Decoupled Graph Convolution (DGC)

Authors: Yifei Wang, Yisen Wang, Jiansheng Yang, Zhouchen Lin

Overview

This repo contains an example implementation of the Decoupled Graph Convolution (DGC) model, described in the NeurIPS 2021 paper Dissecting the Diffusion Process in Linear Graph Convolutional Networks.

DGC, similar to SGC, removes the nonlinearities and collapes the weight matrices in Graph Convolutional Networks (GCNs) and is essentially a linear GCN.

Motivated by the dissection of SGC's limitations from a continuous perspective, DGC further decouples the terminal time T and propagation steps K as two free hyperparameters. In this way, DGC overcomes SGC's limitations and improves SGC by a large margin, making it even comparable to state-of-the-art nonlinear GCNs. Meanwhile, as a linear GCN, DGC is very memory-efficient and saves much training time.

This repo contains the implementation of DGC for citation networks (Cora, Citeseer, and Pubmed) and the performance is shown below. All experiments are conducted with a single NVIDIA GTX 1080ti GPU.

In particular, DGC could complete Cora/Citeseer training with <1s, and training on a large graph (Pubmed) with 900 steps only takes 2.35s.

Dependencies

Our implementation works with PyTorch>=1.0.0 and you can install other dependencies with

$ pip install -r requirement.txt

Usage

We provide the citation network datasets under data/, which corresponds to the public data splits.

To train DGC on the citation networks, simply run main.py with the following commands,

python main.py --dataset <cora, citeseer, or pubmed> --T <T value> --K <K value>

To reproduce the results above, run sh run_citation.sh. It will automatically run 10 trials for each experiment and report the mean and standard deviation.

If you find this repo useful, please cite:

@InProceedings{wang2021dgc,
  title =    {Dissecting the Diffusion Process in Linear Graph Convolutional Networks},
  author =   {Wang, Yifei and Wang, Yisen and Yang, Jiansheng and Lin, Zhouchen},
  booktitle = {Proceedings of the 35th Conference on Neural Information Processing Systems (NeurIPS 2021)},
  year =     {2021}
}

Acknowledgement

This repo borrows a lot from SGC, which is modified from pygcn, and FastGCN.