Naive Bayes
A classifier that applies Bayes' theorem with one wildly unrealistic assumption — and works embarrassingly well anyway, especially on text.
Reviewed July 16, 2026Stable
When not to use it
- When feature dependence carries the signal you need — the independence assumption throws exactly that away.
- When you need well-calibrated probabilities — its estimates are often pushed toward 0 or 1.
- On problems where interactions between features are the whole point.
Reach for something else instead
- Logistic regression — the discriminative counterpart, often better-calibrated.
- Gradient boosting when accuracy matters more than speed and interpretability.
- Transformers for text when you have the data and compute and need the ceiling raised.
Sources & further reading
- Ng & Jordan (2001), On Discriminative vs. Generative Classifiers — the definitive naive-Bayes-vs-logistic-regression comparison.
- Manning, Raghavan & Schütze, Introduction to Information Retrieval — naive Bayes for text classification.
- Rish (2001), An Empirical Study of the Naive Bayes Classifier — why it works despite the assumption.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Forgetting Laplace smoothing, so an unseen feature zeroes out a whole class probability.
- Multiplying raw probabilities instead of summing logs, causing numerical underflow.
- Trusting its probability estimates as calibrated — trust the ranking, not the magnitude.
At a glance
FieldMachine Learning
Based onBayes' theorem + independence assumption
Best attext classification
Superpowerfast, strong baseline
DifficultyBeginner