AUUC (Area Under the Uplift Curve)

1. Definition

• AUUC measures the cumulative incremental gain achieved by an uplift model, integrated over the entire population.
• It is the total area under the uplift curve (incremental gain vs. population proportion).
• It tells you how effective the model is overall at ranking individuals by uplift.

---

2. Key Idea

• In uplift modeling, we compare treatment vs control outcomes.
• The uplift curve shows cumulative incremental responses as you target more of the population (starting from the highest predicted uplift).
• AUUC quantifies the total incremental benefit of following the model’s ranking.

---

3. Computation (Conceptual)

1. Sort customers by predicted uplift score (descending).
2. Partition into bins (e.g., deciles or percentiles).
3. For each bin $k$:
   • Compute incremental gain:
   • $\text{Uplift}_k = \frac{y^{T}_k}{n^{T}_k} – \frac{y^{C}_k}{n^{C}_k}$
   • where:
     • $y^{T}_k$​ = outcomes in treatment
     • $y^{C}_k$​ = outcomes in control
     • $n^{T}_k, n^{C}_k$ = number of samples
4. Plot cumulative uplift vs population proportion.
5. AUUC = numerical area under this curve.

---

4. Formula (Integral Form)

If $U(x)$ is cumulative uplift (incremental response at population fraction $x$):

$\text{AUUC} = \int_0^1 U(x) \, dx$

---

5. Example (Simplified)

• Random targeting yields incremental gain ≈ 0 (curve near baseline).
• A strong model shows a steep uplift curve: targeting top 20% gives 80% of all possible incremental responses.
• AUUC quantifies how much total uplift is captured.

---

6. Relation to Qini Coefficient

• AUUC = raw area measure (depends on dataset size, response rate).
• Qini Coefficient = normalized AUUC (scaled relative to a perfect model and random targeting).
• Think of it like this:

Metric	Analogy
AUUC	Accuracy (raw performance, dataset-dependent)
Qini	AUC (normalized, comparable across models/datasets)

---

Summary

• AUUC = area under cumulative uplift curve → total incremental effect measured.
• Qini coefficient = normalized AUUC → comparability across datasets.
• Both are core metrics for uplift model evaluation.