How to create a confusion matrix.

How to create a confusion matrix for clustering? For this how can I get predicted class and actual class? And how can I get those TP, TN, FP and FN values from it? I am confused please help me.

Thanks in advance.

[Benjamin Großmann]

• Create confusion matrix with

conf_mat = confusion(targets,outputs);

• The predicted class (outputs) is the result of your net for a given input (after training), the "actual class" is your label (targets)
• We can now compare target and output for each sample and for one class "c" against all others:
• target == c & output == c --> TP (The output is positive and that is true)
• target == c & output ~= c --> FN (The output is negative but that is false)
• target ~= c & output == c --> FP (The output is positive but that is false)
• target ~= c & output ~= c --> TN (The output is negative and that is true)

[sreelekshmi ms]

I used confusionmat() also. How can I get the TP, TN, FP, FN values using a MatLab?

Answer 1

[Benjamin Großmann]

Lets use the cifar10 demo included in Matlab for your question

clearvars
close all
clc

load('Cifar10Labels.mat','trueLabels','predictedLabels');
[m,order] = confusionmat(trueLabels,predictedLabels);

figure
cm = confusionchart(m,order);

Please look at the confusionchart and consider the following for one particular class c,

• The diagonal element of class c is the amount of TP
• Everything inside the predicted class column except the diagonal element is falsely predicted as class c --> FP
• Everything inside the true class row except the diagonal element is of class c but not predicted as c --> FN
• Every other diagonal element except the diagonal element of class c itself is TN.

Now, we can put this easily in in code by summing up row, column and diagonal elements and substracting the TP. Lets pick a class, e.g. c=2 "automobile", and calculate TP, FP, FN, TN for that class

c = 2;

TP = cm.NormalizedValues(c,c)           % true class is c and predicted as c
FP = sum(cm.NormalizedValues(:,c))-TP   % predicted as c, true class is not c
FN = sum(cm.NormalizedValues(c,:))-TP   % true class is c, not predicted as c
TN = sum(diag(cm.NormalizedValues))-TP  % true class is not c, not predicted as c

[Benjamin Großmann]

In addition, an approach to get the confusion matrixes for a single class and therefor their TP,FP,FN,TN values

clearvars
close all
clc

load('Cifar10Labels.mat','trueLabels','predictedLabels');
[m,order] = confusionmat(trueLabels,predictedLabels);

figure
cm = confusionchart(m,order);

cmv = cm.NormalizedValues;

% confusion matrix for a single class
conf_mat_sc = @(c) [cmv(c,c), sum(cmv(:,c))-cmv(c,c);
                  sum(cmv(c,:))-cmv(c,c), sum(diag(cmv))-cmv(c,c);];

cms = arrayfun(conf_mat_sc, [1:numel(order)], 'UniformOutput', false);

[sreelekshmi ms]

Thank you Sir.

I have a doubt, based on the c value the confusion matrix will change. If I finding the accuracy

( A=(TP+TN)/(TP+TN+FP+FN)) based on the c value it will change also. From this how can I get a fixed accuracy?

[sreelekshmi ms]

Sir, I tried this in a different data set I got some errors like:

" Index in position 1 exceeds array bounds (must not exceed 2).

Error in shpa (line 68)

TP = cm.NormalizedValues(c,c) ; "

clc;
clear;

data=xlsread('Ecolixl.xlsx');
minpts=6;
epsilon=2;
[idx, corepts] = dbscan(data,epsilon,minpts);
nearEnough = 0.02; 
x = data(:,1);
y = data(:,2);
indexesToKeep = false(1, length(x)); 
for k = 1 : length(x)
    distances = sqrt((x(k) - x).^2 + (y(k) - y).^2);
    if sum(distances > nearEnough) >= 5
        
        indexesToKeep(k) = true;
    end    
end
x = x(indexesToKeep);
y = y(indexesToKeep);
P=[x y];
  dist2 = (data(:,1) - P(:,1).').^2 + (data(:,2) - P(:,2).').^2;
          [~,id] = mink(distances,20,1);
          clusters = data(id);
    
          maximum_num_clusters = 7;
 
Z = linkage(clusters, 'average');
id= cluster(Z, 'Maxclust', maximum_num_clusters);
 
          figure()
          dendrogram(Z)
          uni=length(Z);
          outl=rmoutliers(Z);
          I=data(:,4);
      X= -ones(748,1);
    B=[-ones(length(I) - length(id),1);id];
    Si=[id;X];
% load('Cifar10Labels.mat','trueLabels','predictedLabels');
[m,order] = confusionmat(I,idx);

figure
cm = confusionchart(m,order);
c = 3;

TP = cm.NormalizedValues(c,c) ;          
FP = sum(cm.NormalizedValues(:,c))-TP ;  
FN = sum(cm.NormalizedValues(c,:))-TP ;  
TN = sum(diag(cm.NormalizedValues))-TP;  
      A=(TP+TN)/(TP+TN+FP+FN)*100;

How can I get this? Please help me.