Gaussian Process kernel WhiteKernel has unhelpful "noise_level" documentation
See original GitHub issueDescription
The WhiteNoise
class of kernel has an unhelpful and misleading description. It simply says “noise_level is a parameter”. It doesn’t say it corresponds to variance of a gaussian noise.
Steps/Code to Reproduce
From the docs
noise_level : float, default: 1.0 Parameter controlling the noise level
From the source
The main use-case of this kernel is as part of a sum-kernel where it explains the noise-component of the signal. Tuning its parameter corresponds to estimating the noise-level.
Expected Results
noise_level : float, default: 1.0 Parameter controlling the noise level (variance)
The main use-case of this kernel is as part of a sum-kernel where it explains the noise of a signal as Gaussian noise with variance equal to noise_level. Tuning its parameters corresponds to estimating the noise variance.
Versions
Not relevant
Issue Analytics
- State:
- Created 4 years ago
- Comments:8 (4 by maintainers)
Top GitHub Comments
@bjfar Great work. Please edit the documentation and submit a pull request. Nobody else should have to go through this pain again. I already submitted a pull request to change WhiteKernel’s noise_parameter documentation, so all you need to change is the documentation for that alpha parameter.
I found this StackOverflow question asking the same thing: https://stackoverflow.com/questions/54987985/using-scikit-learns-whitekernel-for-gaussian-process-regression
I think the answer given there is correct: the noise is not added to the prediction variance when you set “alpha” (this only affects the training stage, by considering the training data to be noisy), but it is added to the prediction variance when you use WhiteKernel. So the two approaches are absolutely not equivalent, and the documentation is wrong to claim that they are. When I manually add the noise into the prediction variance I see exact agreement, so this looks to be the problem.
Other than this, it looks like both parameters are indeed variances.