Closed DonIvanCorleone closed 1 year ago
Hi,
Thanks for using this package and glad to see the performance bumps. For your question: Currently model_output="logloss"
is only supported when feature_perturbation="interventional"
(see https://github.com/linkedin/FastTreeSHAP/blob/master/fasttreeshap/explainers/_tree.py#L116, https://github.com/slundberg/shap/blob/master/shap/explainers/_tree.py#L89). Since FastTreeSHAP algorithm (parallelism, v1, v2) is currently only implemented when feature_perturbation="tree_path_dependent"
, it does not bring up computational speedups when setting model_output="logloss"
unfortunately. Hope this answers your question.
Many thanks for your quick & detailed answer! This helps a lot.
Maybe someday in the future feature_perturbation="interventional"
will be supported as well? ;)
Closing the ticket. Cheers
@jlyang1990 thank you for this very useful package!
May I ask - are you planning to implement FastTreeSHAP for feature_perturbation="interventional"
? If so, do you have a timeline in mind! Many thanks!
Hi there,
first: Many thanks for that cool project. We played around with it and we got several nice little performance bumps. Pretty nice.
Unfortunately, we realized that in cases when we set
model_output="logloss"
the parallelism was not working anymore and all outputs of any algorithm [v0, v1, v2] are nearly the same (+/- a couple of seconds)in terms of speed. Is this intended behaviour?Cheers