This is the TensorFlow implementation of our paper accepted by IJCAI 2020:
Hongmin Zhu, Fuli Feng, Xiangnan He, Xiang Wang, Yan Li, Kai Zheng, Yongdong Zhang, Bilinear Graph Neural Network with Neighbor Interactions. Paper in arXiv.
We propose a new graph convolution operator, augmenting the weighted sum with pairwise interactions of the representations of neighbor nodes. We specify two BGNN models named BGCN and BGAT, based on the well-known GCN and GAT, respectively.
The code is tested by server with RTX 1080Ti running in a docker container which includes the following packages:
- python == 3.6.3
- tensorflow == 1.4.0
- numpy == 1.13.3
- scipy == 1.0.0
- networkx == 2.0
In addition, CUDA 8 and cuDNN 6 have been used.
Here are the instruction commands for running the codes on Citeseer in a docker container.
- Command
cd BGCN/1-layer/BGCN-A/
python gcn.py --model bgcn --dropout 0.0 --weight_decay 5e-4 --alpha 0.9 --epochs 2000 --learning_rate 0.005
- Output
Epoch: 0001 train_loss= 1.79177 train_acc= 0.1250 val_loss= 1.79168 val_acc= 0.2840 tst_loss= 1.79169 tst_acc= 0.2600 time= 0.112
Epoch: 0002 train_loss= 1.79154 train_acc= 0.5250 val_loss= 1.79162 val_acc= 0.3820 tst_loss= 1.79162 tst_acc= 0.3660 time= 0.055
...
Epoch: 1999 train_loss= 1.52770 train_acc= 0.8333 val_loss= 1.67561 val_acc= 0.6460 tst_loss= 1.67752 tst_acc= 0.6940 time= 0.049
Epoch: 2000 train_loss= 1.52768 train_acc= 0.8333 val_loss= 1.67560 val_acc= 0.6460 tst_loss= 1.67750 tst_acc= 0.6940 time= 0.048
test_loss= 1.69706 test_acc= 0.7010
early stop #epoch 1004 val_loss= 1.69771 val_acc= 0.6820
- Command
cd BGCN/1-layer/BGCN-T/
python gcn.py --model bgcn --dropout 0.0 --weight_decay 5e-4 --alpha 0.7 --epochs 2000 --learning_rate 0.005
- Output
Epoch: 0001 train_loss= 1.79160 train_acc= 0.2083 val_loss= 1.79143 val_acc= 0.2900 tst_loss= 1.79137 tst_acc= 0.2910 time= 0.150
Epoch: 0002 train_loss= 1.79057 train_acc= 0.6333 val_loss= 1.79115 val_acc= 0.3380 tst_loss= 1.79107 tst_acc= 0.3910 time= 0.071
...
Epoch: 1999 train_loss= 1.57680 train_acc= 0.8583 val_loss= 1.69943 val_acc= 0.6600 tst_loss= 1.69793 tst_acc= 0.6930 time= 0.058
Epoch: 2000 train_loss= 1.57679 train_acc= 0.8583 val_loss= 1.69942 val_acc= 0.6600 tst_loss= 1.69792 tst_acc= 0.6930 time= 0.064
test_loss= 1.71482 test_acc= 0.7080
early stop #epoch 893 val_loss= 1.71812 val_acc= 0.6860
- Command
cd BGAT/1-layer/BGAT-A/
python gat.py --head 1 --feadrop 0.0 --attdrop 0.6 --weight_decay 5e-4 --alpha 0.9 --epochs 2000 --learning_rate 0.005
- Output
Epoch: 0001 train_loss= 1.79173 train_acc= 0.1583 val_loss= 1.79085 val_acc= 0.3160 tst_loss= 1.79142 tst_acc= 0.2620 time= 3.564
Epoch: 0002 train_loss= 1.79165 train_acc= 0.1833 val_loss= 1.79147 val_acc= 0.3700 tst_loss= 1.79160 tst_acc= 0.3200 time= 0.227
...
Epoch: 1312 train_loss= 1.74598 train_acc= 0.4583 val_loss= 1.76272 val_acc= 0.6640 tst_loss= 1.76328 tst_acc= 0.6950 time= 0.222
Epoch: 1313 train_loss= 1.73696 train_acc= 0.4417 val_loss= 1.76269 val_acc= 0.6680 tst_loss= 1.76322 tst_acc= 0.6950 time= 0.230
Early stop!
test_loss= 1.77900 test_acc= 0.7180
early stop #epoch 297 val_loss= 1.77891 val_acc= 0.7200
- Command
cd BGAT/1-layer/BGAT-T/
python gat.py --head 1 --feadrop 0.4 --attdrop 0.6 --weight_decay 5e-4 --alpha 0.9 --epochs 2000 --learning_rate 0.005
- Output
Epoch: 0001 train_loss= 1.79172 train_acc= 0.2333 val_loss= 1.79192 val_acc= 0.1700 tst_loss= 1.79161 tst_acc= 0.1990 time= 3.325
Epoch: 0002 train_loss= 1.79166 train_acc= 0.1667 val_loss= 1.79168 val_acc= 0.1980 tst_loss= 1.79161 tst_acc= 0.2360 time= 0.222
...
