# Confusion Matrix (Contingency Table)

## Confusion Matrix

It is much easier to understand confusion matrix if we use a binary classification problem as an example. For example, we have a bunch of cat photos and the user labelled “cute or not” data. Now we are using the labeled data to train a cute-or-not binary classifier.

Then we apply the classifier on the test dataset and we would only find four different kind of results.

Labeled as Cute | Labeled as Not Cute | |
---|---|---|

Classifier Predicted to be Cute | True Positive (TP) | False Positive (FP) |

Classifier Predicted to be Not Cute | False Negative (FN) | True Negative (TN) |

This table is easy enough to comprehend. We have discussed the Type I and Type II errors in Types of Errors in Statistical Hypothesis Testing . here False Positive (FP) is Type I error and False Negative (FN) is Type II error.

Isn't FN type I error?

Then we loop through all cats in the test dataset, find the results and put the numbers in the table.

There are a few things to look at whenever we have a table. First things first, we would like to know the sum of each row and columns.

- The sum of the row “Classifier Predicted to be Cute” tells us the total number of cats classified as cute, aka,
**Total Predicted Positives**(**PP**); - The sum of the row “Classifier Predicted to be Not Cute” tells us the total number of cats classified as not cute, aka,
**Total Predicted Negatives**(**PN**); - The sum of the column “Labeled as Cute” tells us the total number of cats labeled as cute in the test dataset, aka,
**Total Positives**(**P**); - The sum of the column “Labeled as Not Cute” tells us the total number of cats labeled as not cute in the test dataset, aka,
**Total Negatives**(**N**). - The sum of the diagonal elements tells us the total number of case where the classifier classified the data correctly.

## Measures Defined

### Property of the Dataset itself

Apart from these trivial properties about the test dataset defined above, we can define the **Prevalence**:

### Performance of Classifier

We could define some quite general measures.

**Accuracy**:

Now we recalculate the confusion matrix by dividing the values by some certain sums.

#### Confusion Matrix divided by the Column Sums

As mentioned in the previous sections, the sum of the columns are **P** (**Total Positives**) and **N** (**Total Negatives**). Dividing each column by the sum of the corresponding column gives us the prediction rate for each labels.

Labeled as Cute | Labeled as Not Cute | |
---|---|---|

Classifier Predicted to be Cute | True Positive Rate (aka Recall) = TP/P | False Positive Rate = FP/N |

Classifier Predicted to be Not Cute | False Negative Rate = FN/P | True Negative Rate (aka Specifity) = TN/N |

There are a few names to be emphasized.

**Recall**: True Positive Rate

#### Confusion Matrix divided by the Row Sums

As mentioned, the sum of the rows indicates the

Labeled as Cute | Labeled as Not Cute | |
---|---|---|

Classifier Predicted to be Cute | Positive Predictive Value (aka Precision) = TP/PP | False Discovery Rate = FP/PP |

Classifier Predicted to be Not Cute | False Omission Rate = FN/PN | Negative Predictive Value = TN/PN |

There are a few names to be emphasized.

**Precision**: Positive Predictive Value

#### Ratios, Scores, and More

We also have some other definitions of ratios, please refer to the bottom left corner of the table on the wikipedia page linked in the references.

We will only define the F1 score ($\mathrm F_1$) here. As a F-measure,

## Confused by the Names?

There is a nice chart on this wikipedia page.