Stochastic Training for Graph Convolutional Networks

Code for the paper Stochastic Training for Graph Convolutional Networks. The implementation is based on Thomas Kipf's implementation for graph convolutional networks.

Requirements

• tensorflow (>0.12)
• networkx (<=1.11 for loading GraphSAGE datasets)

Installation

Make sure that you have valid c++11 and cuda compiler.

pip install --upgrade pip
pip install numpy scipy networkx==1.11 tensorflow-gpu scikit-learn
python setup.py install
cd gcn
./build.sh

Data

We support both GCN's data format and GraphSAGE's data format.

To download the Citeseer, Cora, PubMed, PPI and Reddit dataset

mkdir data
cd data
git clone [email protected]:tkipf/gcn.git
mv gcn/gcn/data/* .

wget http://snap.stanford.edu/graphsage/ppi.zip
unzip ppi.zip
mv ppi/* .

wget http://snap.stanford.edu/graphsage/reddit.zip
unzip reddit.zip
mv reddit/* .

Run the demo

We provide recipes for the Citeseer, Cora, PubMed, NELL, PPI, and Reddit dataset in the config folder. Please refer to our paper for the algorithms. Typically, CVD+PP has similar accuracy with Exact, but is faster.

# Exact algorithm
config/cora.config
# NS+PP algorithm
config/cora.config --degree=1 --test_degree=1
# CV+PP algorithm
config/cora.config --cv --test_cv --degree=1 --test_degree=1 
# CVD+PP algorithm
config/cora.config --cv --cvd --test_cv --degree=1 --test_degree=1

Cite

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

@inproceedings{chen2018stochastic,
  title={Stochastic Training of Graph Convolutional Networks with Variance Reduction},
  author={Chen, Jianfei and Zhu, Jun and Song, Le},
  booktitle={International Conference on Machine Learning},
  pages={941--949},
  year={2018}
}