AI Glossary / Data science and analytics / Intermediate
Lift
Definition
Lift is a ratio that compares an observed or model-selected outcome with a stated baseline. In association-rule analysis, lift compares the frequency of two events occurring together with the frequency expected under independence. In ranking evaluation, lift compares the target rate in a selected portion of a ranked dataset with the overall target rate. These uses are related, but the baseline and selection rule must be specified.
What it means
A high success rate can be less impressive when success is already common. Lift provides context by dividing by a baseline rate. A selected group with a 30% target rate has lift 3 relative to a 10% baseline, but lift 1.2 relative to a 25% baseline. The ratio describes relative concentration or association; it is not automatically a causal effect, a probability, or a measure of profit.
How it works
For an association rule A implies B, confidence is P(B given A). With positive event probabilities, lift is confidence divided by P(B), equivalently P(A and B) divided by P(A) times P(B). A value above one indicates more co-occurrence than the independence baseline, below one less, and one matches that baseline for the two events. For a cumulative ranking lift at a chosen fraction, sort records by model score, calculate the target rate in that top fraction, and divide by the overall rate in the same evaluation population.
Three practical examples
- Transaction analysis: a retailer compares the observed co-purchase rate of two items with their individual purchase frequencies.
- Review prioritization: a model's top-scored cases are checked for a greater concentration of relevant outcomes than a random selection of the same size would have on average.
- Campaign planning: analysts compare ranked customer segments using historical response labels, while separately investigating whether an intervention changes behavior.
Think of it this way
Think of comparing a selected basket of fruit with the whole shipment. Finding many apples is more informative when apples are rare in the shipment than when nearly every fruit is an apple. The analogy concerns composition, not causation: selecting a basket does not create the apples, and selecting high-propensity customers does not prove a campaign caused their purchases.
Worked example
Suppose 1,000 transactions include item A in 200 cases and item B in 100 cases. Both appear together in 40. Confidence for A implies B is 40/200 = 0.20. The overall rate of B is 100/1,000 = 0.10, so lift is 0.20/0.10 = 2. Under the independence baseline, the expected joint count is 1,000 times 0.20 times 0.10 = 20. The observed count of 40 is twice that baseline.
Reversing the rule gives confidence 40/100 = 0.40, but its baseline is now A's rate of 0.20, so lift remains 2. Lift is symmetric for this pair even though confidence is not.
For a separate ranking illustration, suppose an evaluation set has 100 positive cases among 1,000 records. The top 200 scored records contain 50 positives. Their positive rate is 25%, compared with 10% overall, so cumulative lift at the top 20% is 2.5. The model captures 50% of all positives there, a different quantity from lift.
Illustrative case study
In a fictional retention project, a team sees strong historical lift among customers most likely to renew and proposes giving that group a discount. Another analyst notes that many selected customers may renew without any discount. The team distinguishes a response-propensity ranking from the incremental effect of the offer and designs an appropriate treatment comparison before claiming additional renewals. The case shows why a model can concentrate likely outcomes without identifying the people whose behavior an intervention will change.
Common misconceptions and mistakes
Lift above one does not establish statistical significance, causation, or commercial value. Rare events can generate large ratios from few records, making supporting counts essential. Different datasets can have different baseline rates, so equal lift values need not imply equal target rates or workload. A cumulative lift curve and a lift value for a single score band answer different questions. Do not confuse a lift ratio of 2 with an increase of two percentage points. A zero baseline makes the ordinary ratio undefined.
Business use
State the baseline population, target definition, evaluation period, and selected fraction beside a lift result. Include absolute counts and target rates so reviewers can judge scale and uncertainty. Compare models on consistent data and selection capacity. If the business can review only 500 cases, performance at that capacity may matter more than an impressive value in a tiny top segment. Evaluate costs, benefits, and intervention effects separately; lift is one analytical input, not a complete business case.
People also ask
Can lift exceed one?
Yes. It is a ratio, not a probability constrained to the interval from zero to one.
Does lift 2 mean a causal doubling?
No. It means twice the specified baseline under the stated calculation.
Is association-rule lift directional?
The lift value is symmetric for the same two events, although the rule's confidence changes when its direction is reversed.
Check your understanding
In the ranking example, why is lift 2.5 rather than 50%?
Answer: Lift divides the selected group's 25% target rate by the population's 10% rate. The 50% figure describes the share of all positive cases captured in the selected group.
Related terms
- Baseline
- Precision
- Uplift modeling
- Selection bias
- Association rule
- support
- confidence
- cumulative gains
- response rate
Learn next
Calculate support, confidence, and lift from one small transaction table. Then build a ranked evaluation table and compare cumulative lift, capture rate, and selected-group target rate. Study causal uplift separately before using a propensity ranking to decide who should receive an intervention.
More in Data science and analytics
- L1-norm principal component analysis - L1-norm principal component analysis refers to PCA-related methods that use absolute-value criteria instead of the squared-value criteria associated with...
- Labeled data - Labeled data are examples paired with target information for a defined learning or evaluation task.
- Latent variable - A latent variable is a variable included in a model but not directly observed in the available data.
- Line chart - A line chart displays values as points connected in an ordered sequence, commonly to show change over time.
- List - A list is an ordered collection of elements. In programming, the exact operations and implementation depend on the language.
- Local outlier factor - Local outlier factor, or LOF, is a method that scores how unusual an observation is relative to the density around its neighbors.
Sources
IBM, association-rule lift formula and independence baseline: https://www.ibm.com/docs/en/db2/11.1.0?topic=SSEPGG_11.1.0%2Fcom.ibm.im.model.doc%2Fc_lift_in_an_association_rule.htm IBM, ranking lift chart and random-selection comparison: https://www.ibm.com/docs/en/db2/11.1.0?topic=cs-lift-chart MIT, original data-mining lecture on association-rule measures: https://www.ocw.mit.edu/courses/15-062-data-mining-spring-2003/d286f7eda9814013279284c97f2f2a53_Lecture_16.pdf
Previous: LIFO queue Next: Lift curve