I am trying to apply the lasso or ridge regression to my data set for the feature selection, but different random seeds produce different models. What is a good or universal way to obtain the final model?

  • Fix a seed, OR
  • combine models from different seeds (if so, then how should I combine them?), OR
  • compare models from several different seeds and find a kind of representative model?

Any comments or references will be really appreciated!

  • $\begingroup$ +1 for a great question. My 1st guess is that neither model is really better than the other--there is some ambiguity given your data. If you want predictions, you could try model averaging. $\endgroup$ – gung - Reinstate Monica Jun 21 '13 at 17:06
  • $\begingroup$ I have to turn it round. Why do you think a trustable final model exists if results depend on seed choice? $\endgroup$ – Nick Cox Jun 21 '13 at 17:16
  • $\begingroup$ Thanks for your replies. I am working on the prediction of survival. Once I set up the final model, I need to validate this model in an independent data set. But I guess that different models (with different seeds) can lead to different results. So I wonder there is a commonly used way to determine a final model... Regarding model averaging, I thought that it could be used when I want to combine models with different "methods" (not models with different seeds). I should do some research. $\endgroup$ – Jenny Jun 21 '13 at 18:06

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.