Multiple responses Random Forest approach

[Jonison]

Hello, all, can I ask a question about random forest. I got SAS miner to play with random forest, but it only can do prediction on one response. Unfortunately, my case has 6 responses. I did some research that multiple responses random forest is applicable. Could you give me some information about how to do RF with multiple responses in SAS environment ?

Many thanks 

[Reeza]

Can you expand on your 6 responses?

Can they be combined into one?

---

@Jonison wrote:

Hello, all, can I ask a question about random forest. I got SAS miner to play with random forest, but it only can do prediction on one response. Unfortunately, my case has 6 responses. I did some research that multiple responses random forest is applicable. Could you give me some information about how to do RF with multiple responses in SAS environment ?

Many thanks 

---

[Jonison]

the 6 responses are produced by 6 individual sensors, and no strong linear correlation among them, if this helps.

[PaigeMiller]

There seems to be an R package named MultivariateRandomForest, but I have never used it.

In my opinion, this is a weakness of the modeling in SAS Enterprise Miner and the "Model Studio" (that might not be the correct name) in SAS Viya, almost all the modeling methods work on a single response variable.

One exception is a modeling technique called Partial Least Squares, which SAS provides and can handle multiple response variables. I don't really know if the interfaces from Enterprise Miner and "Model Studio" allow multiple responses to be specified, but PROC PLS and PROC HPPLS in SAS certainly do allow multiple responses.

[sbxkoenk]

Hello @Jonison ,

Technically speaking, you want to conduct a multivariate multiple RF.  This RF is "multivariate" because there is more than one outcome variable (several dependent variables).  It is a "multiple" RF because there is more than one predictor variable.

There are several statistics and econometrics procedures in SAS that can simultaneously deal with more than one outcome variable (2 or 2+ outcome variables).

You can of course use those procedures in a code node of SAS Enterprise Miner or Model Studio but you may want to avoid that. 😊

I'm going to have a look tomorrow at which standard nodes in EM support multiple dependent variables (simultaneously).
I don't know offhand, but random forest(s) probably doesn't (you're right about that).

But all nodes based on neural networks certainly do! The output layer of the NN can have multiple nodes (and I am not talking about a multi-nomial or ordinal response, but multiple continuous / interval-scaled responses).

Good night,

Koen