Response Versus Net Lift Modeling: How to Balance Veracity and Complexity?

When you’re performing a (retention) campaign, it’s not always possible or even desirable to target all your customers. On the one hand budgetary limitations may stand in the way. On the other hand sending out campaigns to certain customers is just a waste of money. This means that we might want to focus only on, for example, the top 10% most interesting customers.

Today’s solution is called response modeling. The main goal of this approach is to develop a model that selects the customers who are most likely to respond and take action. This action can be, for instance, the validation of a coupon, a purchase, an upgrade or a registration.  Several modeling techniques can be applied such as logistic regression, decision trees, neural networks, etc. in order to provide us with an answer. The key in response modeling is that the model you chose will predict which customers have the highest chance to take action based on historic campaign data. These historic campaign data contain a set of customers with their individual characteristics and of course their response.

However, is this really the goal? One aims not only to contact the customers most likely to respond, but to contact the customers who respond because of the campaign. This leads us to another approach, which goes under many names such as net lift modeling, uplift modeling, differential response analysis, true lift modeling or net score modeling. The net lift modeling approach assumes that there are customers who will always take action and that there are customers who will never take action regardless the campaign. It also assumes that there are even customers who react negatively on a campaign regardless of their original feelings. To illustrate this hypothesis, we sketch the following example. We decide to call up customers to try to convince them of an upgrade of their phone subscription. One group of customers wants to make this upgrade anyway. These are the customers you don’t want to contact. The customers who will react positively will cost you a phone call, but the customers who are annoyed by the call, might cost you a good customer. Another customer group isn’t considering the upgrade. This is the group that interests you, as a marketer. If you can convince them to upgrade by means of the call, you create extra value. If you can’t convince them, nevertheless, you again lose the money of the phone call.

As with response modeling, you can use multiple modeling techniques to build your predictive model. The main differences are determined by the type of historic data you use and the maximization function. Net lift modeling tries to optimize the difference in response of the customers who have received the campaign with the customers who didn’t receive the campaign. In this way, we still end up with customers who are most likely to respond, but also with customers who react specifically to the campaign thus avoiding customers who would respond regardless of the campaign. This saves us the cost of the campaign for customers who are already convinced. Moreover, it avoids that we scare away customers (which is even more costly). The downside of this technique is that we need an elaborated historic dataset. These data need not only to contain customers with their individual characteristics and response to the campaign but also the characteristics and responses of customers who didn’t receive the campaign. Not all companies are willing to establish this control group for the purpose of a net lift modeling approach. Moreover, the complexity of net lift models is higher than the one of response models because of the more complicated maximization function.

In the end, when we compare response modeling with net lift modeling we have to ask ourselves which downsides bring us the highest cost and which advantages deliver the highest profit? The biggest advantage of net lift modeling is that it gives us a true answer since it models exactly what we want to know. Response modeling, on the other hand, only gives us half of the truth. Nevertheless, upon deciding if one is willing to collect the additional data necessary for net lift modeling, one needs to consider the following things:

• Is it possible that customers react negatively on our campaign and might even be scared away? If so, it is important to take on a net lift modeling approach since this approach will identify such customers. Studies in the past have already calculated the high cost related to losing an existing customer.
• What is the cost of your campaign? If you take on a response modeling approach, you will also contact customers who will react positively anyway. In this case you lose the money of the campaign without gaining anything.
• One needs to consider if there is a correlation between response and uplift. If customers targeted by net lift modeling are the same as the customers initially selected by the response model, net lift modeling is unnecessary. This might be the case when the individual customer characteristics that make you want to select a customer are also the characteristics that make him/her react on a campaign.

In conclusion, net lift modeling offers an interesting approach but, nevertheless, is also costly in data required and complexity. The true question you need to ask yourself is what are you trying to model exactly? And whether you’re willing to accept possible downsides of the approach chosen.

References

1. Lo, V.S.Y. (2002). The True Lift Model – A Novel Data Mining Approach to Response Modeling in Database Marketing. Explorations, 4(2): p.78-86.
2. Radcliffe, N.J., Surry, P.D. (2011). Real-World Uplift Modelling with Significance-Based Uplift Trees. Portrait Technical Report TR-2011-1, Stochastic Solutions.
3. Rzepakowski, P., Jaroszwicz, S. (2012). Decision Trees for Uplift Modeling with Single and Multiple Treatments. Knowl. Inf. Syst., 32(2): p. 303-327.
4. Guelman, L., Guillén, M., Pérez-Marín, A.M. (2015). Uplift Random Forests. Cybernetics and Systems, 46(3-4): p. 230-248.