Tanimoto kernel and beyond

[PAI-Performers]

First of all, this is xGPR repo is awesome! With further development, it could be even more powerful practically.
I am wondering, as one possible extension, if the implementation of the Tanimoto kernel can be included? A few other GP based molecular applications use this kernel. It'd be great to benchmark against the standard kernels such as RBF and Matern.
Moreover, graph kernels are increasingly under development recently, e.g., "Kernel Regression Over Graphs Using Random Fourier Features" https://ieeexplore.ieee.org/document/9707857. If there can be additional graph kernels to choose, it'd be super cool!

[jlparkI]

Thanks! Yes, I would like to add the Tanimoto kernel (I saw a preprint about a random features based implementation of this kernel recently) and am hoping to do that soon. We're also open to contributions so if you're interested in writing your own implementation on a fork and submitting that as a PR, can take a look at that too. But like I said, hoping to add an implementation of that kernel.

The "Kernel Regression Over Graphs" paper looks interesting, although a little bit trickier to implement because it has a separate weight vector associated with each graph node, so it has to be initialized with a given graph size. Still, there is probably a way to expand the existing framework to incorporate this, I'll think about this one. There are a couple of other random features-approximated graph kernels we're interested in adding potentially -- this random walk one (https://arxiv.org/pdf/2305.00156.pdf ) and this one (https://arxiv.org/pdf/1911.11119.pdf). So, yes, we're hoping to have more graph kernels in here soon...

[jlparkI]

OK, quick update on this. We didn't include the Tanimoto kernel or the random walk graph kernel in the new release (0.2), but that's only because there were a lot of other changes we wanted to make -- the preconditioner size is now selected automatically so user no longer has to do this, hyperparameter tuning has been greatly simplified and it's now much easier to use e.g. Optuna to tune hyperparameters, xGPR now does approximate kernel classification using an approximate discriminant classifier, etc. etc.

We will try to get these new kernels in the next release -- probably sometime early next year.