Gaussian Process with binned data


#1

I was wondering if there is an implementation or a way to use the existing tools to do GP with binned data for bins with irregular shapes. I want to do something similar to what is discussed here https://arxiv.org/pdf/1809.02010.pdf, but for any kernel and non-rectangular bins.


#2

There’s nothing specifically implemented for this scenario, but in PyMC3 it’s pretty easy to construct your own kernel, or even write your own GP class if necessary. Gradients are handled by Theano so you don’t have to worry about that. Not familiar with that paper, but from a quick skim, it looks like you’ll just need to implement the kernel in Eqs. 2 and 3, which doesn’t look too bad. Here’s an example of a custom covariance function.


#3

Thanks, this might help!