How to get the list of features which are actually used by XGBoost

If the input data have 300 features, and if I setup the XGBoost with 3 trees and 3 level for depth of each tree, the 300 features cannot all be used for sure. But when I call model.get_booster().feature_names, all 300 features were returned.

My guess is that the model.get_booster().feature_names returns all the features in the training data, not the features used by the XGBoost model.

Is there a way to check which variables are actually used by the model? Thank you very much in advance!

The Booster.feature_names attribute is the description of the training dataset - which features, in which order.

In principle, you could query feature importances as model.feature_importances. The thinking goes that "used" features have non-zero values, and "unused" features have zero values.

The correct approach would be to traverse XGBoost tree data structure, and collect node split indices (which correspond to column indices in your training dataset). If your model config is n_estimators = 3 and max_depth = 3 then, by definition, there can be at most 3 * 2^3 unique "used" features.

It's probably hard to implement such tree traversal in Python code, because the internal state of the Booster object is not exposed via public (Python facing-) APIs. But you could dump the Booster object in JSON data format and parse/traverse that.

For reference purposes, you could export your XGBoost model into a PMML document using the JPMML-XGBoost library (or its SkLearn2PMML package front-end), and open the resulting PMML XML file in a text editor. The conversion to the PMML representation only retains "used" features.