getting more positive responders for the money expended, and
reduce customer fatigue by contacting only those customers most
likely to show interest.
From a data mining perspective, the goal is to classify each indi-
vidual as a responder or nonresponder with an associated probability.
Data mining classification algorithms are well suited for this task.
Customers can then be ranked from those with the highest probabil-
ity of response to the lowest probability of response. Choosing those
customers at the top of the list provides a high concentration of
responders. This is reflected in a “lift” chart as depicted in Figure 2-3.
The notion of lift is discussed in detail in Chapter 7, but in short, lift
provides a simple understanding of how much better the predictions
of the data mining model are than a random selection of customers.
To use data mining for response modeling, it is important to have
relevant historical data about which customers responded and did
not respond to campaigns in the past. “Relevant data” means that the
data is for a similar situation, for example, purchase of a product or
type of product, completing a survey, etc. Moreover, there needs to be
sufficient demographic and other customer-related information from
which data mining algorithms can “learn” the patterns or types of
customer that respond. If there is no historical data available, a trial
campaign can be performed on a random subset of the potential cus-
tomers. Assuming there are a sufficient number of responders, the
data mining algorithm can learn what distinguishes a responder from
% of Total Cases
Getting “lift” on responders.