A technical dive into calculating incremental impact, defining customer segments, and visualizing model performance with Propensity and Qini charts.

1 Randomized Experiment

Split customers randomly into a treatment group (receives the campaign) and a control group (receives nothing). This isolation is what lets us measure true causal impact rather than correlation.

Key Definition

Uplift is the difference in behavior caused by the treatment. It answers: "Would this customer have converted anyway, or did the campaign actually move them?"

2 Target Transformation

Standard ML predicts outcomes; uplift modeling predicts the change in outcomes. Two common approaches:

T-Learner: Train separate models for treatment and control, then subtract predictions.

Class Transformation: Create a single transformed target variable that directly encodes uplift, enabling any classifier to learn incremental effects.

3 Scoring

For each customer, compute an uplift score that quantifies how much the treatment shifts their probability of converting:

Uplift = P(Buy | Treatment) - P(Buy | Control)

Uplift scores naturally divide customers into four actionable segments. The goal is to find and target the Persuadables while avoiding the Sleeping Dogs.

Sleeping Dogs
Negative uplift — the campaign actually hurts conversion. These customers are less likely to buy if contacted.
Do Not Disturb
Sure Things
Zero uplift — they will convert regardless of whether they see the campaign. Spending here is waste.
Ignore
Lost Causes
Zero uplift — they will not convert regardless. No amount of marketing will move them.
Ignore
Persuadables
Positive uplift — the campaign causes conversion. These are the customers where marketing dollars create real value.
Target Aggressively
Persuadable
Sleeping Dog
Sure Thing
Lost Cause
How to read this chart: Each dot is a customer plotted by their treatment propensity (x-axis) vs. control propensity (y-axis).
  • The diagonal line represents zero uplift (treatment = control).
  • Points below the diagonal are persuadables — higher probability of converting with treatment.
  • Points above the diagonal are sleeping dogs — treatment actually reduces their conversion probability.
Understanding the Qini Curve: It measures how well your model identifies persuadables when you target customers in order of predicted uplift.
  • The diagonal is random targeting — no model, just chance.
  • A good model shoots up early, capturing most incremental conversions in the first deciles.
  • A bad model dips below the diagonal, meaning it targets sleeping dogs first.
  • The shaded area (AUC) between the curve and the diagonal quantifies model quality.