Closed sfladi closed 3 months ago
I was unaware, that the leave on out cross validation error can actually be computed in a single pass as implemented in rbf.interpolate._loocv(d, K, P).
I still have to read up on the algrebraic tricks that lead to this solution. But that is not an issue with this package.
The documentation talks about automatic estimation of the smoothing parameter sigma and/or kernel scale parameter epsilon for the RBFInterpolatant, using leave-one-out cross validation (LOOCV).
After reading through the implementation, it looks to me like it is performing just a single pass optimization over the whole dataset (instead of looping over cross-validation subsamples).
Could you point me to where the cross validation is happening?