All MCQs

Multiplying many small probabilities can produce numbers too tiny for reliable floating-point representation. Logarithms preserve class ranking while allowing stable addition.

Laplace smoothing adds pseudo-counts before estimating probabilities. This prevents unseen events from forcing a class likelihood to zero.

Multiplying by an estimated probability of zero makes the entire product zero. This zero-frequency problem can eliminate a class based on one unseen feature.

Bernoulli Naive Bayes models binary feature outcomes such as 1 for present and 0 for absent. It explicitly uses both feature presence and absence probabilities.

Multinomial Naive Bayes models discrete counts such as word frequencies within each class. It is widely used for document and text classification.

Gaussian Naive Bayes estimates a mean and variance for each continuous feature within each class. Those normal densities provide the feature likelihoods.

MAP classification chooses the class most probable after combining priors and observed evidence. It directly compares posterior values across candidate classes.

The posterior updates the prior using evidence supplied by the feature values. It is the quantity used to compare candidate classes for a particular case.

This conditional probability measures how compatible a feature is with the spam class. Naive Bayes combines such likelihoods across observed words.

A prior represents the baseline prevalence or belief for each class. Observed features then update that baseline through Bayes’ theorem.

The model factorizes the joint feature likelihood into separate feature contributions within each class. This simplifying assumption is often unrealistic but computationally useful.

Naive Bayes combines class priors with feature likelihoods to calculate posterior class probabilities. The class with the strongest posterior support is typically predicted.