How to read a confusion matrix

Turn a table of predictions into actionable precision, recall, and error patterns.

1. **Rows = true label, columns = predicted label** — locate true positives (TP), false positives (FP), false negatives (FN), true negatives (TN)

2. **Ask the business question** — is a false alarm or a missed case more costly?

3. **Compute metrics**

4. **Check imbalance** — high accuracy can hide useless models on rare classes

5. **Slice errors** — group FPs/FNs by segment (language, product, time) for fixes

| | Pred spam | Pred not spam |

|--|-----------|---------------|

| True spam | TP | FN ← missed spam |

| True not spam | FP ← false alarm | TN |

One number never tells the story — pair the matrix with **threshold tuning** and **class weights**.

**Try the lessons:** `prediction-game` in Lane D · [Glossary: confusion matrix](/reference/confusion-matrix)