Overview of the Uplift Platform

Use the Uplift platform to model the incremental impact of an action, or treatment, on individuals. An uplift model helps identify groups of individuals who are most likely to respond to the action. Identification of these groups leads to efficient and targeted decisions that optimize resource allocation and impact on the individual. See Radcliffe and Surry (2011).

The Uplift platform fits partition models. Although traditional partition models select splits to optimize classification, uplift models select splits to maximize treatment differences.

The uplift partition model accounts for the fact that individuals are grouped by a treatment factor. To determine splits, models are fit to all possible binary splits of each factor. The type of model that is fit is dependent on the type of response. A continuous response is modeled as a linear function of the split, the treatment, and the interaction of the split and treatment. A categorical response is expressed as a logistic function of the split, the treatment, and the interaction of the split and treatment. In either case, the interaction term measures the difference in uplift between the groups of individuals in the two splits. The most significant split is selected and the process repeats.

The Uplift platform selects the most significant split based on the significance of interaction terms in each of the binary split models. However, predictor selection based solely on p-values introduces bias in favor of predictors with many levels that result in many models for the single predictor. For this reason, JMP adjusts p-values to account for the number of levels or models considered. The correction used is based on Monte Carlo simulation. See Sall (2002). The splits are determined by the minimum adjusted p-values for tests of the significance of the interaction effect across models for all possible binary splits across all predictors. The logworth for each adjusted p-value, namely -log10(adj p-value), is reported.