Discovery of Positive and Negative Rules from Medical Databases Based on Rough Sets - Advanced Techniques in Knowledge Discovery and Data Mining - page 246

248 Shusaku Tsumoto

9.8 Discussion

9.8.1 Hierarchical Rules for Decision Support

One of the problems with rule induction is that conventional rule-induction
methods cannot extract rules that plausibly represent experts' decision pro-
cesses [9.12]. The description length of induced rules is too short, com-
pared to the experts' rules. (It may be observed that this length part does
not contribute much to the classification performance.) For example, rule-
induction methods introduced in this chapter induced the following common
rule for muscle contraction headache from databases on differential diagnosis
of headache:
[location = whole] ∧ [Jolt Headache = no] ∧ [Tenderness of M1 = yes]

muscle contraction headache.
This rule is shorter than the following rule given by medical experts:

→

[Jolt Headache = no]
∧([Tenderness of M0 = yes] ∨ [Tenderness of M1 = yes]
∨[Tenderness of M2 = yes])
∧[Tenderness of B1 = no] ∧ [Tenderness of B2 = no] ∧ [Tenderness of B3 = no]
∧[Tenderness of C1 = no] ∧ [Tenderness of C2 = no] ∧ [Tenderness of C3 = no]
∧[Tenderness of C4 = no]
→ muscle contraction headache
These results suggest that conventional rule-induction methods do not reflect
a mechanism of knowledge acquisition of medical experts.
Typically, rules acquired from medical experts are much longer than those
induced from databases, the decision attributes of which are given by the same
experts. This is because rule induction methods generally search for shorter
rules, compared with decision tree induction. In the case of decision tree in-
duction, the induced trees are sometimes too deep, and in order for the trees
to be useful for learning, pruning and examination by experts are required.
One of the main reasons rules are short and decision trees are sometimes long
is that these patterns are generated by only one criteria, such as high accuracy
or high information gain. The comparative study in this section suggests that
experts should acquire rules by usage of several measures. Those character-
istics of medical experts' rules are fully examined not by comparing between
those rules for the same class, but by comparing experts' rules with those
for another class. For example, a classification rule for muscle contraction
headache is given by:
[Jolt Headache = no]
∧([Tenderness of M0 = yes] ∨ [Tenderness of M1 = yes]
∨[Tenderness of M2 = yes])
∧[Tenderness of B1 = no] ∧ [Tenderness of B2 = no]
∧[Tenderness of B3 = no]
∧[Tenderness of C1 = no] ∧ [Tenderness of C2 = no]
∧[Tenderness of C3 = no] ∧[Tenderness of C4 = no]
→ muscle contraction headache