search

chengsoonong / mclass-sky

Multiclass methods for astronomical data
BSD 3-Clause "New" or "Revised" License
9 stars 4 forks source link

Explore what kernels work well #142

Closed chengsoonong closed 7 years ago

nbgl commented 7 years ago

sklearn.gaussian_process.kernels.RationalQuadratic seems to work best with an R^2 of 0.91 (up from 0.88 for the Matern kernel). That said, RationalQuadratic is an infinite sum of RBT kernels, whereas the Matern kernel transform should be much easier to approximate with a finite-dimensional feature space.

I would like to stick with the Matern kernel, approximate it in a finite space, and pursue performance improvement by tweaking hyperparameters.

chengsoonong commented 7 years ago

RationalQuadratic is a good kernel to use. I suggest to not tweak hyperparameters yet. So, use two kernels for now (Matern and RationalQuadratic).