Allow extra_trees by variable

[samFarrellDay]

Summary

Would be nice to allow extra_trees to be variable specific. Would it be possible to allow extra_trees to be a list of variable names? I don't know if this is possible from a technical perspective, I just know it would be a nice feature to have.

Motivation

I am especially interested in the interaction between a specific variable and the model output. I would like lightgbm to split at many different points for this variable. extra_trees gets it to do this, but model accuracy suffers because it's choosing random split points for all other variables too.

[shiyu1994]

@samFarrellDay Thanks for using LightGBM. I think one reason that we don't have feature-wise extra_trees option is that when we allow some features to select threshold by gain, and others to use random thresholds, the former features will be much more likely to be selected since their thresholds are the ones with largest gains.

But in the case of your motivation, this can still be useful sometimes.

We may add this to feature request first. And welcome to contribute to this.

[jameslamb]

Thanks for the request @samFarrellDay and for your feedback on it @shiyu1994 .

I've added this feature request to the backlog of feature requests at #2302. Closing this issue for now.

Anyone reading this who is interested in contributing, just leave a comment here indicating that and this issue can be re-opened as a place to ask maintainers for guidance in implementing this.

[jameslamb]

Sorry, this was locked accidentally. Just unlocked it. We'd still love help with this feature!