SGS-GNN: A Supervised Graph Sparsifier for Graph Neural Networks

SGS-GNN is a novel supervised graph sparsification algorithm that learns the sampling probability distribution of edges and samples sparse subgraphs of a user-specified size to reduce the memory required by GNNs for inference tasks on large graphs.

Installation:

These are the necessary packages for installation from scratch and other related packages.

Python version: 3.11
Pytorch version: 2.0.1
Cuda: 11.7
Cudnn: 8.6
Pytorch-Geometric: 2.3.1

For direct installation, Conda packages are in environment.yml, and PIP packages are in requirements.txt and can be imported as,

conda env create -f environment.yml
pip install -r requirements.txt

Example usage

In the sgs-gnn-batch/Scripts folder, there are bash-scripts that show how our methods can be run. Below are some of the usage examples.

python main.py --dataset SmallCora: Will run dataset SmallCora with default settings.

For a more specific combination,

python main.py --dataset SmallCora --mode 'learned' --runs 5 --epochs 200 --save_csv True --sample_perc 0.2 --edge_mlp_type GCN --GNN GCN --nhid 128 --sparse_edge_mlp True --conditional True --reg1 True --reg2 True

Interpretation:

• mode learned is for SGS-GNN. Options are full, edge, random, learned for fixed distribution samplers, complete graphs version, or our learned sampler.
• runs How many times we want to execute the progam
• epochs refers to maximum number of epochs
• sample_perc sparsity control 0.2 refers to 20% sparsity
• edge_mlp_type GCN means for EdgeMLP probabilty encoding what type neural network we want to use, choices are MLP, GSAGE, and GCN
• GNN what would be downstream GNN for node classification. Choices are, GCN, GAT, GIN, Cheb
• nhid Number of hidden neurons
• sparse_edge_mlp This is EdgeMLP; during the encoding step, do you want to use a sparse graph or not? This is primarily for large-scale graphs.
• conditional Perform conditional updates of EdgeMLP
• reg1 Regularizer 1 is for the assortative loss $L_{assor}$
• reg2 Regularizer 2 is for the consistency loss $L_{cons}$

Some other parameters and settings can be found at sgs-gnn-batch/parser.py

Codes of Related Methods:

• NeuralSparse: https://github.com/flyingdoog/PTDNet
• SparseGAT: https://github.com/Yangyeeee/SGAT
• Mixture of graphs (MOG): https://github.com/yanweiyue/MoG
• GraphSAINT: https://pytorch-geometric.readthedocs.io/en/2.3.1/_modules/torch_geometric/loader/graph_saint.html
• ClusterGCN: https://pytorch-geometric.readthedocs.io/en/2.3.1/_modules/torch_geometric/loader/cluster.html
• DropEdge: https://github.com/DropEdge/DropEdge