Meta-actions as a Tool for Action Rules Evaluation - Feature Selection for Data and Pattern Recognition - page 182

(

,
where the left hand side of the action term took place at a visit that occurred prior to
the right hand side. Furthermore, the transitions weremined from instances belonging
to a unique patients. In summary, the introduction of the instances order and their
identifier with regard to the object eliminated any confusion.
The purpose of meta-actions is to trigger transitions in feature values that will
change the object state. Eventually, the transitions triggered by meta-actions will
trigger an action rule and will cascade into transitioning the decision feature value.
Asmentioned earlier, meta-actions' transitions are a subset of the transitions extracted
in the action rules extraction process. In addition, meta-actions play a passive role,
in the sense that they do not change decision feature values. They rather inform deci-
sion makers of possible transitions. On the contrary, action rules play an active
role that help decision makers drive their strategy, and explicitly look into the
transitions that affect the decision feature. In other words, meta-actions are not
replacing action rules; instead, enhancing the process of selecting the best action
rules.

TSize

,

7

ₒ

6

) ∧ (

CLevel

,

4

) ∧ (

NChemo

,

5

ₒ

6

) ∧ (

State

,

Unstable

ₒ

Stable

)

9.5.1 Action Rules Selection by Meta-actions

In the effort of selecting the best action rules, meta-actions play an important eval-
uation role. After extracting action rules, meta-actions inform us, amongst other
things, on whether the extracted action rules are applicable. Meta-actions also pro-
vide decision makers with the confidence of executing the antecedent side of an
action rule. Given a set of action rules, and an influence matrix describing meta-
actions transitions in our system, we strive to select the best applicable action rules
and their respective triggers. This is done by first dividing the set of action rules
into applicable and non-applicable action rules. Applicable action rules are the rules,
which antecedent sides are covered by the influence matrix. Then for each action
rule, we select the best coverings of the action rule from the influence matrix. The
best coverings of an action rule are the maximal action terms in the influence matrix.
By maximal action term, we mean the action term with the largest number of atomic
terms covering the action rule. This is performed by intersecting the antecedent side
of the action rule with all the action terms in the influence matrix. It is important
to note that the action terms with higher number of atomic action terms will have a
higher confidence. As one can guess, two or more action terms may be required to
cover an action rule. Similarly, two or more meta-actions may be required to cover
an action rule.
We described earlier how to compute the confidence of an action term Te r m -
Conf in 9.2 ; however, we still need to define how to compute the confidence of
multiple action terms, namely global confidence GlobConf . Computing the confi-
dence of multiple action terms depends on whether the action terms belong to the
same meta-action or not. In fact, action terms

that belong to different
meta-actions are independent since they are extracted from different object pop-

{ t 1 , t 2 ,..., t n }