Feature Importance Cover


I am trying to obtain the same feature importance metrics available in Python and R but with Java.
To this end, I parse the model dump which works fine but I don’t fully understand what the cover metric means.
In one of my examples, I have two features and 3000 samples and train only a single tree for simplicity.
In my understanding cover means the number of samples affected by a split, so the root split should have a cover of 3000 but in the model dump, I find that it only has a cover of 1500.

Can you please explain to me what the meaning of cover really is?

Kind regards,



This could be of help.

The root node has all the probabilities set to 0.5 (initial guess).