[FEATURE] (Linear) Quantile RegressionSee original GitHub issue
At some point, I implemented the LADRegression, which basically minimized sum |y_i - model(X_i)|. This also has the effect that the model over- and underestimates 50% of the time.
As in the ImbalancedRegression, we can also penalize over- and underestimations differently, with some parameter
quantile. This would have the effect that the model overestimates a share of
quantile samples and underestimates in
1-quantile of the cases.
This one would be useful for having some kind of nice confidence intervals around predictions by training a model with
quantile=0.05 and another one with
quantile=0.95, for example.
I implemented it here. It’s basically a more general LADregression.
How about we put this into scikit-lego and make the LADRegression just a Quantileregression(quantile=0.5)? Or remove it completely.
- Created 2 years ago
- Comments:5 (4 by maintainers)
Top GitHub Comments
I’d say quantile regression is super useful! indeed AFAIK, scikit-garden is focused on tree-based models so having it in this project makes sense.
I’m okay with having both LADRegression and QuantileRegression around for a while, although we should probably either rewrite LADRegression to use the QuantileRegression, or deprecate the former
Grand. @Garve feel free to get started on a PR then. I think the example you show here will work swell for the docs too 👍
It’d be preferable to have a PR where the
LADRegression uses the
QuantileRegression under the hood. Feel free to let me know if there’s a good reason to delay that though.