What is uplift@k?

uplift@k is a performance metric used in uplift modeling / causal ML.
It measures the incremental effect (uplift) achieved if we only target the top k% of customers (or samples) ranked by the model’s predicted uplift score.
In other words: “If I only contact the top k% of customers, how much additional impact do I get compared to not contacting them?”

Formula

For the top k% of the ranked list:

$\text{Uplift@k} = \frac{y^T_{(k)}}{n^T_{(k)}} – \frac{y^C_{(k)}}{n^C_{(k)}}$

Where:

$y^T_{(k)}$: Total outcomes (e.g., purchases) of the treatment group within the top k%
$n^T_{(k)}$: Number of treatment samples in the top k%
$y^C_{(k)}$: Total outcomes of the control group within the top k%
$n^C_{(k)}$: Number of control samples in the top k%

So uplift@k = difference in average outcome between treatment and control in the top-k% segment.

A random targeting strategy would give a small or zero uplift (since treatment and control would perform similarly).
A good uplift model ensures that in the top k% segment, treatment leads to much better outcomes than control.
That difference is captured by uplift@k.

Suppose you have 10,000 customers, and you target the top 20% (k=20%) = 2,000 customers according to your model.

$\text{Uplift@20\%} = 0.15 – 0.10 = 0.05$

Interpretation: In the top 20% selected by the model, treatment increases the purchase probability by 5 percentage points compared to control.

Marketing: If you send promotions only to the top k% customers, uplift@k tells you how much incremental sales you would generate.
Healthcare: If you administer treatment only to the top k% of patients (those most likely to benefit), uplift@k measures the incremental recovery rate.
Recommendation systems: If you recommend items only to the top k% users predicted to respond, uplift@k measures the extra engagement/CTR gained.

Metric	Uplift	Uplift@k
Definition	Average treatment effect across the entire population	Incremental effect within the top k% segment
Purpose	Measures overall treatment effect	Evaluates the model’s ranking quality (how well it identifies responsive users)
Use Case	Checking whether the treatment works in general	Deciding which subset of users to target for best ROI

uplift = “How effective is treatment overall?”
uplift@k = “How effective is my model at selecting the best subset of users to treat?”