Two paradigms for building classifiers: discriminative models learn the boundary between classes directly, while generative models learn how each class produces data and infer the boundary from that.

Definition

  • Discriminative models learn directly — given features, what is the probability of each class? They focus on what distinguishes the classes from each other.
  • Generative models learn — how likely is this data given each class, and how common is each class? Classification then follows from Bayes’ rule: .

Comparison

DiscriminativeGenerative
Models and
GoalFind the boundaryModel the data-generating process
Examplelogistic-regressionNaive Bayes, Bayesian classifiers
InterpretabilityWeights show which features discriminateCan generate synthetic examples

When Each Is Used

Discriminative models tend to achieve higher classification accuracy when training data is plentiful, because they focus all modelling capacity on the decision boundary. Generative models are more useful when data is scarce (priors help), when you need to detect outliers, or when understanding the data distribution itself is the goal.

This distinction becomes concrete in weeks 7–8 when Bayesian methods are introduced as generative counterparts to the discriminative classifiers covered in weeks 1–5.

Active Recall