Thresholds, ROC & AUC

A spam classifier doesn't output 'spam' — it outputs something like 0.73: probably spam. A threshold turns that score into a decision: 0.73 ≥ 0.5 → spam folder.

Move the line and the product changes:

• Lower the threshold → more spam caught (recall up), more real mail flagged (precision down)
• Raise it → a cleaner spam folder, but more junk reaching the inbox

The model didn't change. The standard of proof did.