Now, based on my understanding, feature scaling should have no impact on my model results due to the fact that XGBoost isn't sensitive to monotonic transformations. Ref

My concern is the model interpretation. While using LIME, I am getting signficantly different "top drivers" for scaled and non-scaled models.


  1. Why would I see different results for both?
  2. Which version should I use?

Any guidance would be much appreciated.

  • 1
    $\begingroup$ xgboost isn't sensitive to scaling but LIME is. $\endgroup$
    – Sycorax
    Dec 2, 2019 at 23:58
  • $\begingroup$ alright, that makes sense. any thoughts on which version would be better to use? $\endgroup$
    – madsthaks
    Dec 3, 2019 at 0:00
  • $\begingroup$ All other things being equal, use the one with better resampling performace. Even if the XGBoost model has exactly the same performance is unlikely that the GLM used by LIME has similar performance (in terms of $R^2$ for example). $\endgroup$
    – usεr11852
    Dec 3, 2019 at 0:11
  • $\begingroup$ Yea, this was my worry. Can I even trust the results that LIME is producing since its built on a completely different, linear model? $\endgroup$
    – madsthaks
    Dec 3, 2019 at 2:02


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Browse other questions tagged or ask your own question.