[Feature] add new pseudo huber quantile loss

RektPunk / MQBoost

Multiple quantiles estimation model maintaining non-crossing condition (or monotone quantile condition) using LightGBM and XGBoost

24 stars 4 forks source link

[Feature] add new pseudo huber quantile loss #17

Closed chulhongsung closed 4 weeks ago

chulhongsung commented 1 month ago

Summary

Add another version of pseudo huber objective function

RektPunk commented 4 weeks ago

@bbeomjin found that if the error term is applied in the Hessian as in your phuber2 objective, the performance may vary depending on the scale of response variable. Therefore, I will close this PR. Thank you for your contribution.