Add penalty when true CATE is known

[arita37]

Typically, we have some past experiments where we have estimates by (A/B test)
of the true CATE.

When using, DECI, we obtain some CATE_estimate, due to wrong causal graph (ie bias).
Is there a way to add penalty from actual True CATE in the loss so the true Graph can be recovered ?

One possibile way would be (ie very rough idea) :

1. Generate some graphs
2. calculate CATE
3. back-propagate error = (CATE- CATE_true)**2
4. and re-train model.

Do you have any suggestion on how to adjust the learning of graph
in DECI ?

[confoundry]

Hi @arita37,

Do you have access to the interventional samples used to generate the CATE? If so, you can create the intervened SEM (using the do function) and train that on the interventional data.

Full conditional samples are not supported at the minute, but if you're conditioning on root nodes then it's the same thing as an intervention. However, if you're conditioning on values in the middle of the DAG it will be difficult.

The way you have suggested would also work, providing you "soft" sample the graphs (as we do in the training).