Uplift Modeling

Abstract

Uplift modeling is a machine learning technique which aims at predicting, on the level of individuals, the gain from performing a given action with respect to refraining from taking it. Examples include medical treatments and direct marketing campaigns where the rate of spontaneous recovery and the background purchase rate need to be taken into account to assess the true gains from taking an action. Uplift modeling addresses this problem by using two training sets: the treatment dataset containing data on objects on which the action has been taken and the control dataset containing data on objects left untreated. A model is then built which predicts the difference between outcomes after treatment and without it conditional on available predictor variables. An obvious approach to uplift modeling is to build two separate models on both training sets and subtract their predictions. In many cases, better results can be obtained with models which predict the difference in outcomes directly. A popular class of uplift models are decision trees with splitting criteria favoring tests which promote differences between treatment and control groups. Ensemble methods have proven to be particularly useful in uplift modeling, often leading to significant increases in performance over the base learners. Linear models, such as logistic regression and support vector machines, have also been adapted to this setting. Dedicated methods, such as uplift or qini curves, are necessary for evaluating uplift models. Application of the methodology to survival data and scenarios with more than one possible action have also been considered.