Simple Multigraph Convolution Networks (SMGCN)

Pytorch implementation of the paper [WWW'2024] "Simple Multigraph Convolution Networks"

TL;DR

Previous methods for multigraph rather lacks cross-view interaction or are too inefficient to be used in practice. We propose a simple and efficient multigraph convolutional networks based on both edge-level and subgraph-level credible graph extraction from multigraph. We show that our method outperforms previous methods on several multigraph datasets and is more parameter efficient and theoretically sound.

Usage

Environment

We use torch_geometric and torch as the main framrwork to conduct our experiment. Mamba is a variant of Conda, which is rewrited in C++. It is also okay to use just conda.

mamba create -n smgcn python==3.10
mamba install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
mamba install pyg -c pyg
mamba install scikit-learn numpy pandas

Data preparation

We use the standard torch_geometric object to represent the multigraph. Run the data.py to download and preprocess the datasets.

cd src/DBLP
python data.py

Take SMGCN (k=2) as an example, simply run the following command to train the model.

python SMGCNk=2.py

Parameters Comparison