Question on feature importance

christophM / interpretable-ml-book

Book about interpretable machine learning

https://christophm.github.io/interpretable-ml-book/

Other

4.78k stars 1.06k forks source link

Question on feature importance #166

Closed huanfachen closed 4 years ago

huanfachen commented 4 years ago

Can we say 'one feature is more important than another' after comparing the feature importance? For example, using the absolute value of the t-statistic in linear regression. It sounds okay, but I would like to confirm this. Is there any literature on this?

christophM commented 4 years ago

Depends how you define "important". For permutation feature importance you can say that one feature causes more performance loss when permuted than another feature. I don't think there is a single clear definition of what importance in the context of model interpretation is.