Nested Graph Neural Network (NGNN) is a general framework to improve a base GNN's expressive power and performance. It consists of a base GNN (usually a weak message-passing GNN) and an outer GNN. In NGNN, we extract a rooted subgraph around each node, and let the base GNN to learn a subgraph representation from the rooted subgraph, which is used as the root node's representation. Then, the outer GNN further learns a graph representation from these root node representations returned from the base GNN (in this paper, we simply let the outer GNN be a global pooling layer without graph convolution). NGNN is proved to be more powerful than 1-WL, being able to discriminate almost all r-regular graphs where 1-WL always fails. In contrast to other high-order GNNs, NGNN only incurs a constant time higher time complexity than its base GNN (given the rooted subgraph size is bounded). NGNN often shows immediate performance gains in real-world datasets when applying it to a weak base GNN.
Stable: Python 3.8 + PyTorch 1.8.1 + PyTorch_Geometric 1.7.0 + OGB 1.3.1
Latest: Python 3.8 + PyTorch 1.9.0 + PyTorch_Geometric 1.7.2 + OGB 1.3.1
Install PyTorch
Install PyTorch_Geometric
Install OGB
Install rdkit by
conda install -c conda-forge rdkit
To run 1-GNN, 1-2-GNN, 1-3-GNN, 1-2-3-GNN and their nested versions on QM9, install k-gnn by executing
python setup.py install
under "software/k-gnn-master/".
Other required python libraries include: numpy, scipy, tqdm etc.
To run Nested GCN on MUTAG (with subgraph height=3 and base GCN #layers=4), type:
python run_tu.py --model NestedGCN --h 3 --layers 4 --node_label spd --use_rd --data MUTAG
To compare it with a base GCN model only, type:
python run_tu.py --model GCN --layers 4 --data MUTAG
To reproduce the added experiments with hyperparameter searching, type:
python run_tu.py --model GCN --search --data MUTAG
python run_tu.py --model NestedGCN --h 0 --search --node_label spd --use_rd --data MUTAG
Replace with "--data all" and "--model all" to run all models (NestedGCN, NestedGraphSAGE, NestedGIN, NestedGAT) on all datasets.
We include the commands for reproducing the QM9 experiments in "run_all_targets_qm9.sh". Uncomment the corresponding command in this file, and then run
./run_all_targets_qm9.sh 0 11
to execute this command repeatedly for all 12 targets.
To reproduce the ogb-molhiv experiment, run
python run_ogb_mol.py --h 4 --num_layer 6 --save_appendix _h4_l6_spd_rd --dataset ogbg-molhiv --node_label spd --use_rd --drop_ratio 0.65 --runs 10
When finished, to get the ensemble test result, run
python run_ogb_mol.py --h 4 --num_layer 6 --save_appendix _h4_l6_spd_rd --dataset ogbg-molhiv --node_label spd --use_rd --drop_ratio 0.65 --runs 10 --continue_from 100 --ensemble
To reproduce the ogb-molpcba experiment, run
python run_ogb_mol.py --h 3 --num_layer 4 --save_appendix _h3_l4_spd_rd --dataset ogbg-molpcba --subgraph_pooling center --node_label spd --use_rd --drop_ratio 0.35 --epochs 150 --runs 10
When finished, to get the ensemble test result, run
python run_ogb_mol.py --h 3 --num_layer 4 --save_appendix _h3_l4_spd_rd --dataset ogbg-molpcba --subgraph_pooling center --node_label spd --use_rd --drop_ratio 0.35 --epochs 150 --runs 10 --continue_from 150 --ensemble --ensemble_lookback 140
To reproduce Appendix C Figure 3, run the following commands:
python run_simulation.py --n 10 20 40 80 160 320 640 1280 --save_appendix _node --N 10 --h 10
python run_simulation.py --n 10 20 40 80 160 320 640 1280 --save_appendix _graph --N 100 --h 10 --graph
The results will be saved in "results/simulation_node/" and "results/simulation_graph/".
We have tried our best to clean the code. We will keep polishing it after the author response. If you encounter any errors or bugs, please let us know in OpenReview. Hope you enjoy the code!
-
Write a doc or plot a graph to explain the NGNN data structure defined in utils.py
-
Make pretransform to NGNN data structure parallel.