We consider classification models of the form:
where p(y=c,x) is the prior over the class labels, and p(x∣y=c,θ) is the class conditional density for c.
- The overall model is a generative model since it specifies the distribution over the feature x, p(x∣y=c,θ)
- By contrast, a discriminative model directly estimates the class posterior p(y=c∣θ,x)