This is fantastic thank you! I also never quite understood before that adding predictors can relieve the GP, but makes a ton of sense since it’s less variation you need to capture for each datapoint (and datapoint = parameter) of a GP!
I originally had Gamma priors as the Kernel priors, but I found that it changed the fitting time from 4 hours to 44 hours which wasn’t really feasible so I fit to a small subset and took the mean of the length-scale prior (I aggressively rounded on these as a v1, I should really improve that). Thinking back, I really should try fitting to fake data at this point to make sure the slow fitting time is due the nature of the model and not due to a misspecified model fitting to the data.
@OriolAbril I’ve got to be honest I’m relatively new to the field so I’m not the most qualified technically, but I am definitely interested in working on it, and it’s one of my long term goals to start contributing to open source (I just imagined I’d know way more before I started).
I just read the paper you linked and it was really manageable and the pseudocode looks straightforward - where do I get started!