Epoch: 1058 train_loss= 1.75115 train_acc= 0.4333 val_loss= 1.76869 val_acc= 0.6780 tst_loss= 1.76848 tst_acc= 0.6680 time= 0.281
Epoch: 1059 train_loss= 1.74892 train_acc= 0.4000 val_loss= 1.76859 val_acc= 0.6820 tst_loss= 1.76831 tst_acc= 0.6790 time= 0.280
Early stop!
test_loss= 1.77275 test_acc= 0.7300
early stop #epoch 582 val_loss= 1.77321 val_acc= 0.7120
- Command
cd BGCN/2-layer/BGCN-A/
python gcn.py --model bgcn --dropout 0.6 --weight_decay 1e-4 --alpha 0.3 --beta 0.7 --epochs 2000 --learning_rate 0.005
- Output
Epoch: 0001 train_loss= 1.79182 train_acc= 0.1333 val_loss= 1.79101 val_acc= 0.3180 tst_loss= 1.79108 tst_acc= 0.2940 time= 0.139
Epoch: 0002 train_loss= 1.79003 train_acc= 0.4000 val_loss= 1.79026 val_acc= 0.4460 tst_loss= 1.79031 tst_acc= 0.4400 time= 0.067
...
Epoch: 1999 train_loss= 0.96810 train_acc= 0.9417 val_loss= 1.36601 val_acc= 0.7200 tst_loss= 1.35519 tst_acc= 0.7130 time= 0.056
Epoch: 2000 train_loss= 0.94014 train_acc= 0.9167 val_loss= 1.36595 val_acc= 0.7200 tst_loss= 1.35513 tst_acc= 0.7130 time= 0.049
test_loss= 1.37182 test_acc= 0.7140
early stop #epoch 1727 val_loss= 1.38217 val_acc= 0.7240
- Command
cd BGCN/2-layer/BGCN-T/
python gcn.py --model bgcn --dropout 0.0 --weight_decay 1e-4 --alpha 0.3 --beta 0.7 --epochs 2000 --learning_rate 0.005
- Output
Epoch: 0001 train_loss= 1.79181 train_acc= 0.1083 val_loss= 1.79105 val_acc= 0.3120 tst_loss= 1.79096 tst_acc= 0.3390 time= 0.383
Epoch: 0002 train_loss= 1.78918 train_acc= 0.6583 val_loss= 1.79016 val_acc= 0.4720 tst_loss= 1.79004 tst_acc= 0.4620 time= 0.086
...
Epoch: 1999 train_loss= 0.78423 train_acc= 0.9833 val_loss= 1.33716 val_acc= 0.7300 tst_loss= 1.31960 tst_acc= 0.7190 time= 0.059
Epoch: 2000 train_loss= 0.78418 train_acc= 0.9833 val_loss= 1.33711 val_acc= 0.7300 tst_loss= 1.31955 tst_acc= 0.7190 time= 0.062
test_loss= 1.31955 test_acc= 0.7190
early stop #epoch 2000 val_loss= 1.33711 val_acc= 0.7300
- Command
cd BGAT/2-layer/BGAT-A/
python gat.py --head 1 --hidden1 8 --feadrop 0.6 --attdrop 0.6 --weight_decay 5e-4 --alpha 0.7 --beta 0.3 --epochs 2000 --learning_rate 0.005
- Output
Epoch: 0001 train_loss= 1.79191 train_acc= 0.1750 val_loss= 1.79127 val_acc= 0.2160 tst_loss= 1.79155 tst_acc= 0.2240 time= 5.875
Epoch: 0002 train_loss= 1.79350 train_acc= 0.1750 val_loss= 1.79081 val_acc= 0.3400 tst_loss= 1.79107 tst_acc= 0.3310 time= 0.480
...
Epoch: 1116 train_loss= 1.41075 train_acc= 0.4417 val_loss= 1.47168 val_acc= 0.7160 tst_loss= 1.47457 tst_acc= 0.7180 time= 0.492
Epoch: 1117 train_loss= 1.45948 train_acc= 0.4250 val_loss= 1.47192 val_acc= 0.7140 tst_loss= 1.47484 tst_acc= 0.7210 time= 0.462
Early stop!
test_loss= 1.46169 test_acc= 0.7410
early stop #epoch 1016 val_loss= 1.46063 val_acc= 0.7440
- Command
cd BGAT/2-layer/BGAT-T/
python gat.py --head 1 --hidden1 8 --feadrop 0.6 --attdrop 0.6 --weight_decay 1e-3 --alpha 0.5 --beta 0.5 --epochs 2000 --learning_rate 0.005
- Output
Epoch: 0001 train_loss= 1.79209 train_acc= 0.1833 val_loss= 1.79096 val_acc= 0.2120 tst_loss= 1.79143 tst_acc= 0.2230 time= 5.741
Epoch: 0002 train_loss= 1.79474 train_acc= 0.1750 val_loss= 1.79024 val_acc= 0.3400 tst_loss= 1.79066 tst_acc= 0.3380 time= 0.457
...
Epoch: 0902 train_loss= 1.44180 train_acc= 0.4667 val_loss= 1.51039 val_acc= 0.7220 tst_loss= 1.50913 tst_acc= 0.7260 time= 0.465
Epoch: 0903 train_loss= 1.48058 train_acc= 0.4167 val_loss= 1.51134 val_acc= 0.7220 tst_loss= 1.50986 tst_acc= 0.7290 time= 0.452
Early stop!
test_loss= 1.54286 test_acc= 0.7400
early stop #epoch 492 val_loss= 1.54243 val_acc= 0.7440
Besides, we provide another possible interaction scope which only considers the interactions between neighbors in graphs (see BGCN-N and BGAT-N).
We utilize three benchmark datasets of citation network---Pubmed, Cora and Citeseer.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent