search

jinlow / forust

A lightweight gradient boosted decision tree package.
https://jinlow.github.io/forust/
Apache License 2.0
56 stars 6 forks source link

Feature/calc init #48

Closed gravesee closed 1 year ago

gravesee commented 1 year ago

@jinlow I couldn't get the xgboost and forust predictions to match greater than tolerance 0.001 for the SquaredLoss with initialize_base_score=True test. It matches to the thousandth's place...

jinlow commented 1 year ago

If we defaulted to true We could warn if init base score was true, and the provided base score is something other than 0.5.

jinlow commented 1 year ago

Or, we set initialize_base_score to false, and then now or in a separate pull request make base score an optional, but only for the new method/init. Then in there we check if base_score is none, and if it is, we set initialize_base_score to true, and use our default base score of 0.5 on the booster, knowing it will be corrected at fit time. So base_score still stays just a straight f64 on the booster.