In discriminative models to predict the label y from the training example x, you must evaluate f(x) = argmax_y p(y|x) Which merely chooses what is the most likely class considering x. It's like we were trying to model the decision boundary between the classes. In generative models instead you must evaluate: f(x) = argmax_y p(x|y) * p(y) so you are using the joint distribution p(x,y) = p(x|y) * p(y) to model each class.