pbiecek / breakDown

Model Agnostics breakDown plots
https://pbiecek.github.io/breakDown/
103 stars 16 forks source link

breakdown to slow for xgboost #25

Open amirva opened 6 years ago

amirva commented 6 years ago

Hi Biecek,

The package is great; but I got a non-sense issue. I have a xgboost binary classification model (objective = "binary:logistic", metric = "auc) with roughly 1000 trees and 200 variables. When I am trying to breakdowm the model for a new observation using "broken", it takes 1.5 hour per each observation!! Is it normal or something is wrong?

Should I specify the prediction function? DMatrix or model.matrix objects could be source of the problem? In this implementation, can I use DMatrix object without using "model.matrix"? If so how?

Thanks,

Amir