In using dataminig models for marketing campaigns, a definition of lift could be:
“lift is a measure of the performance of a targeting model (association rule) at predicting or classifying cases as having an enhanced response (with respect to the population as a whole), measured against a random choice targeting model.” (wikipedia)
Now in plain language for a normal human being:
“lift” means: “how much times is my model better than a random selection”.
Divide you customer database at random in 3 equal parts: part I , part II , and part III
– when you use no model at all: take a random selection A of 10,000 persons out of part I. Send them an email in an effort to sell something. Let’s assume that 2% of them bought you product. Result (no model)=2%
– use your first datamining model (model M1) to select 10,000 persons out of part II. Send them the same email. Let’s assume that 6% of them bought your product. Result (M1)=6%
– use another datamining model (model M2) to select 10,000 persons out of part III. Send them the same email. Let’s assume that 10% of them bought your product. Result (M2)=10%
It is straightforward to see that model M1 does 3 times better than without a model and model M2 does 5 times better.
Formula: lift= ( result of Model X) divided by (result of no model at all)
Lift of first model (M1) => 6% / 2% = 3
Lift of second model (M2) => 10% / 2% = 5