0
$\begingroup$

I am wondering if there is a need to set the beta multiplier in Maxent (species distriubition modeling approach) if one is also reducing features using a contribution threshold. I have seen a number of published SDM papers that reduce the number of features, and I have seen others that combine feature reduction with a beta multiplier that is not the default.

I am using AUC to evaluate the models. In addition, I am using K fold validation to estimate a testing AUC. I am comparing the two AUCs to check for model overfitting.

My understanding of using a beta multiplier is basically to provide a "smoother" to compensate for the possibility of model overfitting. If I am removing features, do I also need to identify the "best" beta multiplier? If so, do I need to do this iteratively (i.e. at every subset of the original full feature dataset)?

$\endgroup$
0
$\begingroup$

1) I suggest to NOT use threshold features as is the default in the recent version. I discourage product features, too. -- unless all potential interactions would make sense biologically/ecologically.

2) I suggest to leave the regularization as it is. Not much to gain and how to ecologically interpret any such changes?

3) I suggest NOT to rely on AUC only. Use a number of metrics such as AUC (train or test?), TSS, ORSS (Stephenson 2000), and others. Also think carefully about what thresholds you use to discriminate yes/no.

4) I suggest using a large number of bootstraps or a small k cross-validation.

$\endgroup$
  • $\begingroup$ 4) I hope "this" helps. What do you mean by "this"? $\endgroup$ – Ferdi Sep 12 '17 at 9:15

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.