[Biomod-commits] RES: contribution of each algorithm to the ensemble

[Frederico Mestre]

Hello,

I already ran my ensemble (selecting and weighting the models above a
pre-defined threshold). 

I wanted to see which models were selected in the ensemble and what were
their contributions. I'll check those tables.

Thanks for your help,

Frederico Mestre


-----Mensagem original-----
De: Robin Engler [mailto:robin.engler at gmail.com] 
Enviada em: sábado, 17 de Dezembro de 2011 01:51
Para: Frederico Mestre
Cc: biomod-commits at r-forge.wu-wien.ac.at
Assunto: Re: [Biomod-commits] contribution of each algorithm to the ensemble

Hi Frederico,

> I’m trying to understand which is the contribution of each of the 
> algorithms to the final ensemble. I’ve selected them by using the 
> qual.th, and weighted them using weight.method and decay in a “global 
> ensemble” (with all the algorithms).

The contributions of each algorithm depend on the type of "ensemble
forecasting" method that you chose, but generally they will be based on
either the AUC or the TSS (or Kappa) scores of each model. In the case of
AUC values, the actual weight used in the ensemble forecast is the AUC value
stretched between 0 and 1 (rather than 0.5 to 1 which is the original scale
of the AUC).

> Is there any way to get a table with the contribution of each of the 
> algorithms?

So if you wish to see the contributions of the different models, you should
look at the tables Evaluation.results.Roc Evaluation.results.TSS
Evaluation.results.Kappa

If you have implemented split-sampling evaulation, then the AUC/TSS/Kappa
values used as weights in the ensemble forecast are those from the
split-sample evaluation. If you have used completely independent data for
evaluation, then the evaluations scores derived from these data are used.


qual.th is the threshold under which a model is discarded from the ensemble
forecast. Set this value to 0 (or 0.5 if using AUC) if you wish to keep all
models regardless of their evaluation scores.

Hope this helps,
Cheers,
Robin