Question - how conformal methods work for predicting many dimensions?

[vitkl]

Thanks for a fantastic series about conformal prediction. I am working on a model that outputs predictions for 100s-1000s of variables. Is it possible to use conformal predictors in that case?

Does that require simply specifying independent conformal intervals for each variable as described in Episode 4B https://www.youtube.com/watch?v=EXo5v2QJxYU?

Do we instead have to consider all variables together (e.g. flatten {observation, variable} matrix of validation predictions and observations, then consider all values to construct conformal quantiles, then unsqueeze the matrixes of quantiles to original input dimensions {observation, variable}?

Any recommendations would be appreciated!

Vitalii

[mtorabirad]

Hello,

Sorry for my late reply.

Yes, conformal prediction can be used for this case and all variables can be considered together.

If you provide more details about your loss function, I can probably give a more detailed answer.