable to figure it out how.

Clarification on the experiment procedure on DBLP dataset

In the paper, it was mentioned that the model was tested against the DBLP dataset. Can I clarify how was this done given that the other datasets were homogeneous but the DBLP dataset is heterogeneous.
Thanks!

How to train the ERGM model mentioned in your paper?

Hi, thanks for your excellent work on graph generation.

I have some problems during the implementation of the baseline method ERGM in your paper. I followed your parameter settings in R, but the simulation results are always bad (such as the explosion of the number of edges).

This is my R code for graph simulation by ERGM:
result.network <- simulate( ergm( input.network ~ edges + degcor + degree1.5 + gwesp + density, control=control.ergm( MCMLE.maxit=20, MCMLE.density.guard=100000, parallel=2, parallel.type="PSOCK" )) )
where 'input.network' is the graph to be simulated, and the function 'simulate' is from package 'ergm'.
I wonder how to have the right hyper-parameters to have the best simulation result like your paper's.

Reproduce

I follow this code to train the results of Graph Generation that can't reproduce your paper. Do you have any other settings?