Uplift modelling
On this page
Definition
Uplift modelling estimates the causal effect of a treatment at the level of an individual user: how much their metric changes because of a campaign. A regular model predicts an outcome; an uplift model predicts the difference between the outcome with and without treatment. This lets you target those who respond, not those who would convert anyway.
How to compute
Models are trained on experimental data where treatment is known: T-learner, S-learner, X-learner, causal forest. Ranking quality is measured with QINI and AUUC — the area under the cumulative-uplift curve over sorted scores. A good model lifts the metric at the same reach.
Pitfalls
Without a control arm the causal effect cannot be separated from correlation. Segmenting on the predicted outcome is not the same as segmenting on the effect: “likely to convert” is not “persuadable”. On small samples QINI and AUUC cannot tell even an oracle from noise, so trust the metric only with enough observations per arm.