[python-package] Where is the code for the 'quantile' objective, and how do I pass a custom objective to LGBMRegressor?

[qinhaohcen]

Description

I want to change quantiles' loss function. Although I can make objective='quantile' , but I can't see a quantile in LGBMRegressor's code. Can you give me some advice? Thank you!

[jameslamb]

Thanks for using LightGBM! I'm confused a bit by your question.

Which of these are you asking?

• "how do I provide a custom objective function for quantile regression with LGBMRegressor in the Python package?"
• "where can I find the code for the quantile objective function, so that I understand how it works?"
• "what parameters can I use to change the behavior of LightGBM's quantile regression?"

Can you please tell me which of those you're asking?

[qinhaohcen]

Thank you for your patient reply！ I would like to ask for advice on this three points you mentioned.

"how do I provide a custom objective function for quantile regression with LGBMRegressor in the Python package?"
"where can I find the code for the quantile objective function, so that I understand how it works?"
"what parameters can I use to change the behavior of LightGBM's quantile regression?"

[qinhaohcen]

Thanks for using LightGBM! I'm confused a bit by your question.

Which of these are you asking?

• "how do I provide a custom objective function for quantile regression with LGBMRegressor in the Python package?"
• "where can I find the code for the quantile objective function, so that I understand how it works?"
• "what parameters can I use to change the behavior of LightGBM's quantile regression?"

Can you please tell me which of those you're asking?

This is my code. When I move the mouse to 'quantite' and press Ctrl+B, I will be prompted to Can't find declaration to go to. But I really want to find out how the quantile method works in LGBMRegressor. And I hope to improve the loss function pinball loss of quantile.
lgb_model_q05 = LGBMRegressor(objective='quantile', metric='quantile', alpha=0.05, learning_rate=0.1, max_depth=7, n_estimators=100)

[jameslamb]

"how do I provide a custom objective function for quantile regression with LGBMRegressor in the Python package?"

See the Note on custom objectives at the bottom of https://lightgbm.readthedocs.io/en/latest/pythonapi/lightgbm.LGBMRegressor.html#lightgbm.LGBMRegressor for information on how to format a custom objective function for LGBMRegressor.

Custom objective functions can be passed to the objective parameter.

---

"where can I find the code for the quantile objective function, so that I understand how it works?"

the built-in quantile loss is defined here

LightGBM/src/objective/regression_objective.hpp

Line 481 in c5391c9

class RegressionQuantileloss : public RegressionL2loss {

---

"what parameters can I use to change the behavior of LightGBM's quantile regression?"

The only quantile-specific parameter is alpha (https://lightgbm.readthedocs.io/en/latest/Parameters.html#alpha).

See https://lightgbm.readthedocs.io/en/latest/Parameters.html for descriptions of the other parameters that control LightGBM's behavior.

[qinhaohcen]

Thank you very much！Now I know why I can't find it.

[qinhaohcen]

Why is hessian directly assigned to 1.0? Is there any theoretical basis? Can you provide a PDF materials for reference？
``

[qinhaohcen]

"how do I provide a custom objective function for quantile regression with LGBMRegressor in the Python package?"

See the Note on custom objectives at the bottom of https://lightgbm.readthedocs.io/en/latest/pythonapi/lightgbm.LGBMRegressor.html#lightgbm.LGBMRegressor for information on how to format a custom objective function for LGBMRegressor.

Custom objective functions can be passed to the objective parameter.

"where can I find the code for the quantile objective function, so that I understand how it works?"

the built-in quantile loss is defined here

LightGBM/src/objective/regression_objective.hpp

Line 481 in c5391c9

class RegressionQuantileloss : public RegressionL2loss {

"what parameters can I use to change the behavior of LightGBM's quantile regression?"

The only quantile-specific parameter is alpha (https://lightgbm.readthedocs.io/en/latest/Parameters.html#alpha).

See https://lightgbm.readthedocs.io/en/latest/Parameters.html for descriptions of the other parameters that control LightGBM's behavior.

If you use pinball loss for quantile，gradient（First derivative）should be 1-alpha or alpha. And hessian should be 0.

Did you consider the combination of hessian and lamda and assign hessian as 1？

Sincerely look forward to your answer!

[jmoralez]

I believe it's an implementation detail where the sum of the hessians is used as a proxy to the number of samples in a node, but hopefully @shiyu1994 or @guolinke can confirm.

[github-actions]

This issue has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.