Understanding python XGBoost model dump output of a very simple tree

Question

I am trying to understand the model dump output from XGBoost. I would like to step through and see exactly how the model arrived at it's prediction. To simplify I trained a model with 1 tree and 1 max depth, and as expected all records get one of two predictions as it has a single split - the values are {0.5386398434638977, 0.5011891722679138}. However, when I look at the model dump I see the following

booster[0]:
0:[f40<70.5] yes=1,no=2,missing=1
    1:leaf=0.00475667231
    2:leaf=0.154868156

I have no idea how to interpret this in a way that makes sense with the prediction. What am I missing? Thanks!

Answer 1

The scores at leaves measure log-odds, not probabilities. (With more trees, these scores get summed to give a final log-odds approximation, then goes through a sigmoid to get probability approximations.)

And indeed, $1/(1+e^{-0.00475667})\approx 0.501189$.