Qini Curve, AUUC & Uplift Evaluation

Uplift Model Offline Evaluation Metrics (Qini Curve, Qini Score, AUUC - Area Under Uplift Curve & Cumulative Gain Charts) solve the fundamental evaluation challenge that individual counterfactual treatment effects are unobservable, rendering classical ROC-AUC/LogLoss metrics inapplicable; Qini construction algorithm: 1) Model Scoring & Sorting: sorting test users in descending order of predicted CATE scores $\hat{\tau}(x)$; 2) Cumulative Segment Tallying: at top $k\%$ cumulative population slice, tallying treatment converters $n_{t, 1}(k)$ and control converters $n_{c, 1}(k)$; 3) Cumulative Qini Metric: $Q(k) = n_{t, 1}(k) - \frac{N_t(k)}{N_c(k)} n_{c, 1}(k)$, quantifying the net incremental conversions directly caused by treatment; 4) AUUC measures the integral area under the Qini curve relative to the diagonal random targeting baseline; higher AUUC indicates superior concentration of persuadable users at the top of the funnel.

Offline Uplift model benchmarking, determining optimal marketing budget cutoff percentiles, and financial incrementality calibration.

Eliminates the fundamental impossibility of computing pointwise MSE against unobserved counterfactuals by measuring group-level incremental lift curves.

Derive the Qini curve formula and explain why the scaling factor $\frac{N_t(k)}{N_c(k)}$ is required to adjust for treatment-control sample imbalance?

Contrast Qini Curve vs AUUC in coordinate normalization and cumulative percentage definitions?

How to construct the random baseline and the theoretical Oracle upper bound on Qini visualization plots?

Why must a calibrated Uplift model exhibit strict monotonic degradation across sorted decile bins?

How to perform Uplift Calibration checks comparing predicted CATE against empirical bucketed treatment-control rate margins?