Is Multi Target Regression (MTR) available to use on RapidMiner?

AayushShah

I am new to this platform and have a use case in mind in which I wanted to predict 4 targets (all continuous) together. 

Somone had already asked this question before on this thread: here
Since that question was asked in 2018, I think its probably a good time to get an update on that. 

So the question is: Is Multi Target Regression possible on RapidMiner?

Thank you.

BalazsBaranyRM

Hi!

Yes, this is possible using the techniques described in that discussion.

Since you are predicting four targets, it's more efficient to implement this using Loop Attributes and selecting the targets there. This will then execute the regression modeling and validation (and whatever you put into the process) for all four attributes.

Regards,
Balázs

AayushShah

So that means, all targets as assumed to be independent and they will be trained individually, isn't it? Thus all same features will be used for all four targets, am I picking up right?

Thank you for the replay!
Aayush Shah

BalazsBaranyRM

Depending on your use case, you should probably exclude the three alternative labels when you're building the models.

Normally, the target attribute (label) is not available when the learning data set is already there. So you should train label X without the other possible labels, unless you expect them to be available later when applying the model (scoring).

Other than the labels, yes, you would predict the four labels from the same data - the other attributes.

Regards,
Balázs

BalazsBaranyRM

Hi!

Yes, this is possible using the techniques described in that discussion.

Since you are predicting four targets, it's more efficient to implement this using Loop Attributes and selecting the targets there. This will then execute the regression modeling and validation (and whatever you put into the process) for all four attributes.

Regards,
Balázs

AayushShah

So that means, all targets as assumed to be independent and they will be trained individually, isn't it? Thus all same features will be used for all four targets, am I picking up right?

Thank you for the replay!
Aayush Shah

BalazsBaranyRM

Depending on your use case, you should probably exclude the three alternative labels when you're building the models.

Normally, the target attribute (label) is not available when the learning data set is already there. So you should train label X without the other possible labels, unless you expect them to be available later when applying the model (scoring).

Other than the labels, yes, you would predict the four labels from the same data - the other attributes.

Regards,
Balázs

AayushShah

Thanks @BalazsBarany
From my understanding to approach this problem I will have to create 4 different models each of them will be trained with the same features.

Model 1 = X1..Xn | Y1
Model 2 = X1..Xn | Y2
Model 3 = X1..Xn | Y3
Model 4 = X1..Xn | Y4

Right? Okay so I will try this way.

BalazsBaranyRM

Hi,

yes, this is right in the technical sense. It also has to be right for your use case logically. Only you can determine that.

Regards,
Balázs