Uplift Modeling: Who to Target, Not Just Who Will Convert

WHY IT EXISTS Standard predictive models tell you who is likely to convert, but not whether your action caused that conversion. You might waste marketing budget on customers who would have bought your product anyway. Uplift modeling was created to solve this by isolating the incremental effect of an intervention.

THE MENTAL MODEL Think of your audience in four groups. First, "Sure Things": they will convert whether you contact them or not. Second, "Lost Causes": they will not convert no matter what. Third, "Persuadables": they will only convert if you contact them. Fourth, "Sleeping Dogs": they will be annoyed by your contact and become less likely to convert. A standard conversion model finds the Sure Things and Persuadables. An uplift model specifically finds the Persuadables.

HOW IT WORKS Uplift modeling requires data from a randomized controlled trial (an A/B test) with a treatment group (gets the intervention) and a control group (does not). The model is trained to predict the difference in outcome probability between these two groups for any given individual. This "uplift score" represents the net change in behavior caused by your action. This can be done with a two-model approach (one for treatment, one for control) or specialized algorithms that optimize for the uplift score directly.

WHEN TO USE IT Use it when you have a costly or limited resource and want to maximize its impact. It's ideal for targeted marketing (who gets a coupon?), proactive churn prevention (who gets a retention offer?), and personalizing user experiences. The goal is to find the individuals most influenced by your action.

WHEN NOT TO USE IT Do not use it if you cannot run a controlled experiment, as you need both treatment and control data. It is also overkill if the intervention is so cheap you can apply it to everyone, or if you only need to predict an outcome without understanding the causal impact of your actions.

ONE CANONICAL EXAMPLE A marketing team wants to send a 20% discount offer to drive sales. A standard model would target users with a high likelihood to buy, many of whom are "Sure Things." An uplift model, trained on past campaign data with a control group, would instead identify users for whom the 20% offer has the largest positive impact on their probability to purchase—the "Persuadables." This maximizes the return on the discounts offered.