When mtry is equal to total number of features, you just get regular bagging
(in the R package -- Breiman & Cutler's Fortran code samples variable
with
replacement, so you can't do bagging with that). There are cases when
bagging will do better than random feature selection (i.e., RF), even in
simulated data, but I'd say not very often.
HTH,
Andy
From: Arne.Muller at sanofi-aventis.com>
> Hello,
>
> I've a question regarding randomForest (from the package with
> same name). I've 16 featurs (nominative), 159 positive and
> 318 negative cases that I'd like to classify (binary classification).
>
> Using the tuning from the e1071 package it turns out that the
> best performance if reached when using all 16 features per
> tree (mtry=16). However, the documentation of randomForest
> suggests to take the sqrt(#features), i.e. 4. How can I
> explain this difference? When using all features this is the
> same as a classical decision tree, with the difference that
> the tree is built and tested with different data sets, right?
>
> example (I've tried different configurations, incl. changing ntree):
> > param <- try(tune(randomForest, class ~ ., data=d.all318,
> > range=list(mtry=c(4, 8, 16), ntree=c(1000))));
> >
> > summary(param)
>
> Parameter tuning of `randomForest':
>
> - sampling method: 10-fold cross validation
>
> - best parameters:
> mtry ntree
> 16 1000
>
> - best performance: 0.1571809
>
> - Detailed performance results:
> mtry ntree error
> 1 4 1000 0.1928635
> 2 8 1000 0.1634752
> 3 16 1000 0.1571809
>
> thanks a lot for your help,
>
> kind regards,
>
> ______________________________________________
> R-help at stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide
> http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.
>
>
