
Table 1: The number of columns in the expanded and mini-
mized DTs.
Data set Columns in Columns in
expanded DT minimized DT
German 6600 11
Bene1 192 14
For example, in the DT for Bene1 (cf. Figure 6),
column blocks {2, 3, 4, 5} and {9, 10, 11, 12}, though
having the same respective action values, are not eli-
gible for contraction, because they differ in more than
one condition entry (viz., with respect to the attributes
‘purpose’ and ‘term’). On the other hand, a deci-
sion diagram, which allows the sharing of one such
instance through multiple incoming edges, might be
smaller than the corresponding tree or table. There-
fore, in addition to the DT, we have built an equivalent
MDD representation based on the extracted rule set,
thereby adhering to the same ordering of attributes as
in the minimum-size DT. Figure 7 presents the re-
sulting diagram for Bene1. It was produced using
the Graphviz graph-drawing software (Gansner et al.,
1993; AT&T, 2003).
Unlike in Figure 6, the part of the MDD represen-
tation that matches the replicated table segment is in-
cluded only once: the subgraph rooted at the right-
most of the two ‘years client’-nodes is effectively
shared through its two incoming edges. Hence, de-
scribing the function in MDD format results in a more
compact representation, because the merging rule, un-
like the DT contraction rule, does apply here. This
empirically confirms why we consider a decision di-
agram to be a valuable alternative knowledge visu-
alization aid. Nevertheless, decision diagrams are so
far seldom considered in this context, despite their be-
ing a graph-based generalization of the far more fre-
quently applied decision tree representation.
We have repeated the same exercise for the Ger-
man credit data set, but in the latter case, no further
size savings could be obtained vis-`a-vis the DT rep-
resentation. In Table 2, we have summarized the re-
sults of the MDD construction process for both data
sets. Note that, because of the aforementioned rela-
tion between a decision tree and a lexicographically
ordered DT, the figures in column (2) also match the
number of splits appearing in the condition entry part
of the corresponding DT. Consequently, the final col-
umn provides a measure of the additional size gains of
MDD reduction over DT contraction (i.e., the added
effect of graph sharing).
Both decision tables and diagrams facilitate the de-
velopment of powerful decision-support systems that
can be integrated in the credit-scoring process. A DT
consultation engine typically traverses the table in a
top-down manner, inquiring the user about the con-
Table 2: MDD size results
Data set Intern. nodes Intern. nodes Size
in MDD (1) in dec. tree (2) saving
German 8 8 0%
Bene1 9 12 25%
dition states of every relevant condition encountered
along the way. A similar decision procedure is in-
duced when consulting the decision diagram repre-
sentation, and following the proper path through the
graph. Hence, both types of representations provide
efficient decision schemes that allow a system imple-
mentation to ask targeted questions and neglect irrel-
evant inputs during the question/answer-dialog. Fur-
thermore, given the availability of efficient condition
reordering operations for both types of representa-
tions, questions can be easily postponed during this
process. For example, in PRO LOG A, the available an-
swer options always include an additional ‘unknown’
option, which allows the user to (temporarily) skip the
question. When that happens, the DT’s conditions are
first reordered internally: moving the corresponding
condition to the bottom of the order and then recon-
tracting the DT may result in new don’t care entries
being formed for it. After that, the session continues
with the next question. If, at some point, a conclu-
sion is reached regarding the DT’s actions, the former
question could effectively be avoided; else, it eventu-
ally pops up again.
In the Bene1 example, suppose that we are deciding
on a particular applicant whose properties will even-
tually be found to match against the condition entries
of column 12, which tells us to accept the loan. Be-
fore arriving at that conclusion, we are required to
provide only 4 of the 7 inputs to make a classifica-
tion decision: ‘term’, ‘owns property’ and ‘income’
successively turn out to be irrelevant for this case. If,
on the other hand, we would consult the rule descrip-
tion shown in Figure 5, we would need to evaluate
every single rule, thereby testing its antecedent until a
condition is found that fails, before we may conclude
that none of the rules applies and that the default class
(applicant = good) must be chosen.
5 CONCLUSIONS
In this paper, we have shown how credit-scoring
knowledge can be compactly visualized either in the
form of decision tables or diagrams. For two real-
life cases, it was first of all shown how a set of
propositional if-then rules, extracted by a neural net-
work rule extraction algorithm, can be represented
as a decision table. The constructed decision tables
ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS
230