This repository is the official implementation of Bilateral Message Passing.
-
Analogous to the bilateral image filter, we propose a bi-MP scheme to address over-smoothing in classic MP GNNs.
-
Our proposed scheme can be generalized to all ordinary MP GNNs (e.g. SOTA MP-GNNs such as GCN, GraphSAGE, and GAT).
Various categories contains scripts from benchmarking-gnns.
📋 Follows the instructions to add a custom bilateral-MP layer.
To install requirements:
# Install python environment
conda env create -f environment_gpu.yml
# Activate environment
conda activate benchmark_gnn
To train & evaluate the model(s) in the paper with specified dataset, and model, run this command:
python main_{DatasetDependentName}.py --config Configpath/Configfname.json --gpu_id 0 --model ModelName
Our model achieves the following performance on :
| Model name | TEST MAE (mean,std) | #Params (#Layers) |
|---|---|---|
| bi-GatedGCN (ours) | 0.166 (0.009) | 511974 (16) |
| GatedGCN | 0.214 (0.013) | 505011 (16) |
| bi-GraphSage (ours) | 0.245 (0.009) | 516651 (16) |
| bi-GCN (ours) | 0.276 (0.007) | 536482 (16) |
| bi-GAT (ours) | 0.277 (0.012) | 535536 (16) |
| GCN | 0.367 (0.011) | 505079 (16) |
| GAT | 0.384 (0.007) | 531345 (16) |
| GraphSage | 0.398 (0.002) | 505341 (16) |
| Model name | TEST F1 (mean,std) | #Params (#Layers) |
|---|---|---|
| bi-GatedGCN (ours) | 0.812 (0.004) | 125832 (4) |
| GatedGCN | 0.808 (0.003) | 97858 (4) |
| bi-GraphSage (ours) | 0.693 (0.016) | 131861 (4) |
| bi-GAT (ours) | 0.675 (0.002) | 115609 (4) |
| GAT | 0.673 (0.002) | 96182 (4) |
| GraphSage | 0.665 (0.003) | 99263 (4) |
| bi-GCN (ours) | 0.642 (0.001) | 118496 (4) |
| GCN | 0.630 (0.001) | 95702 (4) |
| Model name | TEST ACC (mean,std) | #Params (#Layers) |
|---|---|---|
| bi-GatedGCN (ours) | 97.575 (0.085) | 101365 (4) |
| bi-GraphSage (ours) | 97.438 (0.155) | 110400 (4) |
| GatedGCN | 97.340 (0.143) | 125815 (4) |
| GraphSage | 97.312 (0.097) | 114169 (4) |
| GAT | 95.535 (0.205) | 114507 (4) |
| bi-GAT (ours) | 95.363 (0.199) | 104337 (4) |
| bi-GCN (ours) | 90.805 (0.299) | 104217 (4) |
| GCN | 90.705 (0.218) | 110807 (4) |
| Model name | TEST ACC (mean,std) | #Params (#Layers) |
|---|---|---|
| bi-GatedGCN (ours) | 67.850 (0.522) | 110632 (4) |
| GatedGCN | 67.312 (0.311) | 104307 (4) |
| GraphSage | 65.767 (0.308) | 104517 (4) |
| bi-GraphSage (ours) | 64.863 (0.445) | 114312 (4) |
| bi-GAT (ours) | 64.275 (0.458) | 114311 (4) |
| GAT | 64.223 (0.455) | 110704 (4) |
| GCN | 55.710 (0.381) | 101657 (4) |
| bi-GCN (ours) | 54.450 (0.137) | 125564 (4) |
| Model name | TEST ACC (mean,std) | #Params (#Layers) |
|---|---|---|
| bi-GatedGCN (ours) | 76.896 (0.213) | 516211 (16) |
| GatedGCN | 76.082 (0.196) | 504253 (16) |
| bi-GCN (ours) | 71.199 (0.882) | 505149 (16) |
| bi-GAT (ours) | 71.113 (0.869) | 445438 (16) |
| GAT | 70.587 (0.447) | 527824 (16) |
| GCN | 68.498 (0.976) | 501687 (16) |
| bi-GraphSage (ours) | 64.088 (0.182) | 490569 (16) |
| GraphSage | 63.844 (0.110) | 503350 (16) |