Uplift Modelling
Who should get the offer?.
A company built a model to find customers who are likely to leave, and it sent them all a discount. Churn went up. Why? Some had forgotten that they were subscribed, and the email reminded them. Don't target people who are likely to leave. Target people whose behaviour you can change.
- Find likely leavers, send offers
- Churn went up
- Target those you can change
Four kinds of customer
Picture four groups. Persuadables stay only if you act. Those are your target. Sure things stay anyway, so the offer's wasted. Lost causes will leave whatever you do. And sleeping dogs leave because you contacted them, so leave those alone. A normal model can't separate them. Uplift modelling tries to.
Two models
We need data from a randomised test. Then we train two models. One learns from people who got the offer, and the other from people who didn't. The difference between their predictions is the uplift for each person. In this simulation, the offer helps only customers who visit rarely. Run it. On average, it lowers churn by about seven points. But look at the split. Low visitors see a large effect, and high visitors hardly any. So send it to the first group, and keep the discount money from the second.
import numpy as np
import pandas as pd
from sklearn.ensemble import HistGradientBoostingClassifier as Model
rng = np.random.default_rng(0)
n = 20000
visits = rng.gamma(2, 2, n)
offer = rng.integers(0, 2, n)
base = 0.35 - 0.02 * visits.clip(0, 12)
effect = np.where(visits < 4, -0.12, 0.0)
leave = rng.random(n) < (base + effect * offer).clip(.01, .99)
X = visits.reshape(-1, 1)
a = Model(max_iter=60).fit(X[offer == 1], leave[offer == 1])
b = Model(max_iter=60).fit(X[offer == 0], leave[offer == 0])
uplift = a.predict_proba(X)[:, 1] - b.predict_proba(X)[:, 1]
print("Average effect of the offer:", round(uplift.mean(), 3))
for name, mask in [("Low visits", visits < 4), ("High visits", visits >= 4)]:
print(f"{name:<12} {uplift[mask].mean():+.3f}")Average effect of the offer: -0.067 Low visits -0.103 High visits -0.016
Summary
Target people that you can change, not just people who are at risk. You've now learned advanced causal methods that most analysts never reach. Next, let's return to forecasting, and test forecasts properly.
Key takeaways- Four kinds of customer
- Train two models
- Target the persuadable