Why can an unsmoothed zero class-conditional word count destroy a naive Bayes document score?
Classic MLLevel: IntermediateSmoothing unseen events
Why can an unsmoothed zero class-conditional word count destroy a naive Bayes document score?
Hints
- Follow one zero through a product.
- Pseudocounts modify numerator and normalization.
Answer
The likelihood multiplies feature probabilities; a zero factor annihilates the product. Additive smoothing allocates mass to unseen events, with strength controlling the prior's influence.
Source and licence: scikit-learn 1.7.2 — naive bayes · BSD-3-Clause · answer: Quantoros editorial