390005_en-US

liliconcom

Additional randomization for bootstrap forest method

For co-linear input variables we would expect similar variable importance.

 

But we noticed that for co-linear input variables that JMP choses which one to use for splitting based on the order they were given to the platform ... instead of randomly selecting among those split variables if there are ties among the top split candidates. Something that seems small... but ends up being important if we want to tell the full story.

 

We brought this up with JMP support on 5/22/2020 (7613090128) and they mentioned they would pass along to JMP R&D. During the recent Advanced Predictive Modeling workshop I brought to the attention of Chris Gotwalt and he quickly proposed some additional randomization that was also quickly implemented in their code base and should be available in 16.2 or 17.0.

 

 

 

 

Statistics and AnalyticsRe: Additional randomization for bootstrap forest method

Confirming that this will be addressed in JMP 17.

Re: Additional randomization for bootstrap forest method - Status changed to: In DevelopmentRe: Additional randomization for bootstrap forest method

We can close this for voting and probably mark as Delivered (or soon to be delivered).