Xgboost: Meaning of feature importance score

[FabHan]

Do they have an interpretable semantics? How are they calculated? Does higher mean better?

To clarify, I'm using cls.booster().get_fscore() to get the scores.

[FabHan]

Also, get_fscore() returns fewer features than the number of features in the training data. I have 98 features and get_fscores() return scores of 71 features.

[tqchen]

The higher the better, get_fscore returns number of occurance of features in the ensemble

[FabHan]

Does it use their levels in the tree as weights?

Also, do you have an explanation for the situation in my second question?

Thanks.

[tqchen]

that means these feature never get selected into the trees