0
$\begingroup$

I applied a feature selector (with this great python package) in my dataset. This package uses the wrapper approach, where you define a classification model that runs on your data and find the best $k$ features and return the score (e.g. accuracy).

My question is: for my experiments, can i consider this accuracy obtained with the wrapper or should i train a new model with the $k$ features with the wrapper?

Thanks,

$\endgroup$
2
  • $\begingroup$ Neither automated feature selection nor use of accuracy to evaluate classification models generally represent good statistical practice. Please read those discussions, reconsider the modeling approach, and browse this site for suggestions about how to build a model that combines your knowledge of the subject matter, the available data, principled ways to lower the dimension of the predictor set (e.g., ridge regression, lasso), and use of proper scoring rules for probability/classification models. $\endgroup$ – EdM Oct 1 '20 at 21:03
  • $\begingroup$ @EdM thanks for the references. I see that this is a bold discussion over ML field, and i have a lot to study. $\endgroup$ – joann2555 Oct 1 '20 at 22:02

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.