Graph Convolutional Matrix Completion for Bipartite Edge Prediction - Yuexin Wu, Hanxiao Liu and Yiming Yang KDIR 2018. (Best Student Paper)

The framework tries to solve the biparite edge prediction (BEP) problem via decomposing the middle edges as initial node vectors which are later passed through graph convolution neural networks for better hidden representations to the final prediction.

Python == 2.7, lasagne == 0.1, theano == 0.8, numpy, scipy, joblib, scikit-learn, nose-parameterized

If you have Anaconda, you can install the dependencies using the following command: conda create --name gcmc --file requirement.txt

Uncompress data using 7z x data.7z.

If you have installed cuda support, add THEANO_FLAGS='device=cuda0,floatX=float32' before executing the commnad. In experiments, multi-core cpus are more preferable (with --n_jobs option).

python src/run.py --dataset cmu --layers 80 70 5 --iters 1000 --pos_up_ratio 10. --fold 1 --save_log --save_model --binary_graph

This will create logs and model files in save folder. The final result is saved in gcmc.log.

See python src/run.py --help for more option explanations.

NOTICE: This may take LONG time. Consider running single experiments first to estimate time.

General --iters 1000 should be enough for all datasets. In our experiment, the exact parameters can be found in run_all.sh.

Execute the following command to run on all datasets.

bash run_all.sh

If you find the repository useful for your publication, please consider citing our paper:

@conference{wu2018graph,
author={Yuexin Wu and Hanxiao Liu and Yiming Yang},
title={Graph Convolutional Matrix Completion for Bipartite Edge Prediction},
booktitle={Proceedings of the 10th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management - Volume 1: KDIR,},
year={2018},
pages={51-60},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0006900000510060},
isbn={978-989-758-330-8},
}

For any questions and comments, please send your email to [email protected].