Simple Message Passing for Link Prediction (SMPLP)

Support five message passing and two with JKnet: SAGE, LRGA, GCN, SGC, TAGC, SGC+JKnet, SAGE+JKnet

SAGE LRGA GCN SGC TAGC JKnet

Support five link prediction methods: MLP/Dot/Bilinear Dot/MLP Cat/MLP Bilinear

Environment:

python==3.8
torch==1.10.1+cu102
torch-geometric==2.0.4
ogb==1.3.4

The dataset ogbl-vessel can be download and placed in ./dataset/ogbl_vessel/.

The dataset ogbl-collab can be download and placed in ./dataset/ogbl-collab/.

Performance on ogbl-vessel (10 runs):

Methods	Test Acc	Valid Acc	Hardware
TAGConv (1-layers)	51.19 ± 1.74	51.20 ± 1.75	Tesla A100(80GB)
LRGA (1-layers)	54.15 ± 4.37	54.18 ± 4.39	Tesla A100(80GB)
SGC (3-layers)	54.31 ± 23.79	54.33 ± 23.89	Tesla V100(32GB)
SGC (3-layers w/o normalize)	50.09 ± 0.11	50.10 ± 0.11	Tesla V100(32GB)

python gnn.py --encoder='sgc' --decoder='mlp' --hidden_channels=16 --device=1  --lr=1e-6  --num_layers=3   --epochs=80

Performance on ogbl-collab (10 runs):

Methods	Test Hits@50	Valid Hits@50
LRGA (1-layers)	0.6909 ± 0.0055	1.0000 ± 0.0000