uplift modeling

A predictive method that targets the difference an intervention makes rather than the outcome itself: for each person, it estimates how much an action such as a drug, a discount or a call to a wavering customer would shift their expected outcome compared with doing nothing. In causal inferencecausal inference, n.: Working out whether one thing actually causes another, rather than merely occurring alongside it — for example, whether a new feature made users stay longer or just happened at the same time. terms this is the conditional average treatment effect given the person's features. Because no one is observed both treated and untreated, models are trained on data in which treatment was assigned at random, as in an A/B testA/B testing, n.: Comparing two versions of something — two models, two prompts, two designs — by randomly showing each to different groups of real users and measuring which works better., or with careful adjustment for confounding. Approaches include separate outcome models for treated and control groups (the two-model or T-learner), a single model with treatment as a feature (S-learner), X-learners, class-transformation tricks, uplift decision treesdecision tree, n.: A model that makes a prediction by asking a series of yes-or-no questions about the input, like a flowchart, with each answer leading to the next question until it reaches a final answer. whose splits maximize the treatment–control difference, and causal forests. Evaluation uses Qini and uplift curves on held-out randomized data. Marketers classically divide customers into persuadables, sure things, lost causes and 'sleeping dogs', whom an offer would put off.

The uplift model told us to stop calling loyal customers, since the retention call made no difference to them.