Banso D. Wisdom

Posted on May 5, 2019

Understanding the Confusion Matrix (II)

#python #machinelearning #deeplearning #ai

In the first part of this article, I talked about the confusion matrix in general, the 2-class confusion matrix, how to calculate accuracy, precision and other metrics using it and also how to generate a confusion matrix in python.

In this article, we'll be looking at the multi-class confusion matrix.

What is the multi-class confusion matrix?

As the name implies, it is a confusion matrix that deals with multiple classes (i.e. more than 2 classes). Just like the 2-class confusion matrix, it describes the performance of a multi-class classification model.

For the purpose of this article, we'll be assuming that our multi-class classification model is one that classifies images of dogs into the following breeds: Greyhound, Mastiff and Samoyed.

A confusion matrix for this classifier can be visualized as such:

In this visualization, we have two sections which have been outlined. We have the predicted classifications section which contains three subsections for each of the classes we want to classify into and the actual classifications section which has three subsections for each of the classes.

Having visualized this confusion matrix, we can use this visualization to calculate the following metrics:

True Positives.
True Negatives.
False Positives.
False Negatives.
Accuracy.
Precision.
True Positive Rate is also known as Sensitivity or Recall.
True Negative Rate is also known as Specificity.

Before we calculate these metrics, let's define the variables in the visualization:

PGG
This variable represents the number of predictions where images of a Greyhound were correctly classified [as a Greyhound]. This is also the True Positive for the Greyhound class.

PMG
This variable represents the number of predictions where images of a Greyhound were incorrectly classified as a Mastiff.

PSG
This variable represents the number of predictions where images of a Greyhound were incorrectly classified as a Samoyed.

PGM
This variable represents the number of predictions where images of a Mastiff were incorrectly classified as a Greyhound.

PMM
This variable represents the number of predictions where images of a Mastiff were correctly classified [as a Mastiff]. This is also the True Positive for the Mastiff class.

PSM
This variable represents the number of predictions where images of a Mastiff were incorrectly classified as a Samoyed.

PGS
This variable represents the number of predictions where images of a Samoyed were incorrectly classified as a Greyhound.

PMS
This variable represents the number of predictions where images of a Samoyed were incorrectly classified as a Mastiff.

PSS
This variable represents the number of predictions where images of a Samoyed were correctly classified [as a Samoyed]. This is also the True Positive for the Samoyed class.

Now that we have defined these variables, we can now calculate the aforementioned metrics.

True Positives

The definition for the True Positive is the same as in the 2-class confusion matrix. However, here we calculate the True Positives for each class in the confusion matrix unlike the general or absolute True Positives in the 2-class confusion matrix.

The True Positives is the number of predictions where data labelled to belong to a particular class was correctly classified as the said class. E.g Number of predictions where images of a Samoyed was correctly classified as a Samoyed.

From the definition of the matrix variables, we have already identified the True Positives for each of the classes:

True Positives for the Greyhound class is the variable PGG in the confusion matrix.

True Positives for the Mastiff class is the variable PMM in the confusion matrix.

True Positives for the Samoyed class is the variable PSS in the confusion matrix.

True Negatives

The definition of the True Negative is the same as in the 2-class confusion matrix. Here we calculate the True Negatives for each class in the confusion matrix unlike the general or absolute True Negatives in the 2-class confusion matrix.

The True Negatives for a particular class is calculated by taking the sum of the values in every row and column except the row and column of the class we're trying to find the True Negatives for.

For example, calculating the True Negatives for the Greyhound class:

We omit the row and columns belonging to the Greyhound class and sum the variables that are left, which are the rows and columns of the other classes (Mastiff and Samoyed).

Therefore the True Negatives for the Greyhound class is:

TN = PMM + PSM + PMS + PSS

Similarly, we can calculate the True Negatives for the other classes.

Calculating the True Negatives for the Mastiff class:

The True Negatives for the Mastiff class is:

TN = PGG + PSG + PGS + PSS

Calculating the True Negatives for the Samoyed class:

The True Negatives for the Samoyed class is:

TN = PGG + PMG + PGM + PMM

False Positives

The definition of the False Positive is the same as in the 2-class confusion matrix. Here we calculate the False Positives for each class in the confusion matrix unlike the general or absolute False Positives in the 2-class confusion matrix.

The False Positives for a particular class can be calculated by taking the sum of all the values in the column corresponding to that class except the True Positives value.

For example, calculating the False Positives for the Greyhound class:

We sum all the values in the highlighted area, which is the column corresponding to the Greyhound class with the exception of the variable PGG which we had earlier identified to be the True Positives for the Greyhound class.

