Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Already on GitHub? Sign in to your account

[WIP] Prediction variance in Ridge #3417

wants to merge 2 commits into


None yet
3 participants

mblondel commented Jul 17, 2014

I started working on returning prediction variances in Ridge (for the Cholesky solver only). The n_samples > n_features (primal) case works. The n_features > n_samples (dual) case is a work in progress.

mblondel added some commits Jul 17, 2014

@mblondel mblondel Clean up code. 85a8709
@mblondel mblondel Prediction variance in Ridge (WIP).

eickenberg commented Jul 19, 2014

Interesting! Is there somewhere I can read up on this? Will you be working this into all the solvers or just into the cholesky one for now? (Asking because my ridge_path contribution is turning into a the major refactoring I expected to happen, with some associated difficulties and delays. Hope to have it in a satisfactory state soon.)


agramfort commented Jul 19, 2014

I'd be curious to have a ref too. Or maybe you can briefly explain the maths?


mblondel commented Jul 19, 2014

AFAIK it can only be implemented for the cholesky solver or by using matrix
inversion of the covariance matrix. I think we can merge the ridge path
stuff in priority. It'd be nice to cherry pick my clean up commit though.
It extracts the input checking logic and the solver resolving. I don't mind
if you just copy paste the relevant parts if it's easier for you.

For the ref, see section 3.3 in Bishop's book and in particular section
3.3.2 on the predictive distribution.

On Sat, Jul 19, 2014 at 3:54 PM, Alexandre Gramfort <
notifications@github.com> wrote:

I'd be curious to have a ref too. Or maybe you can briefly explain the

Reply to this email directly or view it on GitHub
#3417 (comment)


agramfort commented Jul 22, 2014

thanks for the ref. Looking forward to see a nice example with this :)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment