GNN_vs_GAMLP

The official implementation of ''On Graph Neural Networks versus Graph-Augmented MLPs'' ICLR 2021

Authors: Lei Chen*, Zhengdao Chen*, Joan Bruna.

Dependencies

Core packages:

pytorch 1.5.0
dgl 0.4.3

Scripts

Community detection

Click

• GA-MLP-H (hardness rank 1)

python3.7m main_sbm.py --model swl_gnn_dim10_bh120_mint0_mott0_stack1_idnt1 --verbose --lr 0.001 --n-graphs 6000 --p 4.5 --q 1.5 --n-nodes 250 --n-communities 2

• GA-MLP-H (hardness rank 2)

python3.7m main_sbm.py --model swl_gnn_dim10_bh120_mint0_mott0_stack1_idnt1 --verbose --lr 0.001 --n-graphs 6000 --p 4.73 --q 1.27 --n-nodes 250 --n-communities 2

• GA-MLP-H (hardness rank 3)

python3.7m main_sbm.py --model swl_gnn_dim10_bh120_mint0_mott0_stack1_idnt1 --verbose --lr 0.001 --n-graphs 6000 --p 5 --q 1 --n-nodes 250 --n-communities 2

• GA-MLP-H (hardness rank 4)

python3.7m main_sbm.py --model swl_gnn_dim10_bh120_mint0_mott0_stack1_idnt1 --verbose --lr 0.001 --n-graphs 6000 --p 5.5 --q 0.5 --n-nodes 250 --n-communities 2

• GA-MLP-H (hardness rank 5)

python3.7m main_sbm.py --model swl_gnn_dim10_bh120_mint0_mott0_stack1_idnt1 --verbose --lr 0.001 --n-graphs 6000 --p 6 --q 0 --n-nodes 250 --n-communities 2

• GA-MLP-A (hardness rank 1)

python3.7m main_sbm.py --model swl_gnn_dim10_gin120_mint0_mott0_stack1_idnt1 --verbose --lr 0.001 --n-graphs 6000 --p 4.5 --q 1.5 --n-nodes 250 --n-communities 2

• GA-MLP-A (hardness rank 2)

python3.7m main_sbm.py --model swl_gnn_dim10_gin120_mint0_mott0_stack1_idnt1 --verbose --lr 0.001 --n-graphs 6000 --p 4.73 --q 1.27 --n-nodes 250 --n-communities 2

• GA-MLP-A (hardness rank 3)

python3.7m main_sbm.py --model swl_gnn_dim10_gin120_mint0_mott0_stack1_idnt1 --verbose --lr 0.001 --n-graphs 6000 --p 5 --q 1 --n-nodes 250 --n-communities 2

• GA-MLP-A (hardness rank 4)

python3.7m main_sbm.py --model swl_gnn_dim10_gin120_mint0_mott0_stack1_idnt1 --verbose --lr 0.001 --n-graphs 6000 --p 5.5 --q 0.5 --n-nodes 250 --n-communities 2

• GA-MLP-A (hardness rank 5)

python3.7m main_sbm.py --model swl_gnn_dim10_gin120_mint0_mott0_stack1_idnt1 --verbose --lr 0.001 --n-graphs 6000 --p 6 --q 0 --n-nodes 250 --n-communities 2

Counting attributed walks

Click

• GA-MLP-A (Cora)

python3.7m main.py --dataset cora_path --lr 0.01 --epochs 20000 --num_hops 2 --hid_dim 256 --model swl_gnn --op_base adj

• GA-MLP-A+ (Cora)

python3.7m main.py --dataset cora_path --lr 0.01 --epochs 20000 --num_hops 4 --hid_dim 32 --model swl_gnn --op_base adj

• GIN (Cora)

python3.7m main.py --dataset cora_path --lr 0.01 --epochs 20000 --num_hops 2 --hid_dim 32 --model gin_jk

• GA-MLP-A (RRG)

python3.7m main.py --dataset regular --lr 0.01 --epochs 20000 --num_hops 3 --hid_dim 32 --model swl_gnn --op_base adj

• GA-MLP-A+ (RRG)

python3.7m main.py --dataset regular --lr 0.001 --epochs 20000 --num_hops 6 --hid_dim 256 --model swl_gnn --op_base adj

• GIN (RRG)

python3.7m main.py --dataset regular --lr 0.005 --epochs 20000 --num_hops 3 --hid_dim 16 --model gin_jk

Citation

@article{chen2020graph,
  title={On Graph Neural Networks versus Graph-Augmented MLPs},
  author={Chen, Lei and Chen, Zhengdao and Bruna, Joan},
  journal={arXiv preprint arXiv:2010.15116},
  year={2020}
}