Uplift modeling

Also called incremental response modeling.

Business basics2min read

Uplift modeling estimates how much an action, such as an email or offer, changes the probability of an outcome. It predicts the incremental effect of the action, rather than simply predicting who is likely to buy.

How it works

Randomized treatment and control data are the cleanest starting point. For each customer profile, the model estimates conversion probability with the action minus conversion probability without it. Validate those estimates on held-out experimental data. Purchase history alone cannot establish causation.

For a startup, this helps avoid spending scarce acquisition dollars on people who would buy anyway.

Higher conversion does not mean higher uplift

In this hypothetical randomized campaign, each segment has 1,000 treated customers and 1,000 controls who receive no campaign.

SegmentTreated conversionsControl conversionsUplift
A150, or 15%100, or 10%5 percentage points
B410, or 41%400, or 40%1 percentage point

Segment A produces an estimated 1,000 × (15% - 10%) = 50 extra sales. Segment B produces only 10.

Each extra sale contributes $40 before campaign costs, and each treated segment costs $1,000 to contact:

  • A: (50 × $40) - $1,000 = $1,000 net contribution.
  • B: (10 × $40) - $1,000 = -$600 net contribution.

Targeting the customers most likely to buy would favor B. Targeting incremental profit favors A.

Questions people ask

What is the difference between uplift modeling and propensity modeling?

In this marketing context, propensity modeling predicts how likely someone is to buy. Uplift modeling predicts how much an intervention changes that likelihood. A high-propensity customer might purchase without any intervention, so contacting them adds little value.

What is the difference between uplift and lift?

Usage varies. In targeting, lift often means a model-selected group's response rate divided by a baseline response rate. Uplift modeling estimates a treatment's incremental effect. Always specify the comparison and whether the result is a percentage-point difference or a relative percentage change.

Related terms

Sources

  1. Pierre Gutierrez and Jean-Yves Gérardy, "Causal Inference and Uplift Modelling: A Review of the Literature," Proceedings of Machine Learning Research, 2017.