Random forest nominal variables

[Anonymous]

Is the Random Forest algorithm in Waffles able to handle nominal features with a large number of distinct values (> 100)? If I pass such a dataset to waffles_learn train randomforest, will it be transformed in some way before the randomforest algorithm is called?

Thanks,
Ganesh

[Mike Gashler]

There is no theoretical limit, but I have not done much testing with this case. I suspect that random forest would do much better in this case with binary divisions. Example:

waffles_learn crossvalidate mushroom.arff bag 50 decisiontree -random 1 -binary end

(If it does not work, I would like to have repro steps, so I could debug it and make it work.)

There would be no automatic transformation of the data, since the GDecisionTree implicitly supports both nominal and continuous features and labels.

[Anonymous]

Thanks for the response. Random forest does work in a test example I created that has a nominal variable with nearly 100 levels. Specifically, I did: ~/waffles/bin/waffles_learn train training_data.arff randomforest 100 -samples 2