About evaluation metrics for contextual uplift modeling

In this tech report we discuss the evaluation problem of contextual uplift modeling from the causal inference point of view. More particularly, we instantiate the individual treatment effect (ITE) estimation, and its evaluation counterpart. First, we unify two well studied fields: the statistical ITE approach and its observational counterpart based on uplift study. Then we exhibit the problem of evaluation, based on previous work about ITE and uplift modeling. Moreover, we derive a new estimator for the uplift curve, built on logged bandit feedback dataset, that reduces its variance. We prove that AUUC fails on non randomized control trial (RCT) datasets and discuss some corrections and guidelines that should be kept in mind while using AUUC (re-balancing the population, local importance sampling using propensity score).