Paper: Factorizable Graph Convolutional Networks, NeurIPS'20
See requirment file for more information about how to install the dependencies.
We provide here datasets that are ready to use within this project. Download the datasets and unzip it into ./data dir. These datasets can also be either downloaded from their official websites or generated on the fly.
We use DGL to implement all the GCN models. In order to evaluate the disentanglement performance of GAT model, you need to modify the last line of
dgl/nn/pytorch/conv/gatconv.py
from
return rst
to
return rst, graph.local_var(), graph.edata['a']
The train_*.sh scripts contains the training codes for corresponding datasets and methods.
train_synth.sh for Synthetic dataset;
train_zinc.sh for ZINC dataset;
train_gin.sh for IMDB-B, COLLAB, and MUTAG datasets.
train_pattern.sh for pattern datasets.
The model as well as the training log will be saved to the corresponding dir in ./data for evaluation.
The evaluation dir contains the codes and examples for evaluating the performance of both the task-related performance and the disentanglement performance.
generate_report.get_acc_report reports the accuracy on the Synthetic dataset;
generate_report.get_mae_report reports the MAE on the ZINC dataset;
generate_report.get_10fold_curve_report report the 10-fold cross-validation performance on IMDB-B, COLLAB, and MUTAG datasets.
The performances of disentanglement are evaluated
as 
and C-score.
ged_eval_synth generates the disentanglement performance on the Synthetic dataset;
ged_eval_zinc generates the disentanglement performance on the ZINC dataset.
@article{yang2020factorizable,
title={Factorizable Graph Convolutional Networks},
author={Yang, Yiding and Feng, Zunlei and Song, Mingli and Wang, Xinchao},
journal={Advances in Neural Information Processing Systems},
volume={33},
year={2020}
}
FactorGCN is released under the MIT license. Please see the LICENSE file for more information.