more specialty kernels! check out Julia's implementations of some-kernels-you've-never-heard-of

[frankiethull]

here I am browsing julia and python, checking out what kinds of specialty kernels are already out there.

https://juliagaussianprocesses.github.io/KernelFunctions.jl/stable/kernels/

there are some unique ones here as well: piecewise polynomial, wiener, matern, etc.

a lot of these I haven't heard of before. Taking notes on what's out there as possible enhancements. I think it's worth exploring these as kernlab kernels as well, binding to parsnip once they are working in the ksvm() function.

[frankiethull]

• gabor
• matern
• piecewise poly
• (sm) spectral mixture
• cosine
• sqexponential

[jrosell]

What about nystroem? It's implemented in scikit-learn and also in R in the {mildsvm} package (I haven't tested it yet). Here more information: https://scikit-learn.org/stable/modules/kernel_approximation.html#nystroem-kernel-approx

[frankiethull]

@jrosell - thank you for an'maizing idea.

I will need to look further into this one! additionally, had not came across the {mildsvm} package before, there's some really interesting engines & techniques in there. nystroem is teetering between engine, special kernel, or a feature map, in my mind.

Reading that link & {mildsvm} docs, it seems nystroem is an approximation method of a kernel, {mildsvm} supports nystroem for radial only. Seems like there could be multiple paths in this corn maze 1) extend bindings to {mildsvm}, 2) mimic logic and add a specialty nystroem-radial kernel to maize that works with {kernlab}, or 3) add a nystroem feature map approximation as a recipe step (?) then it might be compatible with other {kernlab}/{maize} kernels too.