Getting the accurate percent of the predicted classes using predictFCN

Dear All

I am using MATLAB 2022 to train a dataset that has two classes 0 and 1.

First of all, I used the classificationLearner then I exported the model to get the predicted classes of the test set.

The predicted classes were either 0 or 1 using predictFCN.

Is there a way to get the percent of the predicted class instead of getting them as integers please?.

I mean for example: if possible to get a prediction of 0.01 for class 0 instead of just 0 and a prediction of 0.89 for the class 1.

Hope I could explain my inquiry.

Thanks in advance.

Answer 1

[Drew]

Editing this answer based on the comments below:

Ok, so it sounds like, for each test observation, you want the score of each output class, and you want those output class scores to be in the form of a probability.

Two steps:

(1) To get the scores, see https://www.mathworks.com/help/stats/export-classification-model-for-use-with-new-data.html#bu4764j-1

[yfit,scores] = C.predictFcn(T)

(2) Whether those scores represent probabilities depends on the type of classifier you have trained, and some other settings. See https://www.mathworks.com/matlabcentral/answers/395526-how-do-i-obtain-scores-as-probabilistic-estimates-using-the-predict-function-on-a-fitcensemble-model for more info for ensembles. Or, refer to the doc pages for the classifier that you have trained.

The initial answer: This appears to not be what you were asking, but if you want to get measures of accuracy over the test set on a per-class basis, you can use the confusionchart command https://www.mathworks.com/help/stats/confusionchart.html. You can use use the "ColumnSummary", "RowSummary" and "Normalization" name-value arguments to confusionchart to get absoluate counts or percentages with your choice of normalization.

[Lama Moualla]

Thanks Drew. I ment that I need the probability of each single predicted value in decimals. In other words instead of getting a probability of 0 for any value I want this probability to be in decimals, like it could be 0.01 or 0.0003...etc.

[Drew]

Ok, so it sounds like, for each test observation, you want the score of each output class, and you want those output class scores to be in the form of a probability.

Two steps:

(1) To get the scores, see https://www.mathworks.com/help/stats/export-classification-model-for-use-with-new-data.html#bu4764j-1

[yfit,scores] = C.predictFcn(T)

(2) Whether those scores represent probabilities depends on the type of classifier you have trained, and some other settings. See https://www.mathworks.com/matlabcentral/answers/395526-how-do-i-obtain-scores-as-probabilistic-estimates-using-the-predict-function-on-a-fitcensemble-model for more info for ensembles. Or, refer to the doc pages for the classifier that you have trained.

[Lama Moualla]

This answer is super fast and perfect!

THANKS A LOT