Discriminative and generative models are different approaches to compute classifiers. Note that the distinction between both approaches is not always clear in the literature.

Discriminative methods:

• aim at estimating $\mathbb{P}(Y | X=x)$ where $Y$ the target variable and $X$ is the observed data.

• does not look at the data distribution.

• examples: Linear Regression (supervised), Decision Trees (supervised), SVM (supervised).

Generative methods:

• consist in learning an underlying distribution.

• aim at estimating the joint distribution $\mathbb{P}(X, Y)$ or the posterior $\mathbb{P}(X | Y=y)$ where $X$ is the target variable and $Y$ is the observed variable.

• try to answer the question: “How were the data generated in order to categorize a signal?”.

• examples: Latent Dirichlet Allocation (unsupervised), Naive Bayes (supervised), Linear Discriminant Analysis (unsupervised).