Feature importance for LightGBM

[sharpwood]

Is there a way to provide a LightGBM Feature importance demo code ?

[codemzs]

Hi @sharpwood , Currently feature importance is not available via API but we are planning on it in the near future.

[sharpwood]

This problem has not been resolved, why is it closed? @codemzs

[Ivanidzo4ka]

@rogancarr Do we have example of how run feature importance for Tree learners?

[rogancarr]

@sharpwood We have per-sample feature importance through the FeatureContributionCalculator. You can find a sample here that uses a linear model, but you can just swap in any of our tree learners, including LightGBM to get importance for those.

Similarly, you can use Permutation Feature Importance to look at global feature importance. You can find the sample here.

Finally, you can also get the Feature Gains for each feature in our tree models through the GetFeatureWeights() method on the models (e.g. in trainedPipeline.LastTransform.Model).

For more explicit examples, I have a PR going through for functional tests that does this every which-way: #2584