Reviews: Uplift Modeling from Separate Labels

This paper proposes an approach to heterogeneous treatment effect estimation (what it calls "uplift modeling") from separate populations. A simple version of the setup of this paper is as follows. We have two populations, k 1, 2, with different probabilities of treatment conditional on observed features, Pk[T X] (the paper also allows for the case where these need to be estimated). We have access to covariate-outcome pairs (X, Y) drawn from both populations, so we can estimate Ek[Y X]. We assume potential outcomes Y(-1), Y(1), and assume that E[Y(T) X] doesn't depend on setup k. What we would really want is to estimate a conditional average treatment effect tau(x) E[Y(1) - Y(-1) X x].