minimum-redundancy maximum-relevancy (MRMR) for feature selection - matlab

Question

I am using a filter measure in feature selection called (MRMR) minimum-redundancy maximum-relevancy. After i run the code below
function testselection
addpath('C:\Users\Desktop\mRMR_0.9_compiled\mi_0.9');
FeaturesFile = dlmread('test.txt')
[x,y]=size(FeaturesFile)
Features=FeaturesFile(:,1:y-1)
classLeble=FeaturesFile(:,y)
[fea] = mrmr_mid_d(Features, classLeble, 5)
end
I got the following error
Undefined function or variable 'estpab'.
Error in mutualinfo (line 21)
[p12, p1, p2] = estpab(vec1,vec2);
Error in mrmr_mid_d (line 17)
t(i) = mutualinfo(d(:,i), f);
Error in mrmr (line 9)
[fea] = mrmr_mid_d(Features, classLeble, 5);
The code for mrmr_mid_d as below
function [fea] = mrmr_mid_d(d, f, K)
% function [fea] = mrmr_mid_d(d, f, K)
%
% MID scheme according to MRMR
%
% By Hanchuan Peng
% April 16, 2003
%
bdisp=0;
nd = size(d,2);
nc = size(d,1);
t1=cputime;
for i=1:nd,
t(i) = mutualinfo(d(:,i), f);
end;
fprintf('calculate the marginal dmi costs %5.1fs.\n', cputime-t1);
[tmp, idxs] = sort(-t);
fea_base = idxs(1:K);
fea(1) = idxs(1);
KMAX = min(1000,nd); %500
idxleft = idxs(2:KMAX);
k=1;
if bdisp==1,
fprintf('k=1 cost_time=(N/A) cur_fea=%d #left_cand=%d\n', ...
fea(k), length(idxleft));
end;
for k=2:K,
t1=cputime;
ncand = length(idxleft);
curlastfea = length(fea);
for i=1:ncand,
t_mi(i) = mutualinfo(d(:,idxleft(i)), f);
mi_array(idxleft(i),curlastfea) = getmultimi(d(:,fea(curlastfea)), d(:,idxleft(i)));
c_mi(i) = mean(mi_array(idxleft(i), :));
end;
[tmp, fea(k)] = max(t_mi(1:ncand) - c_mi(1:ncand));
tmpidx = fea(k); fea(k) = idxleft(tmpidx); idxleft(tmpidx) = [];
if bdisp==1,
fprintf('k=%d cost_time=%5.4f cur_fea=%d #left_cand=%d\n', ...
k, cputime-t1, fea(k), length(idxleft));
end;
end;
return;
%=====================================
function c = getmultimi(da, dt)
for i=1:size(da,2),
c(i) = mutualinfo(da(:,i), dt);
end;
And the code for mutualinfo as follows
function h = mutualinfo(vec1,vec2)
%=========================================================
%
%This is a prog in the MutualInfo 0.9 package written by
% Hanchuan Peng.
%
%Disclaimer: The author of program is Hanchuan Peng
% at <penghanchuan#yahoo.com> and <phc#cbmv.jhu.edu>.
%
%The CopyRight is reserved by the author.
%
%Last modification: April/19/2002
%
%========================================================
%
% h = mutualinfo(vec1,vec2)
% calculate the mutual information of two vectors
% By Hanchuan Peng, April/2002
%
[p12, p1, p2] = estpab(vec1,vec2);
h = estmutualinfo(p12,p1,p2);
Could someone know what is the problem?

Answer 1

If you use this Mutual information computation package, you can see, the estpab function compiled only for some platforms (there are estpab.dll for Win32 platform, estpab.mexglx the ELF32 file and estpab.mexmac for Mac).
So, you have to recompile this package for your platform.
First of all you have to configure mex building system. Call mex -setup and select appropriate compiler.
Than you can recompile the Mutual information computation package by calling makeosmex.
Note: If you get some errors during compilation like
d:\MATLAB work\mi\estcondentropy.cpp(65) : error C2668: 'log' : ambiguous call to overloaded function
you have to modify appropriate line (65 in this example) in appropriate file (estcondentropy.cpp in this example) from the muInf /= log(2); to muInf /= log(2.0);

Answer 2

In the path C:\Users\Desktop\ isn't the user missing?
Usually windows paths are something like C:\Users\Username\Desktop\
You are probably adding the wrong path, so the program cannot find the function estpab at all.