Learning Structural Node Representations using Graph Kernels

Code for the paper Learning Structural Node Representations using Graph Kernels.

Requirements

Code is written in Python 3.6 and requires:

• grakel 0.1b7
• scikit-learn 0.21

Basic Usage

To run SEGK on Barbell graph, execute the following command from the project home directory:
python segk.py --path-to-edgelist datasets/barbell.edgelist --path-to-output-file embeddings/barbell.txt

Hyperparameters

The following three hyperparameters can be specified:

• radius: the maximum radius of the neighborhood subgraphs
• dim: the dimensionality of the generated embeddings
• kernel: the employed graph kernel (either the shortest path kernel or the weisfeiler lehman kernel)

Input

The supported input format is an edgelist where the endpoints are separated by the space character:

node1_id node2_id

Output

The output file contains n lines (where n is the number of nodes of the graph) as follows:

node_id dim1 dim2 ... dimd

where dim1, ... , dimd is the d-dimensional representation learned by SEGK.

Experiments

In the project home directory, there are four jupyter notebooks that reproduce the experiments presented in the paper.

Cite

Please cite our paper if you use this code:

@article{nikolentzos2019learning,
  title={Learning Structural Node Representations using Graph Kernels},
  author={Nikolentzos, Giannis and Vazirgiannis, Michalis},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2019}
}