www.gusucode.com > Ant_Colony_Optimization_feature_selection > Ant_Colony_Optimization_feature_selection.m/
% Editor:Summer
% Institution: Shenzhen University
% E-mail:1348825332@qq.com
% date:2018-12-16
% 30 runs with 12 datasets after ACOFS
% 10-fold cross-validation with k-nn classifier
% Please download the dataset within this link：https://pan.baidu.com/s/1JzBc-erjfYRcruMSUkH7rA
clc;
clear;
%% load classification datasets
switch dataset
case 'isolet'
load('isolet.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'sonar'
load('sonar.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'Hill_Valley_without_noise_Training'
load('Hill.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'Epileptic Seizure Recognitio'
load('Epileptic Seizure Recognitio.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'redwine'
load('redwine.mat')
dn=1:1:10;
dnsize=10;
dnd=1;
case 'whitewine'
load('whitewine.mat')
dn=1:1:10;
dnsize=10;
dnd=1;
case 'MF'
load('MF.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'SPECTHeart'
load('SPECTHeart.mat')
dn=2:2:20;
dnsize=10;
dnd=2;
case 'Statlog'
load('Statlog.mat')
dn=2:2:20;
dnsize=10;
dnd=2;
case 'Madelon'
load('Madelon.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'Libras Movement'
load('Libras Movement.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'LSVT_voice_rehabilitation'
load('LSVT_voice_rehabilitation.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'drivFaceD'
load('drivFaceD.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'Urban land cover'
load('Urban land cover.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'MEU-Mobile KSD 2016'
load('MEU-Mobile KSD 2016.mat')
dn=5:5:50;
dnsize=10;
dnd=5;
case 'ionosphere'
load('ionosphere.mat')
dn=2:2:20;
dnsize=10;
dnd=2;
case 'ORL'
load('ORL.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'COIL20'
load('COIL20.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'orlraws10P'
load('orlraws10P.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'pixraw10P'
load('pixraw10P.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'warpAR10P'
load('warpAR10P.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'warpPIE10P'
load('warpPIE10P.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'Yale'
load('Yale.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'GLIOMA'
load('GLIOMA.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'lung_discrete'
load('lung_discrete.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'colon'
load('colon.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'lung'
load('lung.mat')
data=[X,Y];
dn=5:5:50;
dnsize=10;
dnd=5;
case 'ForestTypes'
load('ForestTypes.mat')
dn=2:2:20;
dnsize=10;
dnd=2;
end
%% Parameters setting
e=10; % e-fold cross-validation
group=data(:,end);
class=unique(data(:,end));
for i=1:length(class)
sa=[];
sa=data((group==class(i)),:);
[number_of_smile_samples,~] = size(sa); % Column-observation
smile_subsample_segments1 = round(linspace(1,number_of_smile_samples,e+1)); % indices of subsample segmentation points
data_group{i}=sa;
smile_subsample_segments{i}=smile_subsample_segments1;
end
RHO=corr(data);
R=abs(RHO);
n=size(data,2)-1; % The number of features without class
CrossOverProb=0.9; % The Cross Over Probability
MutationProb=0.04; % The Mutation Probability
GenerationNo = 100; % The Number of generation
na = 50; % The Size of popuation
cc = 1;
maxrun=30; %run times
Best_fit=zeros(1,dnsize);
for run=1:maxrun
Best_fitness=zeros(1,dnsize);
for dnf=dn
m=dnf; % The number of selected features
for i=1:n
trail(i) = cc;
deltatrail(i) = 0;
end
maxnoiteration = 100;
k = 50;
p = 1;
rho = 0.75;
alpha=1;
beta=1;
counter=1; % The First iteration
%% Initialize population
for j=1:na
Select = randperm(n);
