GNNAsKernel

Official code for From Stars to Subgraphs: Uplifting Any GNN with Local Structure Awareness

Visualizations

GNN-AK(+)

GNN-AK(+) with SubgraphDrop

Setup

# params
# 10/6/2021, newest packages. 
ENV=gnn_ak
CUDA=11.1
TORCH=1.9.1
PYG=2.0.1

# create env 
conda create --name $ENV python=3.9 -y
conda activate $ENV

# install pytorch 
conda install pytorch=$TORCH torchvision torchaudio cudatoolkit=$cuda -c pytorch -c nvidia -y

# install pyg2.0
conda install pyg=$PYG -c pyg -c conda-forge -y

# install ogb 
pip install ogb

# install rdkit
conda install -c conda-forge rdkit -y

# update yacs and tensorboard
pip install yacs==0.1.8 --force  # PyG currently use 0.1.6 which doesn't support None argument. 
pip install tensorboard
pip install matplotlib

Code structure

core/ contains all source code.
train/ contains all scripts for available datasets.

Subgraph extraction is implemented as data transform operator in PyG. See core/transform.py. The transform layer will built the mapping from original nodes and edges to all subgraphs.
The mappings are used directly in GNN-AK(+) to online build the combined subgraphs for each graph, see core/model.py. (For each graph with N node, N subgraphs are combined to a gaint subgraph first. Then for batch, all combined gaint subgraphs are combined again.)
SubgraphDrop is implemented inside core/transform.py, see here. And the usage in core/model.py.
core/model_utils/pyg_gnn_wrapper.py is the place to add your self-designed GNN layer X and then use X-AK(+) on fly~

Hyperparameters

See core/config.py for all options.

Run normal GNNs

See core/model_utls/pyg_gnn_wrapper.py for more options.

Custom new GNN convolutional layer 'X' can be plugged in core/model_utls/pyg_gnn_wrapper.py, and use 'X' as model.gnn_type option.

# Run different normal GNNs 
python -m train.zinc model.mini_layers 0 model.gnn_type GINEConv
python -m train.zinc model.mini_layers 0 model.gnn_type SimplifiedPNAConv
python -m train.zinc model.mini_layers 0 model.gnn_type GCNConv
python -m train.zinc model.mini_layers 0 model.gnn_type GATConv
python -m train.zinc model.mini_layers 0 model.gnn_type ...

python -m train.zinc model.num_layers 6 model.mini_layers 0 model.gnn_type GCNConv # 6-layer GCN

Run different datasets

See all available datasets under train folder.

# Run different datasets
python -m train.zinc 
python -m train.cifar10 
python -m train.counting 
python -m train.graph_property 
python -m ...

Run GNN-AK

Fully theoretically explained by Subgraph-1-WL*.

Use: model.mini_layers 1 (or >1) model.embs "(0,1)" model.hops_dim 0

python -m train.zinc model.mini_layers 1 model.gnn_type GINEConv model.embs "(0,1)" model.hops_dim 0

Run GNN-AK+

At least as powerful as GNN-AK (or more powerful).

Use: model.mini_layers 1 (or >1) model.embs "(0,1,2)" model.hops_dim 16
These are set as default. See core/config.py.

# Run GNN-AK+ with different normal GNNs
python -m train.zinc model.mini_layers 1 model.gnn_type GINEConv            # 1-layer base model
python -m train.zinc model.mini_layers 1 model.gnn_type SimplifiedPNAConv   # 1-layer base model
python -m train.zinc model.mini_layers 2 model.gnn_type GINEConv            # 2-layer base model
python -m train.zinc model.mini_layers 2 model.gnn_type SimplifiedPNAConv   # 2-layer base model

Run with different number of GNN-AK(+) iterations

Changing number of outer layers.

python -m train.zinc model.num_layers 4 
python -m train.zinc model.num_layers 6 
python -m train.zinc model.num_layers 8

Run with different subgraph patterns

See core/transform.py for detailed implementation.

python -m train.zinc subgraph.hops 2      # 2-hop egonet
python -m train.zinc subgraph.hops 3      # 3-hop egonet

# Run with random-walk subgraphs based on node2vec 
python -m train.zinc subgraph.hops 0 subgraph.walk_length 10 subgraph.walk_p 1.0 subgraph.walk_q 1.0

Run GNN-AK(+) with SubgraphDrop

See option sampling section under core/config.py.

Change sampling.redundancy(R in the paper) to change the resource usage.

python -m train.zinc sampling.mode shortest_path sampling.redundancy 1 sampling.stride 5 sampling.batch_factor 4
python -m train.zinc sampling.mode shortest_path sampling.redundancy 3 sampling.stride 5 sampling.batch_factor 4
python -m train.zinc sampling.mode shortest_path sampling.redundancy 5 sampling.stride 5 sampling.batch_factor 4


python -m train.cifar10 sampling.mode random sampling.redundancy 1 sampling.random_rate 0.07 sampling.batch_factor 8 
python -m train.cifar10 sampling.mode random sampling.redundancy 3 sampling.random_rate 0.21 sampling.batch_factor 8 
python -m train.cifar10 sampling.mode random sampling.redundancy 5 sampling.random_rate 0.35 sampling.batch_factor 8 
## Note: sampling.random_rate = 0.07*sampling.redundancy. 0.07 is set based on dataset.

Results

GNN-AK boosts expressiveness

GNN-AK boosts practical performance

Cite

Please cite our work if you use our code!

@inproceedings{
    zhao2022from,
    title={From Stars to Subgraphs: Uplifting Any {GNN} with Local Structure Awareness},
    author={Lingxiao Zhao and Wei Jin and Leman Akoglu and Neil Shah},
    booktitle={International Conference on Learning Representations},
    year={2022},
    url={https://openreview.net/forum?id=Mspk_WYKoEH}
}

LWL-cpu/GNNAsKernel