How to get feature importance through gain?

[hanhanwu]

Hello FLAML team,

I tried getting feature importance through FLAML like what's said here: https://microsoft.github.io/FLAML/docs/FAQ/

But when using model like LGBM, the importance is using the count of splits. I want to check the feature importance through the "gain". Is that possible to get it through FLAML?

I have also tried to set save_feature_importance_type=1 as what's shown here: https://lightgbm.readthedocs.io/en/latest/Parameters.html#saved_feature_importance_type, but when calling FLAML's estimator's feature_importance_, it's still using split counts....

Any suggestion? Thank you!

[qingyun-wu]

If LGBM supports it, you should be able to also get it through FLAML's LGBM if configured properly.

1. What is the LGBM version you are using? As far as I can see, the save_feature_importance_type is only effective after v3.3.0.
2. If you are using LBGM with a version >= v3.3.0. Can you show me your code on how you setsave_feature_importance_type=1 for LGBM while using FLAML?

[hanhanwu]

My LGBM is v3.3.2, here's the code of FLAML settings.

Then I was calling the feature importance through FLAML trained model like what's shown here: https://microsoft.github.io/FLAML/docs/FAQ/, the importance was still ranked by splits not by gain

[qingyun-wu]

@hanhanwu, can you successfully use gain-based feature importance even if you manually configure an LGBMClassifier? According to the documentation you provided, the parameter can be used only in CLI version.

Thanks!

[hanhanwu]

I see. Makes sense, thank you @qingyun-wu