flyingtango / digcn

Implement of DiGCN, NeurIPS-2020

Digraph Inception Convolutional Networks

Paper | Poster | Supplementary

This repository is the official PyTorch implementation of Digraph Inception Convolutional Networks, where we make GCNs available in digraphs (directed graphs) and propose an Inception network to learn multi-scale features in digraphs.

Illustration of digraph conv

 

Illustration of DiGCN model

 

If you find our work useful, please considering citing

@article{tong2020digraph,
  title={Digraph Inception Convolutional Networks},
  author={Tong, Zekun and Liang, Yuxuan and Sun, Changsheng and Li, Xinke and Rosenblum, David and Lim, Andrew},
  journal={Advances in Neural Information Processing Systems},
  volume={33},
  year={2020}
}

Requirements

Our project is developed using Python 3.7, PyTorch 1.5.0 with CUDA10.2. We recommend you to use anaconda for dependency configuration.

First create an anaconda environment called DiGCN by

conda create -n DiGCN python=3.7
conda activate DiGCN

Then, you need to install torch manually to fit in with your server environment (e.g. CUDA version). For the torch and torchvision used in my project, run

conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.2 -c pytorch

Besides, torch-scatter and torch-sparse are required for dealing with sparse graph. For these two packages, please follow their official instruction torch-scatter and torch-sparse.

Other requirements can be set up through:

cd DiGCN
pip install -e .

Run

cd code
python gcn.py --gpu-no 0 --dataset cora_ml
python digcn.py --gpu-no 0 --dataset cora_ml
python digcn_ib.py --gpu-no 0 --dataset cora_ml

Results

License

DiGCN is released under the MIT License. See the LICENSE file for more details.

Acknowledgements

The template is borrowed from Pytorch-Geometric benchmark suite. We thank the authors of following works for opening source their excellent codes. Pytorch-Geometric,Graph2Gauss,GNN-benchmark

digcn's Issues

Can be applied to large-scale graph?

I think that this work cannot be applied to large-scale graphs for the reason that calculating the adj through your method needs eigen decomposition and to_dense() method needs large memory available.
eig_value, left_vector = scipy.linalg.eig(p_ppr.numpy(),left=True,right=False) p_dense = torch.sparse.FloatTensor(edge_index, p, torch.Size([num_nodes,num_nodes])).to_dense()

Is this work extendable using other graph datasets like Pubmed or Reddit-binary?

Dear author,

I have collected my graphs with different format like PubMed/ Reddit-Binary. if your work is extendable, may I know which parts of code I need to change? Thank you in advance,

Differences in results between DiGCN and DGCN

Hi,

I was wondering, what is the large discrepancy in accuracy for benchmarking methods (like GCN/GAT, for example) between Table 1 in the DGCN paper and Table 2 in the DiGCN paper? The accuracy levels seem very different for the same dataset.

Question about equation 4

Dear author,
I'm a little confused about Eq.4. You adopt symmetric normalized directed Laplacian proposed by Chung F (2005), however, It is not clear to me why the graph filter in Eq 4 can reflect directed property? Since it is a symmetric matrix.

您好！请问您在链路预测的时候是如何对有向图进行链路预测的呢？

研一小菜鸡被老板要求做有向图的链路预测，像GAE那样两个节点的嵌入做内积的话还原的是无向图，没什么思路很苦恼，正好看见了您的文章，所以就想问一下您是怎么处理有向图的链路预测的？