Feature importance in machine learning module
I understand that the feature importance is calculated by permutation-based mean dropout loss of a certain performance index, but I could not find which index it is? Is it AUC or accuracy or something else?
Sorry it might be a basic question. I'm working on explaining this method in my article but stuck on this point for a while. Thanks for any reply!
Comments
Hi YCWang,
We use the model_parts function from the DALEX R package to compute the variable/feature importance, more info on this method can be found here: https://ema.drwhy.ai/featureImportance.html. By default, the loss function is 1 - (minus) AUC for binary classification, cross entropy for multilabel classification and RMSE for regression. I will make sure this is added to the output in the next release.
Thank you koenderks. It helps a lot!