EigenGNN

This is a sample implementation of "Eigen-GNN: a Graph Structure Preserving Plug-in for GNNs, TKDE 2021". (Paper)

Requirements

• numpy
• scipy
• networkx

Usage

As indicated by the title, EigenGNN is a plug-in instead of a stand-along model. Therefore, the code in the main file will only generate initial node representations (i.e., eigenvectors of a graph structure matrix), which can be used together with other GNNs.

The initial node representation can be generated as follows:

from EigenGNN import Eigen_multi
features = Eigen_multi(adj, d, adj_normalize, feature_abs)

An example to use EigenGNN with GCN being the GNN backbone for node classification benchmark datasets (e.g., Cora, CiteSeer, PubMed) is provided in example_gcn.py and can be run as:

python example_gcn.py --method eigen --dataset cora

for only using the eigenspace, or

python example_gcn.py --method feat_eigen --dataset cora

for concatenating the eigenspace with input node features.

Another example to use EigenGNN with GIN being the GNN backbone for graph isomorphism test (modeled as graph classification) is provided in example_gin.py and can be run as:

python example_gin.py --cv-fold 0 --model-type EigenGin --num-mlp-hidden 1 --num-gnn-layers 1 --out-weight-dir dir1 --out-log-dir dir2

The last example is use EigenGNN with PyTorch Geometric is provided in example_pyg_gcn.py and can be run as:

python example_pyg_gcn.py --use_eigengnn

Cite

If you find this code useful, please cite our paper:

@article{zhang2021eigengnn,
  title={Eigen-GNN: a Graph Structure Preserving Plug-in for GNNs},
  author={Zhang, Ziwei and Cui, Peng and Pei, Jian and Wang, Xin and Zhu, Wenwu},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2021},
  publisher={IEEE}
}