Uplift Score
1) Meaning
An Uplift Score is the numerical value assigned by an uplift model that estimates the incremental impact (positive or negative) of a treatment (e.g., marketing campaign, discount, medical treatment) on an individual’s likelihood of taking an action.
It tells us: “How much more (or less) likely is this person to act if we intervene compared to if we don’t?”
2) Formula / Concept
At the individual level:
$\text{Uplift Score}_i = P(\text{Outcome} \mid \text{Treatment}, i) – P(\text{Outcome} \mid \text{Control}, i)$
Where:
- $P(\text{Outcome} \mid \text{Treatment}, i)$ = predicted probability of outcome if treated.
- $P(\text{Outcome} \mid \text{Control}, i)$ = predicted probability of outcome if not treated.
- Positive Uplift Score → treatment increases likelihood of desired outcome.
- Negative Uplift Score → treatment decreases likelihood (e.g., backlash).
- Near zero → treatment makes little difference.
3) Example
Marketing campaign for a subscription:
- Customer A:
- Predicted purchase probability if treated = 0.40
- Predicted purchase probability if not treated = 0.25
- Uplift Score = 0.40 – 0.25 = +0.15 (15%) → good target.
- Customer B:
- Treated = 0.70
- Not treated = 0.68
- Uplift Score = +0.02 (2%) → small incremental effect.
- Customer C:
- Treated = 0.10
- Not treated = 0.20
- Uplift Score = –0.10 (-10%) → campaign actually makes things worse (Do-Not-Disturb).
4) How It’s Used
- Ranking customers: Target those with the highest uplift scores (Persuadables).
- Campaign optimization: Avoid spending on Sure Things (high purchase probability even without campaign) and Lost Causes (low probability regardless).
- Risk management: Exclude customers with negative uplift scores (who might churn or unsubscribe if targeted).
5) Evaluation
- Aggregate uplift scores are used to build Uplift Curves / Qini Curves.
- Qini Coefficient / AUUC quantify how well the model ranks individuals by uplift.
Bottom line:
The Uplift Score is the individual-level prediction from an uplift model that measures the causal effect of treatment vs. no treatment. It allows businesses to target Persuadables, avoid wasted spend, and prevent negative responses.