Hello everyone,

I’m currently using Gaussian Process Regression (GPR) for my work and I found this paper: https://ieeexplore.ieee.org/document/9248120 , where they standardize both the features and target values.

Does this mean that if I want to see the results in the original scale, I can just revert them back after making predictions to visualize the plots correctly?

Secondly, when using kernels for GPR, we typically use a constant kernel for the amplitude and an RBF kernel for the length scale. But what about the alpha (noise variance) in the data? Is it a good idea to use a White Kernel for this? If I do, will it affect the entire model? I’m following the Gaussian Process for Machine Learning book by Williams & Rasmussen, and I want to make sure I’m applying the right approach.

Lastly, is it advisable to standardize time-series data when applying GPR? Would it improve numerical stability, or could it affect how the model interprets time dependencies?