Naive Bayes
The classifier that filtered spam for decades. Naive Bayes flips the question with Bayes' rule: instead of asking what class this email is, ask how likely each word would be if it were spam versus if it were normal mail. Every word is a piece of evidence that multiplies the odds — free and prize push toward spam, meeting and invoice pull away. The naive part is assuming words appear independently of each other, which is false, yet the classifier still works remarkably well because it only needs the final odds to land on the correct side.
Tap words to compose an email. Each one multiplies the spam odds up or down; the bar shows the running probability. Mix spammy and normal words and watch the evidence fight.
Build an email by tapping words:
verdict: SPAM
Each selected word multiplies the odds by how much more often it appears in spam than in normal mail. "Naive" = the model pretends the words are independent of each other; wrong, but it works remarkably well.
Check yourself
The word free also appears in plenty of honest emails. Why can the classifier still use it as evidence for spam?
Go deeper (free): StatQuest — Naive Bayes (video) ↗