Accuracy can hide important mistakes. A confusion matrix helps students see what kind of errors a classifier makes, not just how many predictions were correct.
Scenario: school email filter
A student builds a simple classifier that labels messages as "club announcement" or "not club announcement." On a test set of 100 messages, the results are:
| Predicted club | Predicted not club | |
|---|---|---|
| Actually club | 28 | 12 |
| Actually not club | 8 | 52 |
Interpreting the cells
- True positives: 28 club announcements correctly detected.
- False negatives: 12 club announcements missed.
- False positives: 8 non-club messages incorrectly flagged.
- True negatives: 52 non-club messages correctly ignored.
Why the interpretation matters
If the goal is to avoid missing important announcements, false negatives matter more. If the goal is to avoid cluttering a club feed, false positives matter more. Model evaluation depends on the real use case.
Practice prompt
Compute accuracy from the table. Then write a short recommendation: should the student tune the model to catch more club messages or to avoid false alarms?