Question regarding correlated features

[martheveldhuis]

If I were to group feature A and feature B, and the candidate counterfactual only changes the value of A, will that candidate be rejected? Or is this grouping handled by the sampling space you create for this group of values for A and B?

I am not familiar with Julia so I apologize if this could easily be discerned from the code.

[mjschleich]

Hi Marthe,

Thank you for reaching out!

In GeCo there are two ways to encode that two features are correlated:

1. You can use @group(p, A, B) to declare that A and B are correlated. In this case, the sampling algorithm will always sample values for both A and B, you cannot sample values for A or B independently. So the scenario that you described where only A is changed does not happen.

2. The other option is to define a PLAF constraint to ensure that if A changes then B must change as well. For instance:
   @plaf(p, if cf.A != x.A; cf.B > x.B ) states that if A changes then B must increase.

Depending on your application, it may be better to use option 1 or option 2, but in general it will be more efficient if you use the GROUP operator whenever you can.

Please let me know if this clarified you question, or if there are any other bits that are unclear.