reconstruct for kernel PCA

[mateuszbaran]

Where the idea for code for kernel PCA reconstruct in MultivariateStats.jl

MultivariateStats.jl/src/kpca.jl

Line 61 in 30aea0c

function reconstruct(M::KernelPCA{T}, y::AbstractVecOrMat{T}) where {T<:Real}

is taken from? I've found some papers that state that reconstruction in kernel PCA is quite difficult and suggest solving it as an optimization problem: https://arxiv.org/pdf/1207.3538.pdf which is not what MultivariateStats.jl does.

[wildart]

The optimization problem you've mentioned is solved by a kernel ridge regression. See "3.1 Estimation of the Pre-Image Map" from "Learning to find pre-images", Bakır, 2004.

[mateuszbaran]

Thanks, it may be worth adding to documentation.