BUG: XGBRFRegressor invalid number of trees error with shap>=0.42 about shap HOT 3 CLOSED

Thanks for the bug report!

On the master branch of shap, I can reproduce the issue above with xgboost == 1.7.6 but not with xgboost == 2.0.0.

EDIT: It looks like a difference in the loaded tree_limit:

This is then passed to iteration_range, which replaces the deprecated xgboost parameter ntree_limit as of #2987 :

• with XGBoost 1.7.6, self.model.tree_limit is 100. The original model is called with interation_range=(0, 100).
• with XGBoost 2.0.0, self.model.tree_limit is None. The original model is called with iteration_range=(0, 0).

The RFRegressor fails with an iteration_range that is higher than 1, presumably because the trees are in parallel. So, I think iteration_range is not quite a drop-in replacement for ntree_limit. Perhaps the interation_range should be something like tree_limit / num_parallel_trees?

On a side note, it doesn't look like there are any tests for XGBRFRegressor or XGBRFClassifier at the moment, so it would be good to explore if shap can officially support them. @trivialfis , do you by any chance know about this history of XGBoost & shap, if there is any particular reason why the XGB RF models are seemingly not already supported / tested by shap?

from shap.

This may be solved by #3462 , which will improve the XGBoost model loading significantly. As a temporary workaround for now, you could try updating to XGBoost 2+ if possible.

from shap.

This should be solved by #3462, and will be released in 0.45.0.

from shap.