Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

Powered by ReviewNB

Just skimmed but the plots are a true beauty!

Just skimmed but the plots are a true beauty!

You were not wrong!

I have concerns about the out-of-sample prediction in Example1. All that is going on there is that the GP is quickly reverting to the mean function (i.e. zero) as soon as it leaves the range of the observed data. Its not providing anything insightful out of sample, as far as I can tell. This is fine, but the narrative has to reinforce that. There is a lot of good supporting text at the start of each example, but much less supporting the results. This is an instance where it needs to be explained pretty well, or the reader will be mislead (unless they already understand GPs well). I think the only way you can get something interesting out-of-sample there is to have a second covariance function with a much longer lengthscale so that it is not immediately deferring to the prior. Let me know what you think.

We should also add @ferrine as an author (if he's OK with that). He came up with the hierarchical model.

RE what you said @fonnesbeck, that's definitely true. HSGPs are additive, so you can do:

cov = cov1 + cov2
gp = pm.gp.HSGP(m=m, c=c, cov_func=cov)

This can speeds things up because the two GPs share the same basis. Maybe this is a good opportunity to demonstrate this?

Yes, I'm happy to co author

@fonnesbeck: @bwengals and I improved the hierarchical example, to take your feedback into account. Mainly:

• Implemented an additive HSGP on the mean GP, which makes for better out-of-sample predictions and an even more interesting example, code-wise
• Tweaked the data generating process to increase the diversity between the offset GPs, to make them more distinguishable

I think we're in a great position now, so feel free to merge if that looks good to you 🥳

View / edit / reply to this conversation on ReviewNB

fonnesbeck commented on 2024-07-04T19:49:01Z
----------------------------------------------------------------

The intervals don't look quite right here. Plotting issue?

AlexAndorra commented on 2024-07-04T19:53:27Z
----------------------------------------------------------------

No that's because the true GP is randomly centered around -5, while the prior is centered on 0; and at the same time I'm sharing the y-axis on the plot

No that's because the true GP is randomly centered around -5, while the prior is centered on 0; and at the same time I'm sharing the y-axis on the plot

View entire conversation on ReviewNB