Training corpus

Spam docs—

Ham docs—

Vocabulary size—

Add training example

Email text

Classifier settings

Classification threshold: 50% Stream speed: 1.0×

Live stats

Held-out accuracy—

Stream classified0

Stream correct0

Stream accuracy—

Predicted spam0

Predicted ham0

Real multinomial Naive Bayes: P(spam|words) ∝ P(spam) · ∏ P(word|spam). Word likelihoods use Laplace (add-one) smoothing so unseen words never zero out a probability. The log-odds sum of per-word log-likelihood ratios plus the log-prior is converted back to a probability with the sigmoid function.