Information Technology Reference
In-Depth Information
problems within a given domain. The classical knowledge discovery algorithms
have the potential to identify enormous number of significant patterns from data.
Therefore, people are overwhelmed by a large number of uninteresting patterns
which are very dicult to analyze and use to form timely solutions. So, there is
a need to look for new techniques and tools with the ability to assist people in
identifying rules with useful knowledge.
There are two types of interestingness measure: objective and subjective
(see [10], [1], [19], [20]). Subjective interestingness measures include unexpect-
edness [19] and actionability [1]. When a rule contradicts the user's prior belief
about the domain, uncovers new knowledge, or surprises him, it is classified as
unexpected. A rule is deemed actionable, if the user can take action to gain an
advantage based on this rule. Domain experts basically look at a rule and say
that this rule can be converted into an appropriate action.
E-action rules mining is a technique that intelligently and automatically as-
sists humans in acquiring useful information from data. This information can
be turned into actions which can benefit users. The approach gives suggestions
about how to change certain attribute values of a given set of objects in order
to reclassify them according to a user wish.
There are two frameworks for mining actionable knowledge: loosely coupled
and tightly coupled [9]. In the tightly coupled framework, action rules are ex-
tracted directly from a database [7], [8], [22]. In the loosely coupled framework,
proposed in [15], the extraction of actionable knowledge is preceded by classifi-
cation rules discovery. It is further subdivided into:
•
strategies generating action rules from certain pairs of classification rules [18],
[21], [23],
•
strategies generating action rules from single classification rules [16], [24].
This paper relates to a loosely coupled framework. In most of the algorithms
for action rules mining, there is no guarantee that the discovered patterns in the
first step will lead to actionable knowledge that is capable of maximizing profits.
One way to approach this problem is to assign a cost function to all changes of
attribute values [24]. If changes of attribute values in the classification part of an
action rule are too costly, then they can be replaced by composing this rule with
other action rules, as proposed in [23]. Each composition of these rules uniquely
defines a new action rule. Objects supporting each new action rule, let's say
r
,
are the same as objects supporting the action rule replaced by
r
but the cost of
reclassifying them is lower for the new rule.
E-action rule models the actionability concept in a better way than action rule
[15] by introducing a notion of its supporting class of objects. E-action rules are
constructed from certain pairs of classification rules. They can be used not only
for evaluating discovered patterns but also for reclassifying some objects in a
dataset from one state into a new more desired state. For example, classification
rules found from a bank's data can be very useful to describe who is a good
client (whom to offer some additional services) and who is a bad client (whom
to watch carefully to minimize the bank loses). However, if bank managers need
to improve their understanding of customers and seek for specific actions to
