ForceSplit might be ignored if no gain

[tongwu-sh]

Summary

ForceSplit might be ignored if no gain. It is confuse to customer and it is the default behavior without option to disable it. We should consider to provide option on this feature and looks like default off is reasonable.

[jameslamb]

Thanks for your investigation into the use of forced splits, @tongwu-msft ! I think it's definitely a feature of LightGBM that would benefit from more testing and development attention.

Could you please add more details to the summary for this issue?

• Could you provide a small, reproducible example that demonstrates this behavior? Or, if not, provide links to specific places in LightGBM's code that provide evidence that forced splits are only used if the gain is positive?
• Could you please update this to use the word "user" instead of "customer"?

And, to clarify, which of these two things are you proposing?

1. If forced splits are provided, LightGBM should always use them, regardless of gain.
2. LightGBM should continue to, by default, only use forced splits if they provide positive gain. But a parameter forced_splits_strict (or some similar name) should be added, and if users provide forced_splits_strict=True, LightGBM should use forced splits in every tree, regardless of gain.

[shiyu1994]

I think we can allow users to ignore the gain when using forced splits. Either by adding a parameter or by allow a specification in the forced split file.

[jameslamb]

Sure!

In that case Ihink it should be exposed in a parameter, not part of the file, so you can test both True and False in hyperparameter tuning.

[shiyu1994]

Sure!

In that case Ihink it should be exposed in a parameter, not part of the file, so you can test both True and False in hyperparameter tuning.

I agree. Added to our iteration plan of this month.

[guolinke]

any updates on this?