8.7 Gaussian process classification

In Chapter 4, we saw how a linear model can be used to classify data. We used a Bernoulli likelihood with a logistic inverse link function. Then, we applied a boundary decision rule. In this section, we are going to do the same, but this time using a GP instead of a linear model. As we did with model_lrs from Chapter 4, we are going to use the iris dataset with two classes, setosa and versicolor, and one predictor variable, the sepal length.

For this model, we cannot use the pm.gp.Marginal class, because that class is restricted to Gaussian likelihoods as it takes advantage of the mathematical tractability of the combination of a GP prior with a Gaussian likelihood. Instead, we need to use the more general class pm.gp.Latent.

Code 8.7

with pm.Model() as model_iris: 
    ℓ = pm.InverseGamma('ℓ', *get_ig_params(x_1)) 
    cov = pm.gp.cov.ExpQuad(1, ℓ) 
&...