Skip Connections and Training Cost Reduction: A Game Changer in Graph Neural Networks

This repository contains a PyTorch implementation of "Skip Connections and Training Cost Reduction: A Game Changer in Graph Neural Networks".

Dependencies

• CUDA 10.1
• python 3.6.9
• pytorch 1.3.1
• networkx 2.1
• scikit-learn

Datasets

The data folder contains three benchmark datasets(Cora, Citeseer, Pubmed), and the newdata folder contains four datasets(Chameleon, Cornell, Texas, Wisconsin) from Geom-GCN. We use the full-supervised setting as GCN.

Results

Full supervised experimental results for mean classification accuracy and training cost for node classification Improved results are shown in bold. (A full comparison is shown in the paper.)

Dataset	Training Cost(sec)	Accuracy	Dataset	Training Cost(sec)	Accuracy
Cora	480.00	88.03	Cham	79.6865	67.43
Cite	269.95	76.99	Corn	185.32	74.59
Pubm	3168.37	67.43	Texa	369.5642	79.46
Wisc	104.89	77.06

Usage

To replicate the full-supervised results, run the following scripts:

python -u full-supervised.py --data cora --layer 64 --alpha 0.2 --weight_decay 1e-4
python -u full-supervised.py --data citeseer --layer 64 --weight_decay 5e-6
python -u full-supervised.py --data pubmed --layer 64 --alpha 0.1 --weight_decay 5e-6
python -u full-supervised.py --data chameleon --layer 8 --lamda 1.5 --alpha 0.2 --weight_decay 5e-4
python -u full-supervised.py --data cornell --layer 16 --lamda 1 --weight_decay 1e-3
python -u full-supervised.py --data texas --layer 32 --lamda 1.5 --weight_decay 1e-4
python -u full-supervised.py --data wisconsin --layer 16 --lamda 1 --weight_decay 5e-4