Uplift Modeling Lab — Persuadables, Sure Things & Sleeping Dogs

Campaign Population

Population size 400

Persuadable share 30%

Sleeping-dog share 10%

Targeting Policy

Predicted-uplift threshold 0.10

Model noise (score vs true uplift) 0.08

Campaign Result

Targeted (of pop.)

0 / 0

True incr. conversions

0.0

Qini coefficient

0.000

Sleeping-dog backfire

0.0

Persuadable — targeting them creates conversions
Sure Thing — converts either way
Lost Cause — never converts
Sleeping Dog — treatment backfires

How it works

Every customer has two unobservable conversion probabilities: p_C if left alone (control) and p_T if sent the campaign (treated). The individual treatment effect — the uplift — is the difference:

τ(x) = P(convert | treated, x) − P(convert | control, x)
      = p_T(x) − p_C(x)

A real uplift model never sees both outcomes for the same customer (the fundamental problem of causal inference), so it is trained as a T-learner: one classifier fit on the historical control group predicts p̂_C(x), a second fit on the historical treated group predicts p̂_T(x), and the predicted uplift is τ̂(x) = p̂_T(x) − p̂_C(x). This simulator generates the ground-truth p_C and p_T for every synthetic customer directly, then adds Gaussian noise to mimic an imperfect model's predicted score τ̂(x) — the Model noise slider controls how far the score drifts from the truth.

Plotted on the scatter, x = p_C and y = p_T for each customer; the diagonal plane y = x is zero uplift. Four quadrants fall out automatically:

Persuadables (above the diagonal, low p_C) — only convert because of the campaign; targeting them is where budget should go.
Sure Things (near the diagonal, high p_C) — convert regardless; treating them is wasted spend.
Lost Causes (near the diagonal, low p_C) — never convert; also wasted spend.
Sleeping Dogs (below the diagonal) — the campaign actively pushes them away (annoyance, opt-outs); treating them costs conversions.

The Predicted-uplift threshold sets the targeting rule: only customers with τ̂(x) above it receive the campaign. True incremental conversions sums the real τ(x) only over targeted customers — the actual conversions the campaign caused, which can go negative if too many sleeping dogs are swept in. The Qini coefficient is the area between this policy's cumulative-uplift curve (customers ranked by τ̂) and the curve from targeting at random, normalized to roughly [−1, 1] — the standard metric for how much better an uplift model does than blind targeting; it collapses toward 0 as Model noise grows, because a noisier score ranks customers no better than chance.

x = P(convert | control), y = P(convert | treated) Diagonal = zero uplift plane

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.