Usage: custom targeting policy

Often a randomized experiment is not economical, and so the only data available is data collected according to some targeting policy. In other words, you vary probability of treatment \(P(W=1)\), often according either to some business logic or a model output. As long as all individuals are targeted with some non-zero probability, it is still possible to train an unbiased model using this data by simply weighting calculations on each individual according to \(P(W=1)\).

In its current state, pylift supports this kind of correction to an extent. We have added the ability to correct the Qini-style evaluation curves according to a treatment policy (simply add an argument p, defined as \(P(W=1)\). We’ve also adjusted the transformation to allow the policy information to be encoded in the transformation (pylift.methods.derivatives.TransformedOutcome._transform_func). By specifying a column string that in the keyword argument col_policy that specifies the row-level probability of treatment, this encoding is automatically created.

It is therefore possible to write a custom objective function that recovers treatment, outcome, and policy information from the transformed outcome (using the TransformedOutcome._untransform_func function), then adapting the objective function (if possible – this works with xgboost, but not sklearn) accordingly. However, we have not yet explicitly implemented this custom objective function.