# Response Uplift Modeling

In traditional response modeling, a group of treated customers is taken into consideration, and a predictive model is built to distinguish between responders and non-responders, and this is done through the use of the technique of predictive modeling. A decision tree or analysis is used for this purpose. The treated customers are only used for building the model. Whereas, the variation in the difference between a treated and a control group is measured with the help of an uplift model. Different names like incremental impact modeling, true" response modeling, differential response analysis and net modeling are used in discussions of uplift modeling. The true responses are separated from purchasers in an uplift modeling.

High prediction accuracy, for a given data set is aimed, by most of the classification approaches. In most of the practical cases, for example, in mailing an offer, the change in probabilities is taken into account, rather than the class probabilities. Action is taken on the objects which, is found to be the most profitable one. In machine learning literature, uplift modeling, true lift modeling, or differential response analysis is of little importance. A tree based classification can be done in this regard. Pruning method and new splitting criteria can be designed here. A significant improvement over the previous technique of uplift modeling can be done through the proposed approach, and this can be confirmed through experiments.

The change in behavior is not modeled in an ordinary response modeling; the behavior of a person subjected to some influence is modeled. Whereas, the change in behavior is modeled by the uplift models, where change occurs due to an influence, such as how much more he will spend, etc.

In mathematical terms, a response model predicts in the following manner:

P (purchase | treatment),

Which, refers to the conditional probability of purchase, when some specific treatment is present.On the other hand, the prediction done by an uplift model is as follows:

P (purchase | treatment) - P (purchase | no treatment)

The above refers to the change or difference between the conditional probability of purchase, for a specific treatment and the corresponding probability when the customer is not subject to that treatment.

TutorTeddy.com & Boston Predictive Analytics

[ Email your Statistics or Math problems to tutor@aafter.com (camera phone photos are OK) ]

Boston Office (Near MIT/Kendall 'T'):
Cambridge Innovation Center,