graphkernels

graphkernels is a Python package for computing various graph kernels. The backend is coded in C++ and exposed to Python through SWIG.

The Python and R packages are described at:

• M. Sugiyama, M.E. Ghisu, F. Llinares-López and K. Borgwardt. graphkernels: R and Python packages for graph comparison. Bioinformatics, 2017.

The paper can be found here.

Installation

Installing dependencies

graphkernels requires the following dependencies to be pre-installed in your local environment:

• Python bindings for igraph
• NumPy
• Eigen3

During setup, you will also need the following dependencies:

• a C++ compiler (e.g. gcc, XCode)
• pkg-config and the respective Python bindings
• SWIG

We recommend the following steps for installing the dependencies (except for the C++ compiler):

• On Ubuntu 18.04 LTS:

$ sudo apt-get install libeigen3-dev python3-igraph python3-numpy python3-pkgconfig swig

Some of the packages might not be available in older distributions.

• On MacOSX:

$ brew install eigen pkg-config

Installing graphkernels from PyPI

To install the package through PyPI, type:

$ pip3 install graphkernels --user

It is recommended that version 10 or later of pip is used; newer versions of pip can handle missing dependencies better, by automatically downloading and installing them together with graphkernels (this won't work for the Eigen library). To see which version of pip is installed, type:

$ pip3 --version

If you are using an older version, you can locally upgrade pip by typing:

$ python3 -m pip install --user --upgrade-pip

before attempting the graphkernels installation.

In case you already have an older version of graphkernels installed, you can use the --no-cache-dir option so that your local package cache is ignored and the lastest package is downloaded:

$ pip3 --no-cache-dir install graphkernels --user

Build-Installing graphkernels from main repository

Alternatively, the package can be built from source as follows:

1. Download the latest source code from GitHub (please only use the master branch when doing so, other branches are not guaranteed to contain a correct build):

$ git clone https://github.com/Renelvon/GraphKernels.git

2. Build the package; this will also compile the C++ backend and generate the SWIG wrapper:

$ cd GraphKernels
$ python3 setup.py build

3. Install the package for this user only (showed by --user; installing the package system-wide might require sudo -H):

$ python3 setup.py install --user

Usage

The Tutorial folder should help with getting started using the package. There, you can also find an example script for computing graph kernels through our package on a benchmark dataset.

Compatibility

This version of graphkernels aims to be fast, user-friendly and backwards-compatible (in that order). Accordingly, support for Python 2 in graphkernels has been dropped because it is reaching End-Of-Life very soon. If you care about speed enough to use a C++ backend for your project, you should care as much about future portability and give serious thought before continuing development on Python 2. Most of the ML projects will drop support for Python 2 soon.

Citation

If you use the graphkernels package in your projects please cite our work as follows:

@article{Sugiyama-2017-Bioinformatics,
author = {Sugiyama, Mahito and Ghisu, M. Elisabetta and Llinares-López, Felipe and Borgwardt, Karsten},
title = {graphkernels: R and Python packages for graph comparison},
journal = {Bioinformatics},
volume = {34},
number = {3},
pages = {530--532},
year = {2017},
doi = {10.1093/bioinformatics/btx602},
URL = {http://dx.doi.org/10.1093/bioinformatics/btx602},
}

If our project has made your life easier, we will be happy to receive your feedback, through e-mail or an issue on GitHub.