FEAST


A Feature Selection Toolbox for C, Java and Matlab.





FEAST provides implementations of common mutual information based filter feature selection algorithms, and an implementation of RELIEF. All functions expect discrete inputs (except RELIEF, which does not depend on the MIToolbox), and they return the selected feature indices. These implementations were developed to help our research into the similarities between these algorithms, and our results are presented in the following paper:

Conditional Likelihood Maximisation: A Unifying Framework for Information Theoretic Feature Selection
G.Brown, A.Pocock, M.Lujan, M.-J.Zhao
Journal of Machine Learning Research, vol 13, pages 27-66 (2012)

If you're interested in the datasets we used for that paper, you can download them from my advisor's site here (28MB).

All FEAST code is licensed under the BSD 3-Clause License. If you use these implementations for academic research please cite the paper above.

FEAST toolbox: Download at MLOSS.org

Contains implementations of:

mim, mrmr, mifs, cmim, jmi, disr, cife, icap, condred, cmi, relief, fcbf, betagamma

And weighted implementations of:

mim, cmim, jmi, disr, cmi

References for these algorithms are provided in the accompanying feast.bib file (in BibTeX format).

FEAST works on discrete inputs, and all continuous values **must** be discretised before use with FEAST. In our experiments we've found that using 10 equal width bins is suitable for many problems, though this is data set size dependent. FEAST produces unreliable results when used with continuous inputs, runs slowly and uses much more memory than usual. The discrete inputs should have small cardinality, FEAST will treat values {1,10,100} the same way it treats {1,2,3} and the latter will be both faster and use less memory.

MATLAB Example (using "data" as our feature matrix, and "labels" as the class label vector):

>> size(data)
ans =
(569,30) %% denoting 569 examples, and 30 features

>> selectedIndices = feast('jmi',5,data,labels) %% selecting the top 5 features using the jmi algorithm
selectedIndices =

28
21
8
27
23

>> selectedIndices = feast('mrmr',10,data,labels) %% selecting the top 10 features using the mrmr algorithm
selectedIndices =

28
24
22
8
27
21
29
4
7
25

>> selectedIndices = feast('mifs',5,data,labels,0.7) %% selecting the top 5 features using the mifs algorithm with beta = 0.7
selectedIndices =

28
24
22
20
29

The library is written in ANSI C for compatibility with the MATLAB mex compiler, except for MIM, FCBF and RELIEF, which are written in MATLAB/OCTAVE script. There is a different implementation of MIM available for use in the C library.

MIToolbox v3.0.0 is required to compile these algorithms, and these implementations supercede the example implementations given in that package (they have more robust behaviour when used with unexpected inputs).

MIToolbox can be found here.

The C library expects all matrices in column-major format (i.e. Fortran style). This is for two reasons, a) MATLAB generates Fortran-style arrays, and b) feature selection iterates over columns rather than rows, unlike most other ML processes.

Compilation instructions:

Repository:

I've hosted the source on GitHub here for anyone who wishes to browse through it. It will be updated approximately the same time as this page.

Update History