search

amazon-science / quantile-aggregation

Code for Flexible Model Aggregation for Quantile Regression
Apache License 2.0
17 stars 2 forks source link

Integration with Autogluon #1

Open SamuelGabriel opened 1 year ago

SamuelGabriel commented 1 year ago

Hey there :)

Thanks for this nice project and paper.

This project stands on the shoulders of AutoGluon, correct? Thus, I was wondering how this compares to AutoGluons own quantile regression, which seems to rely on Pinball Loss, if I understand it correctly. Unlike yours, which optimizes WIS.

bmeisburger commented 2 months ago

It may be helpful to see (8) here. From the paper:

In summary, we have shown that in training a quantile regression model by optimizing pinball loss as in (4), we are already equivalently optimizing for WIS (7), and approximately optimizing for CRPS (5), where the quality of this approximation improves as the number of discrete quantile levels increases.