| |
|
Science Forum Index » Statistics - Math Forum » Non-prediction error cross-validation?
Page 1 of 1
|
| Author |
Message |
| Guest |
 Posted: Thu Apr 03, 2008 3:41 pm |
|
|
|
|
Something I've been wondering...
In general, people do cross validation by minimising prediction error
with the training set using their tuning parameter. However, some
estimators (e.g. the LASSO) incorporate elements of *variable
selection*, and hence we might want to do things (like do sign-
consistent model selection, say) that normal cross validation will
give bad answers for.
Would it be sensible to cross-validate and optimise, say, the AIC, or
some other value that would penalise high numbers of selected
parameters? There doesn't seem much discussion in the literature about
this. Or am I stating the obvious?
Zhou Fang |
|
|
| Back to top |
|
| |
|
Page 1 of 1
All times are GMT - 5 Hours
The time now is Sun Jul 27, 2008 3:09 am
|
|
