set min_child_weight as a float instead of int

[QuantHao]

Hi, all

Based on xgboost documentation, min_child_weight is treated as number of observation for rmse objective and is usually chosen as int such as https://xgboost.readthedocs.io/en/latest/python/python_api.html#xgboost.XGBRegressor.

However, I recently read a article saying that "With XGBoost, you can specify this parameter as a float as well. If you do, it will use a percentage of samples to determine leaf nods." (https://kevinvecmanis.io/machine%20learning/hyperparameter%20tuning/dataviz/python/2019/05/11/XGBoost-Tuning-Visual-Guide.html#child_weight). I cannot find a second source using min_child_weight as a float.

So I am confused and worried that is it true that we could set min_child_weight as a float?? Any comfirmation would be appreciated.

[QuantHao]

sklearn's min_samples_leaf supports float as a percentage of total sample size which makes sense to me. It would be great if xgboost supports similar functionality.

[trivialfis]

Sorry for the mistake in our doc. Also, the parameter you referred min_samples_leaf is indeed on our roadmap: #5444

[QuantHao]

Hi @trivialfis , thanks so much for your timely reply. May I double check with you that if min_child_weight is set as, for example, 0.01, it will be treated as n_sample * 0.01, meaning 1% of total sample?

Reference: https://kevinvecmanis.io/machine%20learning/hyperparameter%20tuning/dataviz/python/2019/05/11/XGBoost-Tuning-Visual-Guide.html#child_weight

[trivialfis]

No. It's the minimum value of accumulated Hessian for each leaf. Hessian is often used as a proxy for data in gradient boosting.

[QuantHao]

No. It's the minimum value of accumulated Hessian for each leaf. Hessian is often used as a proxy for data in gradient boosting.

I see. Thanks for answering.