Causal inference — uplift and heterogeneous treatment effects.
Estimating who a treatment helps: the difference between the outcome with and without treatment, per individual. Unlike a plain regression, these models take (X, treatment, y) and predict the individual treatment effect, or uplift, which is what decides who to treat.
Algorithms
-
SLearner: One model on
[X, treatment]; uplift is the difference
between the treated and control predictions. Simplest, but the treatment signal can get diluted among the covariates.
-
TLearner: Two models, one per treatment group; uplift is their
difference. Lets each group have its own response surface.
-
XLearner: T-learner plus two "imputed effect" models fitted on the
cross-group residuals, combined with a propensity weight. Often the best of the meta-learners when treatment groups are imbalanced.
-
UpliftTreeClassifier: A single tree that splits directly on uplift gain,
so the leaves are groups with genuinely different treatment effects.
Utilities
-
qini_curve — cumulative incremental
gain of treating the top-ranked individuals.
-
auuc — area under the uplift curve.
-
uplift_at_k — mean uplift of the
top-k predicted group.
Notes
treatment is binary (0 = control, 1 = treated) and must contain both groups. y is a numeric outcome; uplift is E[y | t=1] - E[y | t=0].