|
My understanding is that Gaussian regression requires the all subsets of the inputs to normal random variables. So does that rule out discrete predictors... can we treat categorical variables similar to how they are treated in linear regression (dummy variables with coding)? |
|
GPR can be used with other types of data. There is no requirement that all the inputs must be real-valued. As long as you have a valid kernel/covariance function you are fine. |
|
To add to George's answer: The assumption is that all outputs (jointly) form a multivariate Gaussian. The inputs come into play by defining the covariance of that MVG via a Kernel function. |
