Cumulative uplift curve

For the cumulative uplift curve, we compute for subsets of the test set the cumulative uplift scaled by the number of observations in the subset, then divide it by the total number of test observations ({{ perf.testTotalPopSize }}). Close to the origin, the steeper the curve the better.

The baseline represents the expected effects when following a uniform random assignment. The baseline is increasing as the average treatment effect is positive. The baseline is decreasing as the average treatment effect is negative.

See more in

Related metrics

The AUUC (Area Under the Uplift Curve) is the area between the cumulative uplift curve and the random assignment line. The AUUC is: {{ perf.normalized.auuc | nicePrecision: 5 }}, which means the model performs worse than a random treatment assignment. The AUUC is: {{ perf.raw.auuc | nicePrecision: 5 }}, which means the model performs worse than a random treatment assignment.

The net uplift at {{ perf.netUpliftPoint * 100 }}% is the difference between the values of the cumulative uplift curve and the random assignment line at {{ perf.netUpliftPoint * 100 }}% of the test population. The net uplift at {{ perf.netUpliftPoint * 100 }}% is: {{ perf.normalized.netUplift | nicePrecision: 5 }}, which means the model performs worse than a random treatment assignment. The net uplift at {{ perf.netUpliftPoint * 100 }}% is: {{ perf.raw.netUplift | nicePrecision: 5 }}, which means the model performs worse than a random treatment assignment.

Note that these metrics implicitly assume the treatment is randomized. When it is not the case, they are not a good measure of performance.

Qini curve

For the Qini curve, we compute for subsets of the test set the cumulative uplift scaled by the number of treated observations in the subset, then divide it by the total number of treated test observations ({{ perf.testTreatedPopSize }}). Close to the origin, the steeper the curve the better.

The baseline represents the expected effects when following a uniform random assignment. The baseline is increasing as the average treatment effect is positive. The baseline is decreasing as the average treatment effect is negative.

See more in

Related metric

The Qini score is the area between the Qini curve and the random assignment line. The Qini score is: {{ perf.normalized.qini | nicePrecision: 5 }}, which means the model performs worse than a random treatment assignment. The Qini score is: {{ perf.raw.qini | nicePrecision: 5 }}, which means the model performs worse than a random treatment assignment.

Note that this metric implicitly assumes the treatment is randomized. When it is not the case, they are not a good measure of performance.