Uplift or Persuasion Modeling
A combination of treatment comparisons (e.g. send a sales solicitation to one group, send nothing to another group) and predictive modeling to determine which cases or subjects respond (e.g. purchase or not) to which treatments. Here are the steps, in conceptual terms, for a typical uplift model:
1. Conduct A-B test, where B is control
2. Combine all the data from both groups
3. Divide the data into a number of segments, each having roughly similar numbers of subjects who got treatment A and control. Tree-based methods are typically used for this.
4. The segments should be drawn such that, within each segment, the response to treatment A is substantially different from the response to control.
5. Considering each segment as the modeling unit, build a model that predicts whether a subject will respond positively to treatment A.
The challenge (and the novelty) is to recognize that the model cannot operate on individual cases, since subjects get either treatment A, OR control, but not both, so the “uplift” from treatment Z compared to control cannot be observed at the individual level, but only at the group level. Hence the need for the segments described in steps 3 and 4.
Note: Traditional A-B testing would stop at step 1, and apply the more successful treatment to all subjects.