function lez_weka()
% do it only one time
%javaaddpath('/Users/snrxe/PHD/VideogamesWk/weka-3-6-1/weka.jar')
close all;


%Generate simulated data
% TWO class problem 
%    class1 N(M1, S1)
%    class2 N(M2, S2)

S1= [1,0.9;
    0.9,2];
S2 = [6,0.2;
     0.2,2];
M1 = [5,8];
M2 = [2,7];

% number of sample fro each class N= 2* n total number of sample
n= 100;
        
X = [mvnrnd(M1,S1,n);
      mvnrnd(M2,S2,n)];
  
C = [zeros(n,1)+1;zeros(n,1)+2];

% plotting distribution of data
plot_distribution2(X,[1,2],C);


% K-fold cross-validation ---------------------------------

K = 10; 
dim = 2*n/10;

for k=1:K;
    % selct training data and validation data
    
    %train selection
    val_sel = zeros(1,2*n);
    val_sel(dim * (k-1)+1: dim*k) = 1;
    val_sel=logical(val_sel);
    
    %validation selection
    train_sel = ~val_sel;
    
    Xtr = X(train_sel,:);
    Ctr = C(train_sel);
    
    Xval = X(val_sel,:);
    Cval = C(val_sel);
    
    
    % define a classifier K-NN K=5 (NOTE this K is totally differen from the number K of folds!!!!)
    classifier =  {'weka.classifiers.lazy.IBk', {'-K','5','-W','0'}};
    classifier2 =  {'weka.classifiers.lazy.IBk', {'-K','5','-W','0'}};
    
    % build dataset in Weka format
    train = matlab2weka('tr', {'feat1','feat2','class'}, [Xtr, Ctr]);
    validation  = matlab2weka('vl', {'feat1','feat2','class'}, [Xval, Cval]);
    
    %train classifier on training data
    [cl] = train_weka_classif_affective( train, [], classifier{1,1}, classifier{1,2} );
    [cl2] = train_weka_classif_affective( train, [], classifier2{1,1}, classifier2{1,2} );
    
    
    %Evaluate the classifier on validation data
    %confusion matrix m
    m = zeros(2,2);
    m2 = zeros(2,2);
    for t = 0:validation.numInstances -1 
            class_dist(t+1,:) = cl.distributionForInstance(validation.instance(t));
            pred = round(class_dist(t+1));
            m(Cval(t+1),pred) =  m(Cval(t+1),pred)+1;
            
            
            class_dist2(t+1,:) = cl2.distributionForInstance(validation.instance(t));
            pred2 = round(class_dist2(t+1));
            m2(Cval(t+1),pred2) =  m2(Cval(t+1),pred2)+1;
    end
    
    % compute perfromance as Correct classification rate
    ccr(k) = sum(diag(m)/dim);   
    ccr2(k) = sum(diag(m2)/dim);   
end

n=2*n;

% Evaluate confidence intervals ccr(k)
xbar = mean(ccr); % ccr mean
s = std(ccr); % ccr standard deviation

gamma = 0.95; % P(-A< T < A) =0.95 
t = tinv((1+gamma)/2,n-1); %correspond to  P(T <= A) =(1+gamma)/2


lb = xbar-t*s/sqrt(n);
ub = xbar+t*s/sqrt(n);

fprintf('Final performance classifier 1: %f <= CCR <= %f\n', lb, ub);


% Evaluate confidence intervals ccr2(k)
xbar = mean(ccr2); % ccr mean
s = std(ccr2); % ccr standard deviation

gamma = 0.95; % P(-A< T < A) =0.95 
t = tinv((1+gamma)/2,n-1); %correspond to  P(T <= A) =(1+gamma)/2


lb = xbar-t*s/sqrt(n);
ub = xbar+t*s/sqrt(n);


fprintf('Final performance classifier 2: %f <= CCR <= %f\n', lb, ub);



%%Are the two classifier different?

xbar = mean(ccr)
ybar = mean(ccr2)
sx = var(ccr);
sy = var(ccr2);

%significance of test
alpha = 0.05;

t = tinv(1-alpha/2,n-1);
s = sqrt((sx+sy)/n);

fprintf('t=%.3f\n', t);
fprintf('s=%.3f\n', s);
fprintf('xbar-ybar = %f\n', xbar-ybar); 
lb = -t*sqrt(s/n);
ub = t*sqrt(s/n);

fprintf('IC = (%.3f,%.3f)\n', lb,ub);



if (abs(xbar-ybar)/sqrt(s/n)>= t)
    fprintf('H_0 rejected, Classifier are not equal!\n')
else
    fprintf('H_0 accepted, Classifier are equal \n')
end