Dear RG-community,
I am curious how exactly the training process for a random forest model works when using the caret package in R. For the training process (trainControl()) we got the option to specifiy a resampling method (e.g. oob or boot). Obviously a random forest model is happy without any additional resampling as 1/3 of the training data is supposed to be put aside to achieve an unbiased estimate of the classification error. What if I specifiy another resampling method instead of "oob" (e.g. boot)? How is the classification error estimated? Is the oob-sampling still running in the background? If not, what data is used to be passed down the individuals trees (e.g. 100 bootstrap samples = 100 trees)?
Thanks for your help.
René
Dear Rene,
It is a nice question. Do you have any related paper on this?
Dear Rene,
It is a nice question. Do you have any related paper on this?
If you are asking for more general articles on random forests...yes. E.g.:
Strobl, C., Malley, J., & Tutz, G. (2009). An introduction to recursive partitioning: rationale, application, and characteristics of classification and regression trees, bagging, and random forests. Psychological methods, 14(4), 323.
Also there is quite some literature from the package authors themselves (e.g. http://topepo.github.io/caret/index.html). It is possible that the answer to my question is hidden somewhere in these articles but I did not find it yet.
Cheers,
René
Hi
I usually avoid "oob" because it can only be used with tree-based models, so it is not really useful if you want to compare different models.
I use repeated cross validation or bootstrap cross validation (i.e. "repeatedcv" or "boot"). In both cases the model is estimated on the training set/ inner bootstrap sample and the classification error (Kappa & Accuracy) on the test set/ outer bootstrap.
Stefano
- Badges
- Science topic