scikit-rebate

This package includes a scikit-learn-compatible Python implementation of ReBATE, a suite of Relief-based feature selection algorithms for Machine Learning. These Relief-Based algorithms (RBAs) are designed for feature weighting/selection as part of a machine learning pipeline (supervised learning). Presently this includes the following core RBAs: ReliefF, SURF, SURF*, MultiSURF*, and MultiSURF. Additionally, an implementation of the iterative TuRF mechanism and VLSRelief is included. It is still under active development and we encourage you to check back on this repository regularly for updates.

These algorithms offer a computationally efficient way to perform feature selection that is sensitive to feature interactions as well as simple univariate associations, unlike most currently available filter-based feature selection methods. The main benefit of Relief algorithms is that they identify feature interactions without having to exhaustively check every pairwise interaction, thus taking significantly less time than exhaustive pairwise search.

Certain algorithms require user specified run parameters (e.g. ReliefF requires the user to specify some 'k' number of nearest neighbors).

Relief algorithms are commonly applied to genetic analyses, where epistasis (i.e., feature interactions) is common. However, the algorithms implemented in this package can be applied to almost any supervised classification data set and supports:

Feature sets that are discrete/categorical, continuous-valued or a mix of both
Data with missing values
Binary endpoints (i.e., classification)
Multi-class endpoints (i.e., classification)
Continuous endpoints (i.e., regression)

Built into this code, is a strategy to 'automatically' detect from the loaded data, these relevant characteristics.

Of our two initial ReBATE software releases, this scikit-learn compatible version primarily focuses on ease of incorporation into a scikit learn analysis pipeline. This code is most appropriate for scikit-learn users, Windows operating system users, beginners, or those looking for the most recent ReBATE developments.

An alternative 'stand-alone' version of ReBATE is also available that focuses on improving run-time with the use of Cython for optimization. This implementation also outputs feature names and associated feature scores as a text file by default.

License

Please see the repository license for the licensing and usage information for scikit-rebate.

Generally, we have licensed scikit-rebate to make it as widely usable as possible.

Installation

scikit-rebate is built on top of the following existing Python packages:

NumPy
SciPy
scikit-learn

All 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 Python 3 over Python 2 if you're given the choice.

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

conda install numpy scipy scikit-learn

Once the prerequisites are installed, you should be able to install scikit-rebate with a pip command:

pip install skrebate

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

Usage

We have designed the Relief algorithms to be integrated directly into scikit-learn machine learning workflows. For example, the ReliefF algorithm can be used as a feature selection step in a scikit-learn pipeline as follows.

import pandas as pd
import numpy as np
from sklearn.pipeline import make_pipeline
from skrebate import ReliefF
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score

genetic_data = pd.read_csv('https://github.com/EpistasisLab/scikit-rebate/raw/master/data/'
                           'GAMETES_Epistasis_2-Way_20atts_0.4H_EDM-1_1.tsv.gz',
                           sep='\t', compression='gzip')

features, labels = genetic_data.drop('class', axis=1).values, genetic_data['class'].values

clf = make_pipeline(ReliefF(n_features_to_select=2, n_neighbors=100),
                    RandomForestClassifier(n_estimators=100))

print(np.mean(cross_val_score(clf, features, labels)))
>>> 0.795

For more information on the Relief algorithms available in this package and how to use them, please refer to our usage documentation.

Contributing to scikit-rebate

We welcome you to check the existing issues for bugs or enhancements to work on. If you have an idea for an extension to scikit-rebate, please file a new issue so we can discuss it.

Please refer to our contribution guidelines prior to working on a new feature or bug fix.

Citing scikit-rebate

If you use scikit-rebate in a scientific publication, please consider citing the following paper:

Ryan J. Urbanowicz, Randal S. Olson, Peter Schmitt, Melissa Meeker, Jason H. Moore (2017). Benchmarking Relief-Based Feature Selection Methods. arXiv preprint, under review.

BibTeX entry: