MATLAB Answers

0

How can I plot a confusion matrix for a multi-class or non-binary classification problem?

I want to make a plot similar to the confusion matrix created in the Classification Learner app. This can make a confusion matrix for a multi-class or non-binary classification problem. In addition, it can plot things such as a True Positive or False Negative rates.
How can I do this?

2 Answers

Answer by MathWorks Support Team on 5 Jul 2017
 Accepted Answer

Similar to the binary or two-class problem, this can be done using the "plotconfusion" function. By default, this command will also plot the True Positive, False Negative, Positive Predictive, and False Discovery rates in they grey-colored boxes. Please refer to the following example:
targetsVector = [1 2 1 1 3 2]; % True classes
outputsVector = [1 3 1 2 3 1]; % Predicted classes
% Convert this data to a [numClasses x 6] matrix
targets = zeros(3,6);
outputs = zeros(3,6);
targetsIdx = sub2ind(size(targets), targetsVector, 1:6);
outputsIdx = sub2ind(size(outputs), outputsVector, 1:6);
targets(targetsIdx) = 1;
outputs(outputsIdx) = 1;
% Plot the confusion matrix for a 3-class problem
plotconfusion(targets,outputs)
The class labels can be customized by setting that 'XTickLabel' and 'YTickLabel' properties of the axis:
h = gca;
h.XTickLabel = {'Class A','Class B','Class C',''};
h.YTickLabel = {'Class A','Class B','Class C',''};
h.YTickLabelRotation = 90;

  3 Comments

targetsVec may have input data set with predefined classes(True classes)?
The classification learner app plots the figure with proper class labels (i.e strings too). Is it possible to do the same with plotconfusion?
On the other hand given a vector of targets and outputs (i.e. of the cell type with strings for class labels) the confusionmat command creates a confusion matrix; how can we plot that matrix like MATLAB does? I suspect it should be possible since internally plotconfusion should generate the confusion matrix at some point.
I have updated the above answer to better indicate that the 'TargetsVector' contains the true class labels.
I also included a quick example in the answer showing how to add strings as a name for each class, as I think that is a great easy way to make the plot more easily interpretable

Sign in to comment.


Answer by David Franco on 23 Jan 2018
Edited by MathWorks Support Team on 16 Mar 2018

Implementation code:
Confusion Matrix
function [] = confusion_matrix(T,Y)
M = size(unique(T),2);
N = size(T,2);
targets = zeros(M,N);
outputs = zeros(M,N);
targetsIdx = sub2ind(size(targets), T, 1:N);
outputsIdx = sub2ind(size(outputs), Y, 1:N);
targets(targetsIdx) = 1;
outputs(outputsIdx) = 1;
% Plot the confusion matrix
plotconfusion(targets,outputs)

  2 Comments

is the plotconfusion function available in the normal versions of MATLAB or does it need to have some special toolbox to be attached to the normal version. I have a normal version and it throws an error at the line where the function is written & the error says that undefined funciton.
what are T and Y?
is plotconfusion(targets,outputs) part of the function?

Sign in to comment.