A centralized repository to benchmark ReBATE feature selection algorithm performance across a variety of parameter settings and datasets.

The following directories are included in this repository, all of which can be used to replicate the experimental results of our ReBATE benchmark:

• benchmark-data: Contains all of the supervised classification benchmark datasets used in this benchmark study. All datasets are tab-separated, gzipped, and use the column header Class for the target column.

• model-code: Contains all of the Python scripts used to run the various Relief and other feature selection algorithms on the benchmark datasets.

• post-analysis: Contains the Jupyter notebooks and Python scripts used to analyze and visualize the results of the ReBATE benchmark.

Please see the repository license for the licensing and usage information for the contents of this repository.

This benchmark study uses several existing software packages, including:

• NumPy

• SciPy

• pandas

• scikit-learn

• matplotlib

• Seaborn

• scikit-rebate

Most of the necessary Python packages can be installed via the Anaconda Python distribution, which we strongly recommend that you use. We also strongly recommend that you use of Python 3 over Python 2 if you're given the choice.

NumPy, SciPy, pandas, scikit-learn, matplotlib, and Seaborn can be installed in Anaconda via the command:

conda install numpy scipy pandas scikit-learn matplotlib seaborn

scikit-rebate can be installed with pip via the command:

pip install skrebate

Please file a new issue if you run into installation problems.

If you would like to contribute to the ReBATE feature selection benchmark, please check the existing issues to see if there is an ongoing project you can contribute to.

If you have a suggestion or comment for the benchmark, please file a new issue and thoroughly describe your suggestion or comment. Be sure to first check the existing and closed issues to see if the suggestion or comment hasn't already been discussed.

If you would like to cite this benchmark study, please cite the research papers listed below.

[blank until preprints are posted]