GHCN

This is the PyTorch implementation of our paper: "Tackling Over-Smoothing: Graph Hollow Convolution Network with Topological Layer Fusion"

Requirement

• pytorch >= 1.8.0

• torch-geometric

• torchmetrics

Note that the versions of PyTorch and PyTorch Geometric should be compatible.

Datsetset

The data folder contains three benchmark datasets (Cora, Citeseer, Pubmed).

Cora, CiteSeer, and PubMed are pre-downloaded from torch_geometric.datasets.

We use the same full-supervised setting as FastGCN.

Example

For Cora dataset, please run:

python main.py --dataset=Cora

For CiteSeer dataset, please run:

python main.py --dataset=CiteSeer --weight_decay=5e-3 --k=20

For PubMed dataset, please run:

python main.py --dataset=CiteSeer --weight_decay=5e-5 --k=9 --dropout=0.5

Baselines

The baseline folder contains 6 SOTA GNN methods as follows：

• Semi-Supervised Classification with Graph Convolutional Networks [code]

python baseline/gcn.py

• Graph Attention Networks [code]

python baseline/gat.py

• Simplifying Graph Convolutional Networks [code]

python baseline/sgc.py

• SIGN: Scalable Inception Graph Neural Networks [code]

python baseline/sign.py

• Towards Deeper Graph Neural Networks [code]

python baseline/dagnn.py

• Simple and Deep Graph Convolutional Networks [code]

python baseline/gcnii.py