readme.md

Learning to Distill Graph Neural Networks (LTD)

• Paper link: [https://doi.org/10.1145/3539597.3570480]
• Author's code repo: [https://github.com/BUPT-GAMMA/LTD].

Dataset Statics

Raw Data Statics

Dataset	# Nodes	# Edges	# Classes
Cora	2,708	10,556	7
Citeseer	3,327	9,228	6
Pubmed	19,717	88,651	3

Refer to Planetoid.

Largest Connected Component Statics

Dataset	# Nodes	# Edges	# Classes
Cora	2,485	10,138	7
Citeseer	2,120	7,358	6
Pubmed	19,717	88,651	3

Results

# available dataset: "cora", "citeseer", "pubmed" 
TL_BACKEND="torch" python ltd_trainer.py --dataset cora --model GCN
TL_BACKEND="torch" python ltd_trainer.py --dataset cora --model GAT
TL_BACKEND="torch" python ltd_trainer.py --dataset citeseer --model GCN 
TL_BACKEND="torch" python ltd_trainer.py --dataset citeseer --model GAT
TL_BACKEND="torch" python ltd_trainer.py --dataset pubmed --model GCN 
TL_BACKEND="torch" python ltd_trainer.py --dataset pubmed --model GAT

Dataset	GNN	paper(Largest Connected Component)	Our	Our(Largest Connected Component)
cora	GCN	87.21	83.10±0.41	84.71±0.44
cora	GAT	86.56	82.40±1.38	84.35±0.42
citeseer	GCN	78.51	72.39±1.09	74.69±1.12
citeseer	GAT	77.35	72.85±0.71	76.27±0.62
pubmed	GCN	81.91	79.54±0.70	79.54±0.70
pubmed	GAT	82.74	78.77±0.90	78.77±0.90