s(j,1:m)= Select(1:m);
end
popu=s;
fv=zeros(1,m);
%% Divide training set and test set and calculate fitness value with 10-fold cross-validation
for jdx=1:na
for i=1:e
data_ts=[];data_tr =[];
for j=1:length(class)
smile_subsample_segments1=smile_subsample_segments{j};
sa=data_group{j};
test= sa(smile_subsample_segments1(i):smile_subsample_segments1(i+1) , :); % current_test_smiles
data_ts=[test;data_ts] ; %training set
train = sa;
train(smile_subsample_segments1(i):smile_subsample_segments1(i+1),:) = [];
data_tr =[train;data_tr];%test set
end
val=s(jdx,:);
mdl = fitcknn(data_tr(:,val),data_tr(:,end),'NumNeighbors',4,'Standardize',1);%KNN-classifier
Ac1=predict(mdl,data_ts(:,val));
Fit(i)=sum(Ac1~=data_ts(:,end))/size(data_ts,1);
end
E(1,jdx)=mean(Fit);
end
[MinError,MinErrorIndex]=min(E);
[SO,IX]=sort(E); %Output Individual Optimality
for j=1:k
for l=1:m
for i=1:n
deltatrail(i)=0;
if (s(IX(j),l)==i)
Errorg = max(SO(1:k))-E(IX(j));
Errorhg=0;
for h=1:k
if ( (max(SO(1:k))-E(IX(h))) > Errorhg )
Errorhg = max(SO(1:k))-E(IX(h));
end
end
if (Errorhg ~= 0)
deltatrail(i)=Errorg/Errorhg;
else
deltatrail(i)=1;
end
end
trail(i)= rho * trail(i) + deltatrail(i);
end
end
end
%% looping
while ( (counter < maxnoiteration) & (MinError > 0.001) )
s=popu;
counter = counter+1;
for j=1:na
temp1(1:m)= s(IX(round(rand*(k-1))+1),1:m);
r=randperm(m);
for l=1:m-p
temp2(l)=temp1(r(l));
end
s(j,1:m-p)=temp2(1:m-p);
end
for mm = m-p+1:m
for j=1:na
for i=1:n
flag=0;
for l=1:m-p
if (s(j,l)==i)
flag=1;
end
end
if (flag==1)
USM(i)=0;
else
den=0;
for l=1:m-p
den=den+R(i,s(j,l));
end
if (den~=0)
LI(i)=R(i,n+1)/den;
else
LI(i)=R(i,n+1);
end
vis(i)=LI(i)^beta;
trail_p(i)=trail(i)^alpha;
sigma=0;
for ii=1:n
flag=0;
for l=1:m-p
if(s(j,l)==ii)
flag=1;
end
end
if (flag==0)
den=0;
for l=1:m-p
den=den+R(ii,s(j,l));
end
if (den~=0)
LI(ii)=R(ii,n+1)/den;
else
LI(ii)=R(ii,n+1);
end
sigma=sigma+ (trail(ii)^alpha) * (LI(ii)^beta);
end
end
USM(i)=(vis(i)*trail_p(i))/sigma;
end
end
[maxf,maxfindex]=max(USM);
s(j,mm)=maxfindex;
end
end
for j=1:na
flag=0;
for k1=1:m
for k2=k1+1:m
if(s(j,k1)==s(j,k2))
flag=1;
end
end
end
if (flag==1)
Select=randperm(n);
s(j,1:m)=Select(1:m);
end
end
for i=1:na
for j=i+1:na
if (s(i,1:m)==s(j,1:m))
Select = randperm(n);
s(j,1:m)= Select(1:m);
end
end
end
for jdx=1:na
for i=1:e
data_ts=[];data_tr =[];
for j=1:length(class)
smile_subsample_segments1=smile_subsample_segments{j};
sa=data_group{j};
test= sa(smile_subsample_segments1(i):smile_subsample_segments1(i+1) , :); % current_test_smiles
data_ts=[test;data_ts] ;
train = sa;
train(smile_subsample_segments1(i):smile_subsample_segments1(i+1),:) = [];
data_tr =[train;data_tr];
end
val=s(jdx,:);
mdl = fitcknn(data_tr(:,val),data_tr(:,end),'NumNeighbors',4,'Standardize',1);
Ac1=predict(mdl,data_ts(:,val));
Fit(i)=sum(Ac1~=data_ts(:,end))/size(data_ts,1);
end
fv(1,jdx)=mean(Fit);
if fv(1,jdx)<E(1,jdx) %Update optimal
E(1,jdx)=fv(1,jdx);
popu(jdx,:)=val;
end
end
[MinError,MinErrorIndex]=min(E);
[SO,IX]=sort(E);
fprintf('RUN: %d \t subsetsize: %d \t Iter: %d \t Err: %.4f \t \n',run,dnf,counter,MinError)
for j=1:k
for l=1:m
for i=1:n
deltatrail(i)=0;
if (s(IX(j),l)==i)
Errorg = max(SO(1:k))-E(IX(j));
Errorhg=0;
for h=1:k
if ( (max(SO(1:k))-E(IX(h))) > Errorhg )
Errorhg = max(SO(1:k))-E(IX(h));
end
end
if (Errorhg ~= 0)
deltatrail(i)=Errorg/Errorhg;
else
deltatrail(i)=1;
end
end
trail(i)= rho * trail(i) + deltatrail(i);
end
end
end
%for j=1:na
% temp1(1:m)= s(IX(round(rand*(k-1))+1),1:m);
% r=randperm(m);
% for l=1:m-p
% temp2(l)=temp1(r(l));
% end;
% s(j,1:m-p)=temp2(1:m-p);
%end;
end
Best_fitness(1,dnf/dnd)= MinError;
end
Best_fit(run,:)=Best_fitness;
end
%% Store the final results
BESTFIT=(1-mean(Best_fit))*100;
save('bsofs','Best_fit','BESTFIT')