Number of trees required for reliable results

[researcher1]

Hi,

Suppose I am running RJ on a data set which has more than 500k SNPs.

What is the number of trees required for reliable results?

In your paper you mentioned applying RJ for a number of iterations to
even out the variations of importance values.

What is the number of iterations you would recommend?

Thanks!

[dafr...@googlemail.com]

Hi!

Grow 10000-100000 trees and repeat it 100-1000 times.

Rgrds,
Daniel
Sent via BlackBerry from E-Plus

--
You received this message because you are subscribed to the Google Groups "randomjungle" group.
To post to this group, send email to random...@googlegroups.com.
To unsubscribe from this group, send email to randomjungle...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/randomjungle?hl=en.

[researcher1]

Does having a larger mtry result the need of a smaller number of trees
required?

It seems like that would make sense as it allows a larger sample per
tree.

On Mar 20, 12:45 am, dafre...@googlemail.com wrote:
> Hi!
>
> Grow 10000-100000 trees and repeat it 100-1000 times.
>
> Rgrds,
> Daniel
> Sent via BlackBerry from E-Plus
>
>
>
>
>
>
>
> -----Original Message-----
> From: researcher1 <tanze...@gmail.com>
>
> Sender: random...@googlegroups.com
> Date: Sat, 19 Mar 2011 09:12:30
> To: randomjungle<random...@googlegroups.com>

> Reply-To: tanze...@gmail.com

[dafr...@googlemail.com]

In some way it makes sense, but it is difficult to estimate the full consequences of increasing mtry. For example, a higher mtry value leads to more similar trees (highly correlated trees). On the other hand, a very small mtry means: lesser probability of choosing a specific variable (e.g. SNP) and capturing a specific epistatic 'interaction'. Maybe, take a mtry of M/3 (#SNPs/3), we received promising results with that mtry. But it is not a validated recommendation.
Someone should analyse those dependencies..

Best regards,

Daniel
Sent via BlackBerry from E-Plus

-----Original Message-----
From: researcher1 <tanz...@gmail.com>
Sender: random...@googlegroups.com

Date: Sat, 19 Mar 2011 20:34:26
To: randomjungle<random...@googlegroups.com>
Reply-To: tanz...@gmail.com

[researcher1]

Regarding the testing of mtry

Is there a particular reason for the use of mtry = (M/20 * (1, ....,
19)) in your paper?

Or would any other formula which allows you to cycle through the
possible values would be good?

On Mar 20, 4:13 pm, dafre...@googlemail.com wrote:
> In some way it makes sense, but it is difficult to estimate the full consequences of increasing mtry. For example, a higher mtry value leads to more similar trees (highly correlated trees). On the other hand, a very small mtry means: lesser probability of choosing a specific variable (e.g. SNP) and capturing a specific epistatic 'interaction'. Maybe, take a mtry of M/3 (#SNPs/3), we received promising results with that mtry. But it is not a validated recommendation.
> Someone should analyse those dependencies..
>
> Best regards,
> Daniel
> Sent via BlackBerry from E-Plus
>
>
>
>
>
>
>
> -----Original Message-----
> From: researcher1 <tanze...@gmail.com>
>
> Sender: random...@googlegroups.com

> Date: Sat, 19 Mar 2011 20:34:26
> To: randomjungle<random...@googlegroups.com>

> Reply-To: tanze...@gmail.com

[dafr...@googlemail.com]

How many trees did you grow and what data set did you analysed?

Sent via BlackBerry from E-Plus

---

From: Zehao Tan <tanz...@gmail.com>

Date: Wed, 23 Mar 2011 14:36:45 +0800

To: <dafr...@googlemail.com>

Subject: Re: [randomjungle] Re: Number of trees required for reliable results

Hi,

I tried to cycle through the mtry values using the mtry = M/(1, ... , 20).

I have only managed to try M/20, M/10 and M/11.

All the above mentioned run resulted in the same OOB error.

I was wondering this is normal or am I testing the mtry values wrongly?

Just to check the OOB error is in the .confusion or confusion2 files which results from run?

Best Regards,