Cost-sensitive Active and Proactive Learning of Decision Trees - Data Mining with Decision Trees: Theory and Applications - page 190

with a marketing offer. Based on their response, the learner actively selects
the next customers to be approached and so forth. Exploration does not
come without a cost. Direct costs might involve hiring special personnel
for calling customers and gathering their characteristics and responses to
the campaign. Indirect costs may be incurred from contacting potential
customers who would normally not be approached due to their low buying
power or low interest in the product or service offer.
A well-known concept of marketing campaigns is the explo-
ration/exploitation trade-off. Exploration strategies are directed towards
customers as a means of exploring their behavior; exploitation strategies
operate on a firm's existing marketing model. In the exploration phase,
a concentrated effort is made to build an accurate model. In this phase,
the firm will try, for example, to acquire any available information which
characterizes the customer. During this phase, the results are analysed in
depth and the best modus operandi is chosen. In the exploitation phase
the firm simply applies the induced model with no intention of improving
the model to classify new potential customers and identify the best ones.
Thus, the model evolves during the exploration phase and is fixed during
the exploitation phase. Given the tension between these two objectives,
research has indicated that firms first explore customer behavior and then
follow with an exploitation strategy. The result of the exploration phase is
a marketing model that is then used in the exploitation phase.
Let consider the following challenge. Which potential customers should
a firm approach with a new product offer in order to maximize its net
profit? Specifically, our objective is not only to minimize the net acquisition
cost during the exploration phase, but also to maximize the net profit
obtained during the exploitation phase. Our problem formulation takes
into consideration the direct cost of offering a product to the customer,
the utility associated with the customer's response, and the alternative
utility of inaction. This is a binary discrete choice problem, where the
customer's response is binary, such as the acceptance or rejection of a
marketing offer. Discrete choice tasks may involve several specific problems,
such as unbalanced class distribution. Typically, most customers considered
for the exploration phase reject the offer, leading to a low positive response
rate. However, a simple classifier may predict that all customers in questions
will reject the offer.
It should be noted that the predictive accuracy of a classifier alone
is insucient as an evaluation criterion. One reason is that different
classification errors must be dealt with differently: mistaking acceptance