Confusion matrix

A confusion matrix is a table that counts every prediction by what the model said versus what was actually true.

A single score hides the story. A confusion matrix is the itemised bill: it shows not only how often the model was wrong, but in which direction. For a yes/no problem it has four cells.

                      actually spam    actually not spam
predicted spam             412                  9        ← 9 real emails lost
predicted not spam          31                548        ← 31 spam got through

The diagonal (412 and 548) is everything the model got right. The other two cells are the two different mistakes: false positives (top right) and false negatives (bottom left). Every headline metric is a ratio of these four numbers — accuracy, precision and recall each slice them differently.

With more than two classes, the matrix grows, and its off-diagonal cells become a diagnosis tool. If a digit classifier keeps confusing 4 with 9, one bright off-diagonal cell says so instantly. That tells you what extra training data to collect, which no single number can.

Where to go next#

• Full lesson: Model evaluation
• Related terms: precision, recall, accuracy, f1-score