I'm doing a Gaussian Process regression using a SquaredExponential kernel. Due to very different scales in my multidimensional inputs -some components lie in [0, 0.2] while others in [1000, 1000000]- I decided to standardize the inputs so that all their components lie in [0,1]. I standardized the outputs as well so that they have 0 mean and they lie in [-0.5, 0.5].

Testing it on a known function I can reconstruct it from few points and I am able to rescale it to its original size. However, I don't know what to do to scale back the variance of the predicted points, since I assume it is smaller due to the standardization. How can I scale it back to the size it's supposed to be?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.