During the model selection phase of my (regression) work I noticed that Lars with lasso modification performs way worse than any other model if combined with recursive feature elimination, both in terms of R squared and mse. Comparison was made between lasso-lars and linear, simple lasso and ridge regressors.

Is there any particular theoretical reason for this behaviour, or does it depend on the data? If the latter is the case, what do I have to check in the data?

  • $\begingroup$ Is cross-validation involved in any way? $\endgroup$ – Chamberlain Foncha Mar 15 '18 at 11:05
  • $\begingroup$ @ChamberlainFoncha It is. Actually, I am using sklearn function RFECV which performs cv recursively to identify the optimal number of features to select. $\endgroup$ – sato Mar 15 '18 at 13:49

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.