Advanced Decision Trees - Data Mining with Decision Trees: Theory and Applications - page 170

YRSJOB

<2

2

≥

Option
Node

LT V

≥ 75%

<75%

MARST

M

A

DEPEND

>0

=0

D A M

D

A

Fig. 11.3 Illustration of option tree.

Figure 11.3 illustrates an option tree. Recall that the task is to classify
mortgage applications into: approved (“A”), denied (“D”) or manual
underwriting (“M”). The tree looks like any other decision tree with a
supplement of an option node, which is denoted as a rectangle. If years at
current job (YRSJOB) is greater than or equals two years, then there are
two options to choose from. Each option leads to a different subtree that
separately solve the problem and make a classification.
In order to classify a new instance with an option tree, it is required to
weight the labels predicted by all children of the option node. For example,
in order to classify an instance with YRSJOB=3, MARST=“Married” and
DEPEND=3, we need to combine the fifth leaf and the sixth leaf (from left
to right), resulting in the class “A” (since both are associated with class
“A”). On the other hand, in order to classify an instance with YRSJOB=3,
MARST=“Single” and DEPEND=3, then we need to combine the fourth
leaf and sixth leaf, resulting in either class “A” or “D”. Since in simple
majority voting, the final class is selected arbitrarily, it never makes sense
to create an option node with two children. Consequently the minimum
number of choices for an option node is three. Nevertheless, assuming that
a probability vector is associated with each leaf, it is possible to use a
Bayesian combination to obtain a non-arbitrary selection. In fact, the model