Generative vs discriminative models comparison table

What's actually being asked

A structured, point-by-point comparison of generative and discriminative models across six named dimensions, with algorithm examples.

Answer

Both paradigms are solving the same underlying classification/regression problem, but from opposite ends: one builds a full model of how the data could have been produced and reads a prediction off it, the other fits the prediction rule straight from data with no detour through "how was this data made."

Dimension	Generative Model	Discriminative Model
Learning objective	Learn how the data itself is produced — model P(x,y) (or P(x|y) and P(y)), then derive P(y|x) via Bayes' theorem	Learn the decision rule directly — model P(y|x) (or just a direct mapping x → y) without modeling how x arises
Probability distribution modeled	Full joint P(x,y), which implicitly also gives P(x) = Σ_y P(x|y)P(y) — so the model can generate new x	Only the conditional P(y|x); no model of P(x), so it cannot generate realistic new x
Training process	Estimate class priors P(y) and class-conditional likelihoods P(x|y) from data (e.g., fit per-class Gaussians)	Directly optimize a loss (e.g., cross-entropy, hinge loss) that penalizes wrong y predictions given x, via gradient descent
Decision boundary	Implicit — derived indirectly from comparing P(x|y)P(y) across classes; tends to be smoother, shaped by distributional assumptions	Learned explicitly and directly to separate classes; can be tightly fit to the actual class boundary in the training data
Data requirements	Can often work with less labeled data (also models P(x), so unlabeled data helps); more sensitive to distributional assumptions being wrong	Usually needs more labeled data for comparable accuracy, but makes fewer assumptions about how the input data is distributed
Applications	Generation tasks: image/text synthesis, data augmentation, outlier/anomaly detection via density estimation	Pure classification/regression tasks where only a label is needed: spam filtering, medical diagnosis prediction, risk scoring
Algorithm examples	Naïve Bayes, Gaussian Mixture Models, Hidden Markov Models, VAEs, GANs	Logistic Regression, Support Vector Machines, standard feed-forward Neural Networks, Conditional Random Fields

The key one-liner

Generative models learn "how the data was made" and can therefore generate new samples; discriminative models learn "how to tell classes apart" and are usually more accurate at that narrow task when enough labeled data is available, but can't generate anything. Neither is universally "better" — the right choice depends on whether the end goal is classification accuracy alone or classification plus the ability to generate/explain the data.

Exam tip

This question is almost always graded against a table like this — don't write it as prose only. Mentioning that both are connected by P(x,y) = P(y|x)P(x) = P(x|y)P(y) in the "learning objective" row shows you understand why a generative model can derive a discriminative-style prediction but not vice versa.