How should I implement a confusion matrix in Matlab?

Question

I know what is a confusion matrix.

Given N classes we have a NxN matrix M where - each row is one of the classes - each column is one of the classes

M(X,Y) = number of elements which have been classified in class X and should have classified in class Y (of course if X=Y the classification is correct).

Now I have a set of "compound classes" and a variable number of classes

• with "COMPOUND" I mean that one class corresponds to a vector (of variable length). For example if the class is represented by a 2 element vector,the possible classes will be:

  [0,0]

  [0,1]

  [1,0]

  [1,1]

Mi goal is to define a function with:

• INPUT: a NxM matrix (where N=number of classified elements,M=number of elements of each class) of classified elements , a NxM matrix of expected class for each element.
• OUTPUT: NXN confusion matrix corresponding to input classes

Example of OUTPUT:

[1 0 0]
[0 2 0]
[0 1 0]

This matrix is 3x3 matrix.

This means the the total number of classes is 3 and number of classified elements are 4:

• one element should have been classified in class 1 and has been classified in class 1
• two elements should have been classified in class 2 and have been classified in class 2
• one element should have been classified in class 3 and has been classified in class 2

(it doesn't matter how the elements are composed of. Let's image the classes may be:

[0,0,1]
[1,0,1]
[1,1,1]

** Example of input matrix (EXPECTEDclasses): **

[0,0,1]
[1,0,1]
[1,0,1]
[1,1,1]

** Example of input matrix (OBTAINED classes): **

[0,0,1]
[1,0,1]
[1,0,1]
[1,0,1]   <-- this is the element(line) incorrectly classified

How should I do? (the classification is made by a Perceptron or Adaline neuronal network)

Thank you in advance for any hint!

Answer 1

The key is to use ismember in the following form: [Lia,Locb] = ismember(A,B,'rows') The second output argument tells you which row of matrix B that each row of matrix A matches. Use this information to construct the confusion matrix (which is a built-in function in the stats toolbox).

classes = [0 0 1;
            1 0 1; 
            1 1 1];
 expected = [0 0 1;
             1 0 1;
             1 0 1;
             1 1 1];
 obtained = [0 0 1;
             1 0 1;
             1 0 1;
             1 0 1];    
>> [~, ex] = ismember(expected, classes,'rows')    
ex =    
     1
     2
     2
     3

>> [~, ob] = ismember(obtained, classes,'rows')    
ob =

     1
     2
     2
     2

>> confusionmat(ex,ob,'order',[1 2 3])    
ans =

     1     0     0
     0     2     0
     0     1     0

If you don't have access to confusionmat, you can use accummarray to do it yourself.

num_observations = length(ex);
num_classes = size(classes,1);
accumarray([ex,ob],ones(num_observations,1),[num_classes,num_classes])