uplift modeling

A marketing model that targets the customers most likely to buy will waste budget on people who would have bought anyway, and may even reach 'sleeping dogs' whom contact annoys into leaving. Uplift modeling targets a different quantity entirely: not who will respond, but who will respond because of the intervention. It ranks individuals by their personal incremental effect, so a campaign spends only where it actually changes behavior.

Uplift is the individual-level reading of the conditional average treatment effect: the difference between the outcome probability under treatment and under control for the same person. Because that difference is never directly observed, models estimate it from experimental or quasi-experimental data using the same meta-learner toolkit, two-model, class-transformation, and tree-based approaches that split on effect heterogeneity rather than on outcome. Evaluation cannot use ordinary accuracy; it uses uplift-specific curves such as the Qini and uplift curves that score how well the ranking concentrates true incremental gains.

Conceptually uplift modeling is causal inference wearing an industrial hat: a CATE estimation and targeting problem dressed for marketing, churn prevention, credit, and medicine. Its practical edge is also its danger, ranking by an unobservable effect makes it acutely sensitive to overlap, sample size, and the quality of the underlying randomization.