Therefore the False Positives for Greyhound class is:

FP = PGM + PGS

Similarly, we can calculate the False Positives for the other classes.

Calculating the False Positives for the Mastiff class:

The False Positives for the Mastiff class is:

FP = PMG + PMS

Calculating the False Positives for the Samoyed class:

The False Positives for the Samoyed class is:

FP = PSG + PSM

False Negatives

The definition of the False Negative is the same as in the 2-class confusion matrix. Here we calculate the False Negatives for each class in the confusion matrix unlike the general or absolute False Positives in the 2-class confusion matrix.

The False Negatives for a particular class can be calculated by taking the sum of all the values in the row corresponding to that class except the True Positives value.

For example, calculating the False Negatives for the Greyhound class:

We sum all the values in the highlighted area, which is the row corresponding to the Greyhound class with the exception of the variable PGG which we had earlier identified to be the True Positives for the Greyhound class.

Therefore the False Negatives for Greyhound class is:

FN = PMG + PSG

Similarly, we can calculate the False Negatives for the other classes.

Calculating the False Negatives for the Mastiff class:

The False Negatives for the Mastiff class is:

FN = PGM + PSM

Calculating the False Negatives for the Samoyed class:

The False Negatives for the Samoyed class is:

FN = PGS + PMS

Accuracy

Accuracy is calculated as the ratio of the number of correct classifications to the total number of classifications. From our confusion matrix, the correct classifications are the True Positives for each class and the total number of classifications is the sum of every value in the confusion matrix, including the True Positives.

Therefore, the accuracy is:

AC = (PGG + PMM + PSS) / (PGG + PMG + PSG + PGM + PMM + PSM + PGS + PMS + PSS)

Precision

Precision is a multi-class confusion matrix is the measure of the accuracy relative to the prediction of a specific class. It is calculated as the ratio of the True Positives of the class in question to the sum of its True Positives and False Positives.

For example, calculating the Precision of the Greyhound class:

Precision (G) = TP / (TP + FP)
= PGG / (PGG + (PGM + PGS))

Similarly, we can calculate the Precision of other classes.

Calculating the Precision of the Mastiff class:

Precision (M) = TP / (TP + FP)
= PMM / (PMM + (PMG + PMS))

Calculating the Precision of the Samoyed class:

Precision (S) = TP / (TP + FP)
= PSS / (PSS + (PSG + PSM))

True Positive Rate

The True Positive Rate (also known as Recall or Sensitivity) is calculated as the ratio of the True Positives of a specific class to the sum of its True Positives and False Negatives.

For example, calculating the True Positive Rate of the Greyhound class:

TPR (G) = TP / (TP + FN)
= PGG / (PGG + (PMG + PSG))

Similarly, we can calculate the True Positive Rate of other classes.

Calculating the True Positive Rate of the Mastiff class:

TPR (M) = TP / (TP + FN)
= PMM / (PMM + (PGM + PSM))

Calculating the True Positive Rate of the Samoyed class:

TPR (S) = TP / (TP + FN)
= PSS / (PSS + (PGS + PMS))

True Negative Rate

The True Negative Rate (also known as Specificity) is calculated as the ratio of the True Negatives of a specific class to the sum of its True Negatives and False Positives.

For example, calculating the True Negative Rate of the Greyhound class:

TNR (G) = TN / (TN + FP)
= (PMM + PSM + PMS + PSS) / ((PMM + PSM + PMS + PSS) + (PGM + PGS))

Similarly, we can calculate the True Negative Rate of other classes.

Calculating the True Negative Rate of the Mastiff class:

TNR (M) = TN / (TN + FP)
= (PGG + PSG + PGS + PSS) / ((PGG + PSG + PGS + PSS) + (PMG + PMS))

Calculating the True Negative Rate of the Samoyed class:

TNR (S) = TN / (TN + FP)
= (PGG + PMG + PGM + PMM) / ((PGG + PMG + PGM + PMM) + (PSG + PSM))

Phewww, that was a lot. 😅😅

Now, a quick example

Suppose we have the image below as the visualized confusion matrix for our classifier, we can use the information in the visualization and the metrics defined above to evaluate its performance.

From this confusion matrix, we can identify that:

The variable PGG and the True Positives of the Greyhound class is 250.

The variable PMM and the True Positives of the Mastiff class is 320.

The variable PSS and the True Positives of the Samoyed class is 180.

The variable PMG is 25.

The variable PSG is 18.

The variable PGM is 21.

The variable PSM is 24.

The variable PGS is 22.

The variable PMS is 12.

Using this information that we've "extracted", we can calculate the metrics mentioned earlier and thus evaluate the performance of the classifier.