Number of attributes for tree construction in RandomForest

Question

Hi,

how can I define the number of attributes (the subset of all attributes
in the learning set) randomly chosen to grow the trees in the
forest?

Could you give me an example how to specify that n^1/2 
attributes are considered.

Thanks.

Cheers,
Stephan

IngoRM · Answer

Hi,
If I'm wrong no doubt heads more pointy than mine will intervene...

no need for that  :D

Everything was perfectly right.

All the best,
Ingo

haddock · Answer

Hi Stephan,

I think that a setting of 1.0 would mean "use all the attributes" and a setting of 0.0 would mean "use none of the attributes", neither of which are of any interest to random foresters, as far as I remember.

A setting of 0.1 might be more like it, meaning "use a random selection of 10% of the attributes".

So what does "-1" mean? In general this convention is used to indicate a specific setting that is calculated, rather than part of the selection range, so -0.8 would fuse the system and cause springs to fall out of the back of your machine.

You see this convention used with quite a number of the operators.

If I'm wrong no doubt heads more pointy than mine will intervene... :D

Legacy User · Answer

Thank you for your answer.

But I still didn't catch what is meant with "(-1: use log(m) + 1 features) ":
Does it mean that with a total of 'm' attributes the setting of
subset_ratio=-1.0 yields log(m)+1 attributes used for the test?

So, what does then the setting 1.0 mean?

And I still don't see a way to define a specific number of attributes like my square example.