Feature importance

[florian6973]

Hi!

For feature analysis and interpretability, it could be nice to be able to compute easily and/or export feature importances for the xgboost model.

Thank you!

[mmcdermott]

This seems viable to do with XGBoost proper. @florian6973, can this be done by re-instantiating the model returned by MEDS-Tab and computing the feature importance directly? If not, what are the barriers to that? Just so I better understand the issue and scope here. Thanks!

[florian6973]

I agree that reinstantiating the model returned by MEDS-tab and computing the feature importance should work. However, the barrier I faced is that I have not found the mapping between the features and their meaning. Do we already know to which disease code and aggregation type each column corresponds to? Or do we need to add some code to keep track of it?

[mmcdermott]

How do the features appear to you in this usage @florian6973 ? Are they integer indices only? @Oufattole or @teyaberg , can we add a vocabulary of feature names to xgboost or something?

[teyaberg]

Is something like this get_column_names useful, or would it be useful to have this accessible in some other way?

[mmcdermott]

we should be passing the feature names to the XGBoost objects upon construction, I think, via feature_names here? https://xgboost.readthedocs.io/en/release_0.82/python/python_api.html#module-xgboost.core