Anyway you could update using this in python using XGBoost? about shap HOT 3 CLOSED

from shap.

I saw your update on XGBoost so you could close this if you wanted.

I did have another question, not really an issue or anything so I could email you if preferred to speak further.

Say I had 500 features in a model and 400 of them pushed the model score down or had little to no importance, is it safe to drop those features and re-train the model? Just wondering if there is any negative effects of doing this.

from shap.

The XGBoost PR was just merged so I'll close this issue now. Check out the example on the homepage for instructions on how to use it! (it assumes you have a version of master newer than 10/12/17)

As for the feature question, if they push the model output down then they are important to the model. If you mean they push the model accuracy down, then sure leave them out. The only way to know for sure is to actually do it and then see if the performance on a test dataset improves. Typically XGBoost is good at selecting only the features that matter (it is a stage-wise expansion, which has connections to L1 regularization), so I would be surprised if dropping them helped significantly unless the dataset is very small.

from shap.