Uplift modeling (also called incremental response modeling or true-lift modeling) predicts how much a specific treatment — a variant, an offer, an email — changes an individual's probability of converting. It answers a sharper question than a standard propensity model: not "who is likely to convert?" but "whose decision is actually changed by this treatment?"
Those are not the same people. Someone who would have converted anyway gives you no incremental lift; showing them the treatment is wasted effort or even counterproductive. Uplift modeling isolates the visitors whose behavior the treatment genuinely moves.
The Four Customer Segments
Uplift modeling implicitly sorts people into four groups based on how they respond to treatment vs. no treatment:
| Segment | Converts if treated? | Converts if not treated? | Worth targeting? |
|---|---|---|---|
| Persuadables | Yes | No | Yes — this is the prize |
| Sure things | Yes | Yes | No — you'd convert them anyway |
| Lost causes | No | No | No — nothing works |
| Sleeping dogs | No | Yes | No — treatment hurts |
The goal is to find the persuadables and concentrate spend on them, while avoiding sleeping dogs whom the treatment actively pushes away.
How It Works
Uplift requires an experiment, not just historical data, because you need to observe both treated and untreated outcomes:
- Run a randomized test with a treatment group and a holdout (control) group.
- Model each outcome as a function of user features — typically the predicted conversion probability with treatment and without.
- Take the difference — the predicted uplift is
P(convert | treated) − P(convert | not treated)for each individual. - Target by uplift, sending the treatment to those with the highest predicted incremental effect.
When to Use Uplift Modeling
- You're running targeted offers or discounts and want to avoid giving them to people who'd buy anyway.
- You're doing personalization at scale and want to allocate a treatment to whoever benefits most.
- Your treatment has a real cost (a discount, a sales touch, an email send) so incremental impact matters more than raw conversion likelihood.
Limitations
- Data-hungry. Estimating a difference of two probabilities is noisier than estimating one, so it needs substantial experiment data.
- Requires a holdout. You cannot learn uplift without withholding treatment from some users, which has a short-term cost.
- Harder to validate. You never observe both outcomes for the same person, so evaluating an uplift model requires specialized metrics (e.g., Qini or uplift curves) rather than standard accuracy.
Uplift modeling sits behind sophisticated personalization and contextual bandit systems, where the aim is not just to predict conversion but to act only where the action changes the outcome.