GraKeL is a library that provides implementations of several well-established graph kernels. The library unifies these kernels into a common framework. Furthermore, it provides implementations of some frameworks that work on top of graph kernels. Specifically, GraKeL contains 16 kernels and 2 frameworks. The library is compatible with the scikit-learn pipeline allowing easy and fast integration inside machine learning algorithms.
In detail, the following kernels and frameworks are currently implemented:
- Vertex histogram kernel
- Edge histogram kernel
- Shortest path kernel from Borgwardt and Kriegel: Shortest-path kernels on graphs (ICDM 2005)
- Graphlet kernel from Shervashidze et al.: Efficient graphlet kernels for large graph comparison (AISTATS 2009)
- Random walk kernel from Vishwanathan et al.: Graph Kernels (JMLR 11(Apr))
- Neighborhood hash graph kernel from Hido and Kashima: A Linear-time Graph Kernel (ICDM 2009)
- Weisfeiler-Lehman framework from Shervashidze et al.: Weisfeiler-Lehman Graph Kernels (JMLR 12(Sep))
- Neighborhood subgraph pairwise distance kernel from Costa and De Grave: Fast Neighborhood Subgraph Pairwise Distance Kernel (ICML 2010)
- Lovasz-theta kernel from Johansson et al.: Global graph kernels using geometric embeddings (ICML 2014)
- SVM-theta kernel from Johansson et al.: Global graph kernels using geometric embeddings (ICML 2014)
- Ordered decompositional DAG kernel from Da San Martino et al.: A Tree-Based Kernel for Graphs (SDM 2012)
- GraphHopper kernel from Feragen et al.: Scalable kernels for graphs with continuous attributes (NIPS 2013)
- Propagation kernel from Neumann et al.: Propagation kernels: efficient graph kernels from propagated information (Machine Learning 102(2))
- Pyramid match kernel from Nikolentzos et al.: Matching Node Embeddings for Graph Similarity (AAAI 2017)
- Subgraph matching kernel from Kriege and Mutzel: Subgraph Matching Kernels for Attributed Graphs (ICML 2012)
- Multiscale Laplacian kernel from Kondor and Pan: The Multiscale Laplacian Graph Kernel (NIPS 2016)
- Core framework from Nikolentzos et al.: A Degeneracy Framework for Graph Similarity (IJCAI 2018)
- Weisfeiler-Lehman optimal assignment kernel from Kriege et al.: On Valid Optimal Assignment Kernels and Applications to Graph Classification (NIPS 2016)
To learn how to install and use GraKeL, and to find out more about the implemented kernels and frameworks, please read our documentation. To learn about the functionality of the library and about example applications, check out our examples in the examples/ directory and our tutorials in the tutorials/ directory.
In case you find a bug, please open an issue. To propose a new kernel, you can open a feature request.
Installation
The GraKeL library requires the following packages to be installed:
- Python (>=2.7, >=3.5)
- NumPy (>=1.8.2)
- SciPy (>=0.13.3)
- Cython (>=0.27.3)
- cvxopt (>=1.2.0) [optional]
- future (>=0.16.0) (for python 2.7)
To install the package, run:
$ pip install grakelRunning tests
To test the package, execute:
$ pytestRunning examples
$ cd examples
$ python shortest_path.py
Cite
If you use GraKeL in a scientific publication, please cite our paper (http://jmlr.org/papers/volume21/18-370/18-370.pdf):
@article{JMLR:v21:18-370,
author = {Giannis Siglidis and Giannis Nikolentzos and Stratis Limnios and Christos Giatsidis and Konstantinos Skianis and Michalis Vazirgiannis},
title = {GraKeL: A Graph Kernel Library in Python},
journal = {Journal of Machine Learning Research},
year = {2020},
volume = {21},
number = {54},
pages = {1-5}
}
License
GraKeL is distributed under the BSD 3-clause license. The library makes use of the C++ source code of BLISS (a tool for computing automorphism groups and canonical labelings of graphs) which is LGPL licensed. Futhermore, the cvxopt package (a software package for convex optimization) which is an optional dependency of GraKeL is GPL licensed.
Acknowledgements
We would like to thank @eddiebergman for modernizing our CI and extending our python support.