Cost-sensitive Active and Proactive Learning of Decision Trees - Data Mining with Decision Trees: Theory and Applications - page 197

lifecycle of that customer. Since the exact LTV of a customer is revealed
only after the customer stops being a customer, managing existing LTVs
requires some sort of prediction capability. While data mining algorithms
can assist in deriving useful predictions, the CRM decisions that result from
these predictions (for example, investing in customer retention or customer-
service actions that will maximize her or his LTV) are left in the hands of
humans.
In proactive data mining, we seek automatic methods that will not
only describe a phenomenon, but also recommend actions that affect the
real world. In data mining, the world is reflected by a set of observations.
In supervised learning tasks, which are the focal point of this topic,
each observation presents an instance of the explaining attributes and the
corresponding target results. In order to affect the real world and to assess
the impact of actions on the world, the data observations must encompass
certain changes. We discuss these changes in the following section.
12.6.1 Changing the Input Data
We consider the training record, <x 1 ,n ,x 2 ,n ,...,x k,n ; y n > for some
specific n . This record is based on a specific object in the real world. For
example, x 1 ,n ,x 2 ,n ,...,x k,n may be the explaining attributes of a client,
and y n , the target attribute, might describe a result that interests the
company, whether the client has left or not.
It is obvious that some results are more beneficial to the company than
others, such as a profitable client remaining with the company rather than
leaving it or that clients with high LTV are more beneficial than those with
low LTV. In proactive data mining, our motivation is to search for means
of actions that lead to desired results (i.e., desired target values).
The underlying assumption in supervised learning is that the target
attribute is a dependent variable whose values depend on those of the
explaining attributes. Therefore, in order to affect the target attribute
towards the desired, more beneficial, values, we need to change the
explaining attributes in such a way that target attributes will receive the
desired values.
Consider the supervised learning scenario of churn prediction, where a
company observes its database of clients and tries to predict which clients
will leave and which will remain loyal. Assuming that most of the clients
are profitable to the company, the motivation in this scenario is churn
prevention. However, the decision of a client about whether to leave or not