AI Glossary / Governance, ethics and safety / Intermediate
Lift curve
Definition
A lift curve shows how effectively a model concentrates positive cases compared with random selection as the selected fraction of a dataset changes. In a cumulative lift curve, the highest-scored cases are included first and performance is compared at the same selection depth.
What it means
A ranking model is often used when a team can examine only part of a population. The question is not just whether the model classifies well overall, but whether its top-ranked cases contain more relevant positives than an equally sized random group would contain on average.
Lift expresses that enrichment as a ratio. A lift of 2 at a specified depth means twice the positive concentration of the random baseline at that depth. It does not mean twice the profit, twice the overall accuracy or a causal doubling of outcomes caused by an intervention.
How it works
Score a labeled evaluation dataset and sort examples from highest to lowest predicted likelihood of the positive class. At each selected fraction q, count how many actual positives occur in the selected group.
Cumulative lift can be calculated as the selected group's positive rate divided by the overall positive rate. Equivalently, divide the fraction of all positives captured by q. Random selection has expected lift 1. When the full dataset is selected, cumulative lift is 1, provided the dataset contains positives.
Tie handling, sampling and score direction need to be specified. A decile lift chart may report separate bins rather than cumulative groups, so read the axes and definitions carefully. Very small selected groups can give unstable estimates, especially when positives are rare.
Three practical examples
- Quality review: A team with limited inspection capacity ranks items for checking and examines whether the highest-scored group contains more actual defects than random selection.
- Support prioritization: Analysts evaluate a model that identifies requests likely to need additional assistance, checking lift at the team's available review capacity.
- Model comparison: Two ranking models are compared at several selection depths because one may be stronger for the top 5% while another performs better when 30% can be reviewed.
Think of it this way
Think of sorting a mixed basket so that desired items appear nearer the top. Lift asks how much richer the top portion is than an equally sized handful drawn at random. The analogy does not account for differing item values, review costs or the effects of taking action after selection.
Worked example
A fictional evaluation set contains one thousand items, including one hundred actual positives. A model's top one hundred items contain thirty positives.
The selected fraction is 100/1,000=0.1. The selected group's positive rate is 30/100=0.3, compared with the overall rate 100/1,000=0.1. Lift is 0.3/0.1=3. Equivalently, the model captures 30% of all positives in 10% of the population, giving 0.3/0.1=3.
The selected group still contains seventy negatives, and seventy positives remain outside it. Lift therefore needs to be read alongside precision, recall, workload and the consequences of errors.
Illustrative case study
Fictional case: Review Harbour chooses a model to prioritize manual checks. Model A has higher lift in the top 5%, but the team can inspect 20% of cases each day. At that depth, model B captures more positives.
The team evaluates both at the relevant capacity and checks whether performance varies across important categories. It keeps the ranking test separate from any claim that intervention changes outcomes. This invented case illustrates choosing an operating point rather than treating one headline lift number as decisive.
Common misconceptions and mistakes
Lift is not a causal effect. A model may identify people likely to respond without identifying those whose response would change because of a campaign. That latter question requires a different evaluation design.
Do not compare lift numbers at different selection depths as if they were directly equivalent. Do not evaluate on training data or choose the best-looking depth after repeatedly inspecting the final test set. Changing class prevalence and population composition can also affect comparisons. Finally, a high lift on a tiny group may correspond to very few cases and considerable uncertainty.
Business use
Lift curves help connect ranking quality with limited operational capacity. They are useful when teams need to decide how many cases to review and where a model provides value beyond random selection.
Translate the chosen operating point into expected workload, missed positives and false alarms using representative data. Add cost and outcome analysis separately rather than assuming lift equals financial return. Monitor the relevant depth over time and retain absolute counts, since ratios alone can hide changes in scale and prevalence.
People also ask
What does lift 1 mean?
The selected group's positive concentration matches the random-selection baseline at that depth.
Why does cumulative lift reach 1 at 100% selection?
The selected group becomes the whole dataset, so its positive rate equals the overall rate.
Is a lift curve the same as a gains curve?
No. Gains typically shows the fraction of positives captured; lift divides that fraction by the selected population fraction.
Check your understanding
The top 20% of cases contains 40% of all positives. What is cumulative lift at that depth?
Answer: 0.40/0.20=2. The result does not by itself specify profit or a causal intervention effect.
Related terms
- Precision
- Recall
- Uplift modeling
- Cumulative gains
- ranking model
- selection depth
- class prevalence
- random baseline
Learn next
Sort a small labeled dataset by model score and calculate cumulative positives at several depths. Plot both gains and lift, keeping their y-axes distinct. Mark the actual review capacity and count false positives and missed positives there. Repeat on held-out data before using the curve to guide decisions.
More in Governance, ethics and safety
- Learning - Learning, in AI, is the process through which a system uses experience or data to develop or improve its ability to perform a task.
- Least privilege - Least privilege is the security principle of giving a person, account, or software component only the access and capabilities needed for its function.
- Lemmatization - Lemmatization is the process of mapping a word form to its dictionary-style base form, called a lemma.
- Linking Bias - Linking bias describes systematic differences in the networks inferred from people's recorded connections, interactions or activity.
- Local interpretable model-agnostic explanations - Local Interpretable Model-agnostic Explanations, usually called LIME, is a method for explaining an individual model prediction by fitting a simpler...
- Loss Of Situational Awareness Bias - Loss of situational awareness bias is a source-taxonomy term for a problem in which automation leaves people insufficiently aware of the current situation...
Sources
ITU-T Y Supplement 97, term 3.1.65, consulted: https://www.itu.int/rec/T-REC-Y.Sup97-202511-I/en. SAS Visual Analytics Reference, Lift Plot page, consulted: https://documentation.sas.com/api/collections/vacdc/v_008/docsets/varef/content/varef.pdf?locale=en. SAS note 41683 also supplied an indexed primary excerpt describing cumulative lift: https://support.sas.com/kb/41/683.html; direct page access timed out. The entry explicitly uses cumulative selection depth and distinguishes separate-bin charts.
Previous: Lift Next: Lifting lemma