Cost-sensitive Active and Proactive Learning of Decision Trees - Data Mining with Decision Trees: Theory and Applications - page 198

may depend on other considerations, such as her or his price plan. The
client's price plan, hardcoded in the company's database, is often part of
the churn-prediction models. Moreover, if the company seeks for ways to
prevent a client from leaving, it can consider changing the price plan of the
client as a churn-prevention action. Such action, if taken, might affect the
value of an explaining attribute towards a desired direction.
When we refer to “changing the input data”, we mean that in proactive
data mining we seek to implement actions that will change the values of
the explaining attributes and consequently lead to a desired target value.
We do not consider any other sort of action because it is external to the
domain of the supervised learning task. To look at the matter in a slightly
different light, the objective in proactive data mining is optimization ,and
not prediction. In the following section we focus on the required domain
knowledge that results from the shift to optimization, and we define an
attribute changing cost function and a benefit function as crucial aspects of
the required domain knowledge.

12.6.2 Attribute Changing Cost and Benefit Functions
The shift from supervised learning to optimization requires us to consider
additional knowledge about the business domain, which is exogenous to
the actual training records. In general, the additional knowledge may
cover various underlying business issues behind the supervised learning
task, such as: What is the objective function that needs to be optimized?
What changes in the explaining attributes can and cannot be achieved?
At what cost? What are the success probabilities of attempts to change
the explaining attributes? What are the external conditions, under which
these changes are possible? The exact form of the additional knowledge may
differ, depending on the exact business context of the task. Specifically, in
this topic we consider a certain form of additional knowledge that consists
of attribute changing costs and benefit functions. Although we describe
these functions below as reasonable and crucial considerations for many
scenarios, nevertheless, one might have to consider additional aspects of
domain knowledge, or maybe even different aspects, depending on the
particular business scenario being examined.
The attribute changing cost function, C : D R , assigns a real

value cost for each possible change in the values of the explaining attributes.
If a particular change cannot be achieved (e.g., changing the gender of
a client, or making changes that conflict with laws or regulations), the

× D →