如何通过“信息增益”获得 xgboost 中的特征重要性？

Question

we can get feature importance by 'weight' :

model.feature_importances_

But this is not what i want. I want by importances by information gain.

we can get feature importance by 'gain' plot :

xgboost.plot_importance(model, importance_type='gain')

However, I don't know how to get feature importance data from above plot. Or

if there is function like model.feature_importances_ to give gain feature importance? Either of the two ways will work. Please let me know in comments if the question is not clear

Answer 1

You can get it from

model.booster().get_score(importance_type='gain')

http://xgboost.readthedocs.io/en/latest/python/python_api.html

Answer 2

Actually, I am a bit unclear about your question but still I'll try to answer this.

I guess you need something like feature selection. If I am right, then you can check sklearn.feature_selection.

Following is the URL: http://scikit-learn.org/stable/modules/feature_selection.html

There are many important functions like chi2, SelectKBest, mutual_info_classif, f_regression, mutual_info_regression, etc..

Answer 3

In the current version of Xgboost the default type of importance is gain , see importance_type in the docs .

The feature importance can be also computed with permutation_importance from scikit-learn package or with SHAP values. You can read details on alternative ways to compute feature importance in Xgboost in this blog post of mine.